Materials Science & AI, where deep learning meets the periodic table.

Materials science is the discipline that connects atomic-scale physics and chemistry to macroscopic engineering properties. It sits between physics (which describes how atoms and electrons behave) and engineering (which uses materials to build things), and it is the natural language for understanding why steel is strong, why silicon is the substrate of modern electronics, why ceramics resist heat, and why polymers are flexible. Modern materials science is also one of the most-active AI-for-Science application areas: the Materials Project and similar databases provide millions of computed materials properties as training data, machine-learned interatomic potentials are now standard tools for atomic-scale simulation, generative models propose novel crystal structures conditioned on target properties, and autonomous laboratories close the design-make-test loop with robotic synthesis. This chapter develops both the working materials-science vocabulary an AI reader needs (Sections 2–9) and the AI methodology that has reshaped the field (Sections 10–19). Section 10 frames what makes materials-AI methodologically distinctive from an ML perspective; this section maps the materials-science territory itself.

The structure-property-processing triangle

The most useful framing of materials science for an AI reader is the structure-property-processing triangle. The structure of a material — atomic arrangement, defects and microstructure, phase composition — determines its properties (mechanical strength, thermal and electrical conductivity, optical response, magnetic behaviour). The processing route (synthesis, heat treatment, deformation, additive manufacturing) determines the structure. Materials science is the systematic study of these three sides and the relationships between them, and AI methods are increasingly central at every vertex: structure prediction (Sections 2–4 of this chapter, plus AI methods in Sections 11, 15), property prediction (Section 6, plus AI methods in Sections 11–14), and processing-aware optimisation (autonomous laboratories, Section 16).

The length-scale hierarchy

Materials phenomena span enormous scale ranges. Atomic-scale physics (~Å) governs bonding, electronic structure, and the formation energies that DFT computes (Section 8). Defect-scale phenomena (~nm) — point defects, dislocations, grain boundaries (Section 4) — control mechanical strength and diffusion. Microstructural features (~µm) — grain structure, phase distributions, precipitates — set most engineering properties. Macroscopic behaviour (~mm and beyond) is what engineering applications care about. Different theories and computational methods apply at different scales, and bridging between them — using ML methods to connect ab-initio calculations with mesoscale dynamics with macroscopic engineering models — is one of the most-active research frontiers. The materials-AI chapter develops the toolkit for the atomic-to-microstructural piece of this hierarchy.

Why this is one chapter, not two

The vocabulary and the methods are tightly intertwined. Crystal-graph neural networks (Section 11) only make sense once crystal structures and their symmetries are understood (Section 2). ML interatomic potentials (Sections 13–14) only make sense once defects and the energy landscape are understood (Section 4). Property prediction (Section 12) only makes sense once the targets — band gaps, formation energies, elastic moduli (Sections 6–7) — are understood. Generative crystal design (Section 15) only makes sense once the chemistry of the periodic table is understood. Reading just the AI half without the vocabulary leaves an AI practitioner unable to evaluate methodological choices; reading just the vocabulary leaves a materials scientist unaware of how the field is being reshaped. The 19-section structure is therefore deliberate: §2–9 develop the vocabulary, §10 bridges to the methodology, §11–19 develop the methods.

Crystals are solids in which atoms are arranged in periodic patterns that extend throughout the material. The periodic structure dramatically simplifies analysis (calculations on infinite crystals reduce to calculations on a single repeating unit) and enables most of the powerful theoretical machinery — from X-ray diffraction to band theory — that modern materials science uses.

The unit cell and the lattice

The fundamental concept is the unit cell: the smallest repeating unit that, translated periodically in three dimensions, generates the entire crystal. The unit cell is defined by three lattice vectors a, b, c and the three angles between them; together these define the Bravais lattice. There are exactly 14 distinct Bravais lattices in three dimensions (Bravais, 1850), classified by the symmetries that leave them invariant: cubic, tetragonal, orthorhombic, monoclinic, triclinic, hexagonal, rhombohedral, with additional centred variants (face-centred, body-centred, base-centred). The lattice describes only the translational symmetry; the actual atomic positions within the unit cell are described by the basis (the set of atoms at specified positions within the unit cell). A crystal structure is fully specified by the lattice plus the basis.

Common structures worth knowing

A handful of crystal structures recur across thousands of materials. Face-centred cubic (FCC): atoms at the cube corners and face centres; the most-densely-packed structure for spheres and the structure of Cu, Al, Au, Ag, Ni, and many more. Body-centred cubic (BCC): atoms at corners and the cube centre; the structure of Fe (below 911°C), W, Mo, Nb. Hexagonal close-packed (HCP): the other close-packed structure; the structure of Mg, Ti, Zn, Co. Diamond cubic: the structure of diamond, Si, Ge, and a substantial fraction of semiconductor materials; effectively two interpenetrating FCC lattices. Rock salt: alternating cation/anion FCC sublattices; NaCl, MgO, many other ionic compounds. Perovskite: ABO₃ structure with B at the centre, A at corners, and O at face centres; the structure of many functional oxides including high-temperature superconductors and ferroelectrics. Wurtzite and zincblende: the structures of compound semiconductors like GaN and GaAs. Recognising these by name is essential vocabulary for materials-AI applications.

Space groups and symmetry

Beyond translational symmetry, crystals have point-group symmetries: rotations, reflections, and inversions that leave the structure invariant. Combining the 14 Bravais lattices with the 32 crystallographic point groups (and adding "screw" and "glide" symmetries that combine rotations and reflections with translations) produces the 230 space groups that classify all possible 3D crystal structures. The space groups are tabulated in the International Tables for Crystallography, the canonical reference. Each space group is identified by a number (1–230) and a symbol (e.g., Fd-3m for diamond, Pm-3m for cubic perovskite). Modern crystal-structure databases routinely include space-group designations, and ML methods that engage with crystal structures often use space-group equivariance as an architectural prior.

Miller indices and crystallographic planes

Miller indices (h,k,l) label crystallographic planes by the reciprocals of where they cross the unit-cell axes (with appropriate scaling and sign conventions). The (100), (110), and (111) planes are the most-commonly-discussed in cubic crystals — their relative atomic densities, their surface reconstructions, and their roles in mechanical deformation are all distinctive. Crystallographic directions are labelled with square brackets [hkl] (referring to the vector connecting the origin to (h,k,l) in the lattice) and the equivalence class of all symmetry-related directions is labelled with angle brackets ⟨hkl⟩. The vocabulary is essential for discussing mechanical anisotropy, electronic-structure features along specific directions, and surface science.

Reciprocal lattice and the Brillouin zone

For computational solid-state physics, the most-important construction is the reciprocal lattice: the Fourier-conjugate of the real-space lattice. Each real-space lattice has a corresponding reciprocal lattice in k-space (wavevector space), and the first Brillouin zone is the Wigner-Seitz cell of the reciprocal lattice — the region of k-space closest to the origin. Electronic states, phonon modes, and various other physical quantities are most-naturally described as functions of wavevector k in the Brillouin zone, with high-symmetry points (Γ at the origin, plus K, X, M, L, etc. at zone-boundary points depending on the lattice) playing distinguished roles. Section 7 develops band theory, which is fundamentally a description of how electron energies depend on k throughout the Brillouin zone.

Beyond perfect crystals

Real materials are rarely perfect crystals. Polycrystals are aggregates of many small crystallites (grains) with different orientations; amorphous materials (glasses, many polymers) lack long-range periodic order entirely; quasicrystals have ordered but non-periodic structures (Shechtman 2011 Nobel Prize). For AI methods, the perfect-crystal idealisation is the substrate of most training data (DFT calculations are typically run on idealised crystals), but real-world performance often depends on departures from perfection — a tension that Section 4 (defects) and Section 10 (the AI frontier) return to.

Materials science traditionally organises its subject matter into broad classes — metals, ceramics, polymers, semiconductors, composites — that share characteristic structural, bonding, and property profiles. The classification is imperfect (the boundaries blur, modern materials often combine features), but the categories carry substantial empirical regularities and recur throughout the field.

Metals

Metals are characterised by metallic bonding: delocalised electrons that move freely throughout the structure, producing high electrical and thermal conductivity, optical reflectivity at visible wavelengths, mechanical ductility, and the characteristic lustrous appearance. About three-quarters of the elements in the periodic table are metals. Pure metals are rarely used in engineering; alloys — solutions of multiple metals (or metals plus non-metallic elements) — dominate practical applications. Steels (Fe with C plus various alloying elements) are the most-used materials by mass in modern civilisation. Aluminium alloys (with Cu, Mg, Zn, Si) provide the strength-to-weight ratios essential for aerospace. Titanium alloys serve high-strength corrosion-resistant applications. Nickel-based superalloys dominate high-temperature applications (jet engines). Copper alloys (brass, bronze) are the substrate of much electrical and plumbing infrastructure.

Ceramics

Ceramics are inorganic non-metallic compounds, typically with ionic or covalent bonding. They are characteristically hard, brittle, electrically insulating, and resistant to high temperature and chemical attack. Traditional ceramics (clay, porcelain, brick) date to prehistory. Engineering ceramics include alumina (Al₂O₃, used as bearings, cutting tools, dental implants), silicon carbide (SiC, used in high-temperature electronics and abrasives), zirconia (ZrO₂, used as structural ceramic and oxygen sensor), silicon nitride (Si₃N₄, used in bearings and engine components), and various carbides and nitrides. Functional ceramics include the perovskite-structure ferroelectrics (BaTiO₃, PbZrTiO₃, used in capacitors and transducers), high-temperature superconductors (the cuprates discovered in 1986), and various oxide-electronics materials. Glasses are a closely-related category: amorphous oxides (silica-based glasses, fluoride glasses, chalcogenide glasses) with their own substantial application landscape.

Polymers

Polymers are long-chain molecules formed from repeating monomer units, typically with covalent bonding within chains and weaker (van der Waals or hydrogen-bond) interactions between chains. Thermoplastics (polyethylene, polypropylene, PET, polystyrene, nylon) soften when heated and can be reshaped repeatedly; they dominate consumer plastics. Thermosets (epoxies, polyurethanes, phenolics) cross-link irreversibly during curing and form rigid networks. Elastomers (rubbers, silicones) have the elastic-deformation behaviour of conventional polymers but with substantial reversible stretch — natural rubber stretches several hundred percent before failure. Engineering polymers (PEEK, PTFE, polycarbonate) span specialised high-performance applications. Biopolymers (cellulose, chitin, the various polypeptides covered in Ch 03) blur into the realm of biology. The polymer literature is enormous and the AI-for-polymers methodology has been substantially less mature than for crystalline materials, but is rapidly catching up.

Semiconductors

Semiconductors are materials whose electrical conductivity is intermediate between metals (high) and insulators (low), with the property that conductivity can be substantially modified by impurities (doping), temperature, or applied fields. The dominant semiconductor is silicon (Si), the substrate of essentially all modern electronics. Compound semiconductors (GaAs, GaN, InP, SiC, the various II-VI compounds) provide higher-performance alternatives for specific applications: GaN for power electronics and high-efficiency LEDs, GaAs for high-frequency electronics and laser diodes, SiC for high-temperature/high-voltage applications. 2D semiconductors (transition-metal dichalcogenides like MoS₂, plus the various 2D materials beyond graphene) are an active research frontier. Section 7's band theory provides the conceptual machinery for understanding what makes a semiconductor a semiconductor; this chapter's AI sections (§11–19) has substantial overlap with semiconductor design.

Composites and hierarchical materials

Composites combine multiple material types to achieve property combinations that no single material can. Fibre-reinforced composites (carbon-fibre-reinforced polymer, glass-fibre-reinforced polymer, the various aerospace composites) embed stiff fibres in a softer matrix to achieve high strength-to-weight ratios. Particle-reinforced composites (concrete, cermets) embed hard particles in a softer matrix. Functional composites include the various electromagnetic-shielding, thermal-management, and structural-electronic-multifunctional materials. Hierarchical materials (bone, wood, mollusc shells, the various biological-inspired materials) combine structure across multiple length scales to achieve property combinations that random structures cannot. The methodology of designing composites is substantially mature for some applications (aerospace) and immature for others (biological-inspired functional materials).

Soft and biological materials

The traditional metals/ceramics/polymers classification has gradually expanded to include soft matter (colloids, gels, liquid crystals, foams, granular materials) and biological materials (the various structural proteins, mineralised tissues, plant cell walls). These don't fit neatly into the traditional categories — many are composites of polymers and inorganic phases at multiple length scales — but they have substantial industrial and biomedical importance. AI methods for soft matter and biological materials are emerging research areas, with substantial overlap with the polymer-AI and biology-AI methodologies of Ch 03 and the recent generative-soft-matter literature.

Real materials are not perfect crystals. They contain defects at every length scale — missing atoms, atoms in wrong positions, line defects (dislocations), planar defects (grain boundaries, stacking faults), and bulk defects (voids, inclusions). The defects are not nuisances to be eliminated but the dominant determinants of mechanical, transport, and many other properties. Materials science is, in substantial measure, the science of defects.

Point defects

Point defects are zero-dimensional irregularities in the crystal. Vacancies are sites where an atom is missing; their concentration in equilibrium is determined by thermodynamics (vacancy formation energy and temperature). Interstitials are atoms in non-lattice positions, squeezed between regular sites. Substitutional defects are foreign atoms at lattice positions (this is the dominant defect type in most alloys — alloying elements are substitutional defects). Antisite defects in compound materials are atoms of one element on the other sublattice. Frenkel pairs (vacancy-interstitial pairs) and Schottky defects (cation-anion vacancy pairs in ionic crystals) are coupled-defect motifs. Point defects substantially affect ionic conductivity, optical properties, diffusion, and many other characteristics.

Dislocations and plastic deformation

The most-impactful defect type is the dislocation: a one-dimensional line defect where part of a crystal is shifted relative to the rest. Dislocations are the carriers of plastic deformation in metals — when metal yields, what's actually happening is dislocations propagating through the crystal, allowing one plane of atoms to slip past another at much lower stresses than would be required to break all the bonds simultaneously. The discovery (Orowan, Polanyi, Taylor 1934) resolved a substantial puzzle: theoretical strength predictions assuming defect-free crystals exceeded measured yield strengths by a factor of ~1,000. The dislocation-mediated picture has been spectacularly empirically confirmed and is the conceptual substrate for nearly all of mechanical metallurgy. Edge dislocations, screw dislocations, and mixed dislocations are the basic types; their interactions with each other, with grain boundaries, and with second-phase particles are what determine the strength of structural metals.

Grain boundaries and microstructure

Most engineering materials are polycrystalline: aggregates of many small crystallites (grains, typically µm to mm in size) separated by grain boundaries. Grain boundaries are 2D defects where the crystal orientation changes discontinuously. Their structure depends on the misorientation between adjacent grains; low-angle boundaries (≲15° misorientation) consist of arrays of dislocations; high-angle boundaries have more-complex atomic structure. Grain boundaries strongly affect mechanical properties (the Hall-Petch relationship: yield strength ∝ 1/√(grain size)), transport properties (grain boundaries scatter electrons and phonons), and many other characteristics. Microstructure — the detailed arrangement of grains, phases, and defects at micrometre scales — is one of the central concerns of metallurgy and materials engineering. Modern characterisation methods (Section 9, particularly EBSD and 3D X-ray methods) provide quantitative microstructural data that AI methods increasingly process.

Phase transformations

Materials can change phase — change crystal structure, change composition profile, develop precipitates — through various transformations. Diffusional transformations (precipitation hardening, spinodal decomposition) proceed by atomic diffusion at finite temperature. Diffusionless transformations (martensitic transformations in steel, the shape-memory transformations in NiTi) proceed by coordinated atomic shifts without long-range diffusion. Recrystallisation and grain growth happen when a heavily-deformed metal is annealed and replaces its strain-energy-rich microstructure with new strain-free grains. Each transformation type has its own kinetics and resulting microstructures, and the engineering control of transformations is what materials processing is fundamentally about.

Defects as functional features

Defects are not always to be minimised. Doping in semiconductors uses controlled substitutional defects (B in Si for p-type, P in Si for n-type) as the basis of all modern electronics. Vacancy ordering in oxides produces the ionic conductivity essential for solid-oxide fuel cells and oxygen sensors. Defect engineering in oxide thin films creates two-dimensional electron gases, ferroelectric polarisation, and various other functional features. Colour centres (single defects that absorb specific wavelengths) are the basis of nitrogen-vacancy centres in diamond used as quantum sensors. The methodology of designing materials by controlling defects is increasingly central to functional-materials development, and AI methods that handle defects beyond the perfect-crystal idealisation are an emerging research area.

The challenge for AI

A specific tension worth flagging: most ML-for-materials training data (Materials Project, OQMD, etc.) consists of perfect-crystal DFT calculations, but real-world properties depend substantially on defects. ML methods that operate purely on perfect-crystal data may produce predictions that don't transfer to defect-rich real materials. The 2024–2026 generation of methodology has begun to address this with explicit defect-handling methods, large-cell DFT calculations including representative defects, and hybrid approaches that combine perfect-crystal predictions with defect-energetics models. Section 10 returns to this challenge.

A phase diagram maps which crystal structures, microstructures, or compositions are thermodynamically stable as a function of temperature, pressure, and composition. Phase diagrams are the core organising tool of metallurgy, materials processing, and computational materials thermodynamics — and they are the substrate of much AI-for-materials work.

Phases and components

A phase is a region of material with uniform structure and composition. Pure water at room temperature is one phase; ice and water in equilibrium are two phases (same composition, different structure); a salt-water solution is one phase; a salt-water solution with undissolved salt at the bottom is two phases. Components are the chemical species that can vary independently in composition. Pure water is a one-component system; salt water is two components (water and salt); steel is at least two components (iron and carbon, plus various alloying elements). The phase rule (Gibbs, 1878): F = C − P + 2, where F is the degrees of freedom (number of intensive variables that can vary independently), C is the number of components, and P is the number of phases. The rule constrains how many phases can coexist at given conditions and is the substrate of all phase-diagram interpretation.

Single-component phase diagrams

The simplest phase diagrams are single-component, plotting pressure vs temperature for a pure substance and showing the regions where solid, liquid, and gas (and various distinct solid phases) are stable. The water phase diagram shows the liquid-gas boiling curve, the solid-liquid melting curve, and the solid-gas sublimation curve, all meeting at the triple point (273.16 K, 611 Pa) and ending at the critical point (647 K, 22 MPa) above which liquid and gas become indistinguishable. Ice itself has many crystal phases — ice Ih (ordinary ice), ice II, ice III, etc. — that appear at high pressures, with their own phase boundaries. Carbon's phase diagram has graphite stable at low pressure and diamond stable at high pressure, with the boundary intersecting room temperature at ~5 GPa. The vocabulary (triple point, critical point, sublimation, allotropy) is essential.

Binary phase diagrams

The most-used phase diagrams in metallurgy are binary: composition-vs-temperature for two-component systems at fixed pressure (typically 1 atm). Reading binary phase diagrams is a substantial skill: the liquidus and solidus curves bound the all-liquid and all-solid regions; eutectic points are compositions where liquid solidifies into a mixture of two solid phases at a single temperature; peritectic reactions are L + α → β reactions on cooling; eutectoid reactions (γ → α + Fe₃C in steel) are the all-solid analogue of eutectic. The lever rule uses geometric reasoning on the phase diagram to compute phase fractions at any temperature. Specific binary diagrams worth recognising: Fe-C (iron-carbon, the substrate of all steel metallurgy), Cu-Zn (brass), Al-Cu (precipitation-hardenable aluminium alloys), Cu-Ni (the textbook example of complete solid solubility).

Ternary and higher-order systems

Ternary phase diagrams handle three components and are visualised on triangular composition diagrams (typically isothermal cuts at fixed temperature). The complexity grows quickly: a binary diagram is 2D (T vs composition), a ternary at fixed T is 2D (a triangle in composition space), and a full ternary across temperature is 3D (a triangular prism). Modern alloys are typically four or more components (most superalloys have 8+, modern high-entropy alloys have 5+ in roughly equal proportions), and visualising the relevant phase diagrams is intractable directly. The methodology of CALPHAD (CALculation of PHAse Diagrams) addresses this through computational thermodynamics: model the Gibbs free energy of each phase as a function of composition and temperature, then compute equilibrium phase fractions by minimisation. Modern CALPHAD databases (TCNI for nickel-base superalloys, TCAL for aluminium alloys, etc.) cover hundreds of components and thousands of phases, and they are the substrate of much industrial alloy design.

Gibbs free energy and stability

The thermodynamic basis of phase diagrams is Gibbs free-energy minimisation (Ch 07 §4): at constant temperature and pressure, equilibrium occurs at the minimum of G = H − TS where H is enthalpy, T is temperature, and S is entropy. Each phase has its own G(T,P,composition), and the equilibrium phase distribution at given conditions is the one that minimises total G of the system. The common-tangent construction identifies coexisting phases as having equal slopes of their G-vs-composition curves; the resulting compositions are the equilibrium phase compositions. The methodology connects directly to Ch 02 §6's chemical-thermodynamics machinery and to the statistical-mechanics framework of Ch 07 §4. ML methods for free-energy prediction are increasingly central to computational materials thermodynamics.

Metastability and kinetics

Phase diagrams describe equilibrium, but real materials often persist in metastable states for engineering-relevant timescales. Diamond is metastable at room temperature and pressure (graphite is the equilibrium phase) but the kinetic barrier to graphite formation is high enough that diamonds survive billions of years. Quenched steel is metastable martensite, kinetically prevented from transforming to the equilibrium ferrite + cementite. Glass is supercooled liquid that has fallen out of equilibrium during cooling. Engineering materials processing fundamentally exploits metastability: heat treatments, quenching, and various other processing steps produce metastable structures with engineering properties that the equilibrium phases cannot match. Modeling metastability requires kinetic modelling beyond pure thermodynamics, and AI methods for kinetic processes (nucleation rates, growth rates, transformation kinetics) are active research areas.

Materials properties are how materials manifest in engineering applications. This section surveys the major property classes — mechanical, thermal, and transport (electrical, thermal, ionic) — that recur across most materials-AI applications and that bridge the structural vocabulary of the previous sections to engineering use.

Elastic and plastic mechanical response

The mechanical response of a material to applied stress is conventionally divided into two regimes. The elastic regime is reversible: removing the stress returns the material to its original shape. The relationship is captured by the Young's modulus E (stress per unit strain in uniaxial tension), the shear modulus G, the bulk modulus K, and Poisson's ratio ν. The four are interrelated for isotropic materials (only two are independent); for anisotropic crystals, the full elastic tensor has 21 independent components in the most-general case (3 for cubic crystals, 5 for hexagonal). The plastic regime is permanent: the material yields, deforming irreversibly. The transition point — the yield strength — is one of the most-important engineering properties, and is determined by dislocation dynamics (Section 4) rather than by clean continuum elasticity. The ultimate tensile strength is the maximum stress before fracture; ductility is how much plastic strain a material can absorb before fracture.

Hardness and toughness

Hardness measures resistance to surface deformation; the various hardness scales (Vickers, Rockwell, Brinell, Mohs) measure it different ways but agree qualitatively. Hardness correlates roughly with yield strength but is much easier to measure on small samples or thin films. Toughness measures energy absorbed before fracture and is essentially independent of hardness — there exist hard-but-brittle materials (ceramics, glass) and soft-but-tough materials (rubbers). The combination of hardness and toughness is what most engineering applications need; achieving both simultaneously is one of the long-standing challenges in materials design. Fracture toughness (K_IC) is the quantitative measure of resistance to crack propagation and is particularly important for structural applications.

Thermal properties

Thermal expansion describes how much a material expands per unit temperature change; mismatched thermal expansion between joined materials (different metals in a thermostat, different layers in microelectronic packaging) creates thermal stresses that drive substantial engineering design. Thermal conductivity describes heat flow per unit temperature gradient; metals are good thermal conductors (because the same delocalised electrons that carry charge also carry heat), ceramics and polymers are typically poor. Heat capacity describes how much energy is needed to raise the temperature of a unit mass; the Debye model from Ch 07 §4's statistical mechanics gives the temperature dependence at low temperatures. Phase-change properties (latent heats of melting, boiling, solid-solid transitions) are essential for thermal management in everything from food storage to advanced heat exchangers.

Electrical conductivity and the metal-insulator distinction

The simplest property classification of materials by electronic behaviour: conductors (mostly metals) have high electrical conductivity; insulators have very low; semiconductors are intermediate, with conductivity that can be substantially modified by doping or applied fields. The conductivity spans 25 orders of magnitude across this range — from copper (~6 × 10⁷ S/m) to fused silica (~10⁻¹⁸ S/m). The deeper picture comes from band theory (Section 7), which explains the differences in terms of how electron states fill the available energy levels. Superconductivity — exactly zero electrical resistance below a critical temperature — is a distinct regime requiring quantum-mechanical phenomena beyond simple band theory. The 1986 discovery of high-temperature superconductors (cuprates, T_c above liquid-nitrogen temperature) was a substantial empirical surprise; the theory of high-T_c superconductivity remains incomplete despite decades of work, and AI methods for predicting candidate superconductors are an active frontier.

Diffusion and ionic transport

Diffusion describes how atoms move through a material in response to concentration gradients. Fick's laws describe the macroscopic phenomenology; the Arrhenius temperature dependence reflects the underlying activation energies. Diffusion is essential for nearly all materials processing — phase transformations, precipitation, segregation, sintering. Ionic conductivity is the analogous transport for charged ions; ionic conductors are essential for batteries (lithium-ion, beyond-lithium), fuel cells (the various solid-oxide systems), and oxygen sensors. The methodology of designing fast ionic conductors is increasingly AI-augmented — the methodology of predicting ionic-mobility from crystal structure connects directly to the materials-AI work in §11–19.

Optical and magnetic properties

Materials' interaction with light and with magnetic fields produces additional property classes. Optical properties (absorption, reflection, refraction, photoluminescence) determine appearance and many functional applications (LEDs, lasers, photovoltaics, optical sensors). Magnetic properties (paramagnetism, ferromagnetism, antiferromagnetism, ferrimagnetism) underlie permanent magnets, magnetic-recording media, magnetic resonance imaging, and various sensor applications. Both depend on electronic structure, with quantum mechanical foundations developed in Ch 07 §7 and extended to solids by the band theory of Section 7. AI methods for predicting these properties from crystal structure are increasingly mature and are central to functional-materials discovery.

Band theory is the framework for understanding electrons in crystalline solids. It explains why some materials are metals and others insulators, why semiconductors can be doped to tune their conductivity, why optical absorption begins sharply at characteristic wavelengths, and many other observations. The conceptual machinery is essential for any AI reader engaging with electronic-materials applications.

Electrons in periodic potentials

Electrons in a crystal experience a periodic potential: the potential energy as a function of position has the same periodicity as the crystal lattice. The Schrödinger equation for an electron in such a potential has solutions of a specific form, given by Bloch's theorem: ψ_k(r) = u_k(r) exp(ik·r), where u_k(r) has the same periodicity as the lattice and k is a wavevector that labels the state. The wavevector k lives in the Brillouin zone (Section 2), and each value of k gives an infinite ladder of energy eigenstates ε_n(k) labelled by the band index n. The full band structure ε_n(k) — energy as a function of wavevector for each band — is the central object of solid-state electronic-structure theory.

Bands, gaps, and the Fermi level

The states fill in order of increasing energy, two electrons (spin up and spin down) per state. The energy below which all states are filled at zero temperature is the Fermi level ε_F. What matters for electronic behaviour is where ε_F falls in the band structure. Three cases:

Metals: ε_F falls in the middle of a band, with partially-filled states immediately above and below. Electrons can easily move into nearby unoccupied states under applied fields, producing high conductivity. Insulators: ε_F falls in a band gap — an energy range with no allowed electronic states — between completely-filled bands (the valence band) and completely-empty bands (the conduction band). Electrons cannot conduct unless thermally excited across the gap, which is impractical for typical room temperature in materials with band gaps above ~3 eV. Semiconductors: same band-gap structure as insulators but with smaller gaps (~0.5–3 eV), so thermal excitation produces measurable carrier concentrations even at room temperature, and doping or applied fields can substantially modify conductivity.

The semiconductor framework

Semiconductors get extended treatment because they are the substrate of modern electronics. Pure (intrinsic) semiconductors have small thermal carrier populations: electrons promoted across the gap leave behind holes (empty states in the valence band that behave like positive charge carriers). Doping adds impurities that contribute extra carriers: n-type doping (P or As in Si, group-V donors) adds extra electrons; p-type doping (B in Si, group-III acceptors) adds extra holes. The resulting p-n junctions are the building blocks of all modern electronics — diodes, transistors, photovoltaics, light-emitting diodes. The detailed band structure of specific semiconductors (Si, Ge, GaAs, GaN, etc.) including their band gaps, effective masses, and direct-vs-indirect transitions, determines the device physics.

Direct vs indirect band gaps

A specific structural feature: the maxima of the valence band and minima of the conduction band may or may not occur at the same k value in the Brillouin zone. Direct band gaps have valence maximum and conduction minimum at the same k (typically k = 0, the Γ point). Indirect band gaps have them at different k values. The distinction matters substantially for optical applications: photons carry essentially zero momentum on the scale of the Brillouin zone, so direct-gap absorption and emission are efficient (silicon's indirect gap is why silicon LEDs are inefficient and why GaAs is used for laser diodes). The methodology of predicting direct vs indirect from crystal structure is an active AI-for-materials application.

Beyond simple band theory

Simple band theory describes non-interacting electrons: each electron sees the average potential from all others but doesn't interact with them directly. Many materials behave well in this approximation; many do not. Strongly correlated electron systems — materials where electron-electron interactions dominate — include high-temperature superconductors, many transition-metal oxides, and various other functional materials. Mott insulators are materials that band theory would predict are metals but which are insulators because of strong correlations; their understanding requires extensions like the Hubbard model. Topological materials — materials with topologically non-trivial band structures — include topological insulators (insulating in bulk but conducting on the surface) and Weyl semimetals (with band-touching points that act like relativistic Weyl fermions). The 2016 Nobel Prize recognised the topological-materials theoretical work; the experimental landscape continues to evolve. AI methods that engage with electronic structure beyond simple band theory are an active research direction.

Computing band structures

Practical band-structure calculations almost always use density functional theory (Section 8) plus various extensions. The output of a DFT calculation is the band structure ε_n(k) along high-symmetry lines in the Brillouin zone, plus various derived quantities: the density of states (states per unit energy), the Fermi surface (the locus of states at ε_F in metals), the effective masses (curvatures of bands at extrema, governing carrier dynamics), and the band gaps. The standard DFT functionals systematically underestimate band gaps; more-sophisticated methods (hybrid functionals, GW approximation) provide better agreement at higher computational cost. AI methods for fast band-gap prediction from crystal structure are increasingly central to high-throughput materials screening.

Density functional theory (DFT) is the dominant computational method for the electronic structure of materials. It is the substrate of the Materials Project and similar databases, the basis of most computational materials-design workflows, and the immediate predecessor of the machine-learned interatomic potentials that §11–19 develop. Understanding DFT at a working-vocabulary level is essential for any AI reader engaging with materials science.

The motivating problem

The many-electron Schrödinger equation for N interacting electrons in a fixed nuclear potential is exact but intractable: it requires storing the wavefunction Ψ(r₁, r₂, ..., r_N), which lives in 3N-dimensional space. For N = 100 electrons (a small molecule or unit cell), this is impossibly large. Wavefunction-based methods (Hartree-Fock plus various corrections — coupled-cluster, configuration interaction, etc.) treat this approximately and can be very accurate but scale poorly with system size: ~N⁴ for Hartree-Fock, ~N⁷ for high-accuracy coupled cluster. Practical materials applications need ~100s to ~1000s of atoms, beyond the reach of wavefunction methods.

The Hohenberg-Kohn theorems

The conceptual breakthrough that makes DFT possible is the Hohenberg-Kohn theorems (1964): the ground-state electron density n(r) — a function of just three coordinates rather than 3N — uniquely determines all ground-state properties, and the ground-state density minimises a universal energy functional E[n]. This is a remarkable result: the high-dimensional wavefunction problem reduces in principle to a low-dimensional density problem. The theorems prove that the reduction is possible; they do not give the energy functional. Finding the right functional remains the central challenge.

The Kohn-Sham approach

The practical implementation is the Kohn-Sham equations (Kohn & Sham 1965): replace the interacting-electron problem with a fictitious non-interacting-electron problem that has the same density. The fictitious electrons satisfy single-particle Schrödinger-like equations with an effective potential that includes the external potential, the classical electron-electron Coulomb repulsion (the Hartree potential), and an exchange-correlation potential that captures all the quantum-mechanical many-body effects. The exchange-correlation potential is the only unknown — and it is where all the difficulty of the many-body problem now lives. The methodology converts the interacting problem to a single-particle problem at the cost of needing an approximate exchange-correlation functional.

Exchange-correlation functionals

The choice of exchange-correlation functional determines DFT's accuracy and computational cost. Local Density Approximation (LDA): the simplest, treats the exchange-correlation energy as a local function of density, calibrated against the homogeneous electron gas. Surprisingly accurate for many properties; fails for systems with strong density gradients. Generalised Gradient Approximation (GGA): adds gradient corrections; the dominant family includes PBE (Perdew-Burke-Ernzerhof, 1996, the most-cited GGA). Hybrid functionals (B3LYP, HSE06): mix exact Hartree-Fock exchange with GGA exchange; substantially more accurate for band gaps and chemical reaction barriers but ~10× slower. Meta-GGAs (TPSS, SCAN): include kinetic-energy-density information; current-generation accuracy targets. Range-separated and double-hybrid functionals push further but at substantial computational cost. The methodology is a hierarchy of approximations, with users choosing based on required accuracy and available computational resources.

Pseudopotentials and basis sets

Two practical ingredients matter substantially for DFT calculations. Pseudopotentials replace the strong nuclear potential and tightly-bound core electrons with smoother effective potentials representing valence electrons; this dramatically reduces the computational cost. Common pseudopotential frameworks include PAW (Projector Augmented Wave), norm-conserving, and ultrasoft. Basis sets represent the wavefunctions as combinations of basis functions: plane waves (the natural choice for periodic systems, used in VASP, Quantum ESPRESSO, ABINIT), localised orbitals (Gaussian-type orbitals dominate quantum chemistry; numerical orbitals appear in some codes), or all-electron methods (LAPW, FLAPW). The choice of pseudopotential and basis-set substantially affects both accuracy and computational cost.

What DFT does and doesn't do well

DFT's empirical track record is substantial. It reliably predicts crystal structures at low temperature within ~1% of lattice parameters. It computes total energies with chemical accuracy (~kcal/mol) for many systems. It predicts elastic constants within ~10% for most materials. It captures band-structure features at the qualitative level for most semiconductors and insulators. But it has known systematic problems: it underestimates band gaps (typically by ~50% with GGA functionals); it misses van der Waals interactions (without dispersion corrections); it fails for strongly-correlated systems where standard functionals are insufficient. Modern functional development (the various 2020s SCAN-family and r²SCAN-family meta-GGAs) addresses some of these issues. AI-augmented DFT — using ML methods to learn corrections to DFT results, train better functionals, or replace DFT entirely with ML interatomic potentials (§13–14) — is one of the most-active frontiers.

Production codes and the ecosystem

The dominant DFT codes worth knowing: VASP (Vienna Ab initio Simulation Package, the most-used commercial code), Quantum ESPRESSO (open-source, plane-wave plus pseudopotential), ABINIT (open-source), CP2K (open-source, mixed Gaussian-and-plane-wave), FHI-aims (open-source, all-electron), and the various commercial-and-academic alternatives. The output of these codes feeds the major materials databases (Materials Project, OQMD, AFLOW) and constitutes the substrate of much of AI for materials. Production workflows use Materials Project's pymatgen, ASE (Atomic Simulation Environment), or the various other Python ecosystems for high-throughput automation. The methodology has matured to the point that DFT calculations are increasingly automated, with quality control, error catching, and database insertion handled by sophisticated workflow infrastructure.

Materials characterisation is the experimental side of materials science: the techniques that produce the data on which both theoretical work and AI methods rest. This section surveys the major techniques for AI-relevant characterisation across structural, microstructural, chemical, and electronic-property measurements.

X-ray diffraction

X-ray diffraction (XRD) has been the workhorse structural-characterisation technique since the 1910s. The methodology: shine an X-ray beam at a crystalline sample, record the diffraction pattern produced as the X-rays scatter off planes of atoms, and analyse the pattern to determine the crystal structure (lattice parameters, space group, atomic positions). Powder XRD averages over many crystallite orientations, producing 1D patterns of intensity vs scattering angle that fingerprint specific phases. Single-crystal XRD uses an oriented single crystal to fully determine atomic positions and is the canonical method for structure determination. Modern synchrotron sources enable sub-millisecond time resolution and atomic-scale spatial resolution; XRD-derived databases (ICSD with ~270K entries, COD with ~500K) are central to materials informatics. ML methods for XRD pattern analysis (phase identification, quantitative phase analysis, structure refinement) are increasingly deployed in production.

Electron microscopy

Scanning Electron Microscopy (SEM) images sample surfaces using a focused electron beam; provides topographic and compositional contrast at nanometre to micrometre resolution. Transmission Electron Microscopy (TEM) shoots electrons through thin samples; provides atomic-resolution imaging in modern aberration-corrected instruments and is the gold standard for direct observation of atomic structure, defects, and interfaces. Scanning Transmission Electron Microscopy (STEM) combines elements of both with scanning capability. Modern instruments include detectors for various analytical signals: EDX (Energy-Dispersive X-ray spectroscopy) for chemical composition, EELS (Electron Energy-Loss Spectroscopy) for electronic structure and chemical bonding, and EBSD (Electron Backscatter Diffraction) for crystallographic orientation mapping. The data volumes are enormous (modern STEM produces tens of terabytes per session) and AI methods are increasingly central to automated analysis.

X-ray photoelectron spectroscopy

XPS (X-ray Photoelectron Spectroscopy) measures the kinetic energies of photoelectrons emitted when X-rays strike a sample, providing surface-sensitive information about chemical composition and oxidation states. The methodology produces characteristic peaks at energies that fingerprint specific elements and bonding environments — Si in silicon dioxide vs Si in silicon nitride vs metallic Si all produce distinguishable peaks. XPS is essential for surface science, catalysis, and thin-film characterisation. ML methods for XPS spectral analysis (peak identification, deconvolution, background subtraction) are increasingly deployed.

Neutron scattering

Neutron scattering complements X-ray techniques in important ways. Neutrons interact with nuclei rather than electrons, making them sensitive to light elements (hydrogen, lithium) that X-rays barely see. Neutrons have spin and so probe magnetic structure directly. Neutron sources are typically large facilities (reactors at ILL Grenoble, Oak Ridge National Lab; spallation sources at SNS Oak Ridge, ISIS in the UK, J-PARC in Japan, ESS Sweden under construction), and the data is correspondingly precious and well-curated. AI methods for neutron-scattering data analysis are an active area, particularly for inelastic scattering (which probes phonon dispersions and magnetic excitations) where the signal-to-noise is challenging.

Mechanical and thermal testing

Engineering-property characterisation includes tensile testing (stress-strain curves, yield strength, ultimate strength, ductility), hardness testing (Vickers, Rockwell, etc.), thermal analysis (DSC for differential scanning calorimetry, TGA for thermogravimetric analysis), thermal expansion measurement, and various others. The methodology is mature and well-standardised (ASTM and ISO standards specify procedures); modern instruments produce well-curated digital outputs that integrate naturally with data-driven workflows.

Atom probe and 3D characterisation

Specific frontier techniques worth knowing. Atom probe tomography (APT) provides 3D atomic-resolution composition maps over volumes of ~10⁵ atoms, with single-atom chemical identification. 3D X-ray imaging at synchrotrons provides micrometre-resolution 3D microstructural data without sectioning. Cryogenic electron tomography extends TEM to 3D, with substantial recent advances in resolution and throughput. The combined "3D characterisation" toolkit produces data volumes that overwhelm traditional analysis methods, and ML methods are increasingly central — segmentation of atomic features in 3D APT data, automated phase identification in 3D XRT, and the various other applications.

Open data and databases

Modern materials characterisation increasingly produces openly-shared data. ICSD for crystal structures, CSD for organic crystals, Crystallography Open Database for both, RIDA and JARVIS-DFT for computed properties, NIST property databases for engineering data — together they provide the empirical substrate that ML methods consume. The combined open-data ecosystem has substantially democratised materials informatics over the past decade, and the methodology is increasingly accessible to academic and small-industry researchers without proprietary database access.

The previous nine sections established the materials-science vocabulary: crystal structures and symmetry, material classes, defects and microstructure, phase diagrams, mechanical and transport properties, band theory, density-functional theory, and characterisation techniques. This section is the bridge to the methodology that follows. Materials AI has produced one of the cleanest empirical records of any AI-for-Science domain, but the methodological landscape has its own distinctive properties — an unusually-mature data infrastructure (Materials Project, OQMD, AFLOW, NOMAD), well-defined prediction targets, substantial economic incentives, and a structure-encoding problem that the field has progressed through several methodological generations to solve. This section orients the ML practitioner; Sections 11–19 develop the methods within that frame.

The data substrate, from an ML perspective

Materials AI inherits an unusually-clean training-data substrate. Materials Project (Berkeley, ~150K materials with DFT-computed structures, formation energies, band gaps, elastic moduli, X-ray diffraction patterns, and various other properties) is the most-used database. OQMD (Open Quantum Materials Database, Northwestern, ~1M entries) provides higher coverage but somewhat less curation. AFLOW (Duke, ~3M entries) has the largest catalog. JARVIS-DFT (NIST) integrates DFT-computed properties with experimental measurements where available. NOMAD (Berlin) is the European federation point. ICSD and the Cambridge Structural Database provide hundreds of thousands of experimentally-determined structures. The combined substrate is among the largest, cleanest, most-organised in any AI-for-Science domain, and most successful methods build on it rather than collecting bespoke training data.

The structure-encoding problem

Representing crystal structures for ML methods is non-trivial. Crystals are infinite: the unit cell is repeated periodically in three dimensions, so any representation must respect this periodicity. Crystals have explicit symmetries: rotations and translations that leave the structure invariant. Crystals are heterogeneous: different atomic species with different properties at different positions. The methodology has matured through several generations of representations — composition-only descriptors (just elemental proportions), hand-crafted structural features (radial distribution functions, bond-angle distributions), graph-based molecular representations adapted for crystals, and increasingly equivariant graph neural networks that respect the periodic-and-symmetric structure by construction. Section 2 develops representations in detail; the choice substantially shapes what kinds of methods work.

The simulation-vs-experiment validation gap

A specific tension shapes the field. Most ML training data is DFT-computed: the methodology produces predictions calibrated against DFT, not against experiment. DFT itself has known systematic errors — band gaps are typically underestimated by ~50% with standard GGA functionals, formation energies have ~0.1 eV/atom uncertainty, and various edge cases produce larger errors. ML methods trained on DFT inherit DFT's biases. Experimental data is sparse, noisy, and expensive to acquire, and the measured quantity is often not exactly what DFT predicts (real materials have defects, grain boundaries, finite temperature, processing-history-dependent microstructures). The methodology of bridging the simulation-experiment gap — multifidelity training, transfer learning from DFT to experiment, careful uncertainty quantification — is one of the central methodological challenges. Section 10 returns to this.

Symmetry as architectural prior

Crystals have explicit symmetries: 230 space groups in three dimensions, plus the lattice translation group. Equivariant neural networks respect these symmetries by construction, producing predictions that transform appropriately under symmetry operations. The methodology connects to the broader equivariance machinery (Ch 01 §8, Ch 03 §11–12 for proteins) and has been substantially developed for materials applications. The dominant modern MLIPs (NequIP, MACE) are E(3)-equivariant: rotations and reflections in 3D space act on the network's internal representations the way they act on physical positions and forces. The architectural prior produces meaningful empirical improvements — equivariant networks consistently outperform non-equivariant baselines at fixed parameter counts and data budgets.

The economic stakes

Unlike many AI domains, materials AI has direct industrial applications. Battery materials: improvements in cathode materials (higher capacity, longer cycle life, lower cost) are worth billions in market value; ML-driven candidate discovery is increasingly central at major battery companies. Permanent magnets: rare-earth-free or rare-earth-light magnet candidates are the substrate of national-security and supply-chain concerns. Semiconductors: discovering new wide-bandgap semiconductors, photonic materials, and quantum-device materials has substantial industrial implications. Catalysts: AI-driven catalyst discovery for hydrogen production, CO₂ reduction, and the various energy applications has both economic and climate-policy implications. Structural alloys: the ICME (Integrated Computational Materials Engineering) initiatives at major aerospace and automotive companies are increasingly AI-augmented. The methodology in this chapter is shaped by these stakes.

The discovery-vs-prediction spectrum

Like AI for physics (Ch 08), materials AI spans a spectrum from prediction within known frameworks (faster property prediction, ML-replacement of expensive simulations) to discovery of new materials (generative design of unprecedented compositions and structures). Most production deployments live near the prediction end, where empirical wins are clearer and methodology more mature. Discovery-end applications — A-Lab's autonomous synthesis of novel oxides, MatterGen's generation of property-conditioned crystal candidates, the various "ML found a new battery" announcements — are higher-stakes and require more-careful evaluation methodology. The empirical record on discovery-end successes is mixed: many papers claim materials wins that don't reproduce, while others have produced concrete improvements that have entered industrial pipelines. Section 10 returns to this evaluation challenge.

Every materials-AI method begins with a representation of the crystal structure. The choice of representation substantially constrains what the model can learn and how it generalises. Several representations dominate the field, each with characteristic strengths.

Composition-only descriptors

The simplest representation discards structural information entirely and uses only the elemental composition. Magpie (Materials Agnostic Platform for Informatics and Exploration, Ward et al. 2016) computes feature vectors from elemental properties (atomic number, electronegativity, melting point, etc.) and statistics over composition. Roost (Goodall & Lee 2020) uses a graph-based representation where each element in the formula is a node and message-passing produces composition embeddings. The methodology is fast and works surprisingly well for many properties (formation energy correlates substantially with composition, even before structure is considered), but cannot distinguish polymorphs (compositions with multiple stable structures) and cannot capture structure-sensitive properties like band gaps directly. Composition-only methods serve as strong baselines and are deployed in production for high-throughput initial screening before more-expensive structural methods.

Hand-crafted structural descriptors

Pre-deep-learning methods used hand-crafted features computed from the crystal structure. Smooth Overlap of Atomic Positions (SOAP, Bartók et al. 2013) represents the local atomic environment as a Gaussian-smeared density expanded in spherical harmonics; the resulting feature vector is invariant to rotations and translations. Behler-Parrinello symmetry functions (2007) use radial and angular distribution functions computed in atom-centred local environments. Coulomb matrices represent the molecule or unit cell through pairwise charge-distance matrices. These descriptors were the substrate of the first wave of materials ML and remain useful baselines; they have been substantially superseded by graph-based methods for most applications since 2018.

Crystal graphs

The dominant modern representation is the crystal graph: nodes are atoms (with element, position, and various other attributes), edges connect atoms within a cutoff radius (with bond-length and angle attributes), and the periodic boundary conditions are explicitly handled by edges that connect across unit-cell boundaries. CGCNN (Crystal Graph Convolutional Neural Network, Xie & Grossman 2018) was the foundational paper: a GNN-based architecture with convolutional message-passing on crystal graphs, achieving competitive results for formation energy and band gap prediction on the Materials Project dataset. MEGNet (Materials Graph Network, Chen et al. 2019) extended the methodology with global state attributes (temperature, pressure) and substantially-improved performance. The crystal-graph paradigm has dominated property prediction since 2018, with successive generations refining the architecture rather than replacing the representation.

Equivariant graph representations

The 2022 wave introduced equivariant crystal graphs: graph representations where node and edge features carry explicit geometric information (vectors, tensors) that transforms appropriately under rotation. NequIP (Batzner et al. 2022) was the foundational paper: an E(3)-equivariant message-passing architecture where intermediate features include not just scalars but also vectors and higher-order tensors, with operations that respect rotational symmetry by construction. The empirical case is substantial: equivariant architectures consistently outperform non-equivariant baselines at fixed parameter and data budgets, particularly for force predictions and energy gradients (where rotational symmetry matters most). Subsequent work — MACE, Allegro, EquiformerV2 — has continued the trajectory. Section 9 develops the equivariance machinery in detail.

Periodic boundary conditions

A specific structural challenge that distinguishes crystal representations from molecular representations is periodic boundary conditions. A molecule is finite; a crystal is infinite (or, more pragmatically, the unit cell repeats periodically forever). The representation must explicitly handle this: when computing edges within a cutoff radius, an atom near the unit-cell boundary has neighbours both within the same cell and in adjacent cells. The methodology of computing periodic neighbour lists is straightforward (handled by ASE's neighbour-list utilities, pymatgen's local environment finders, etc.) but the architecture must be designed to handle the resulting graphs without artefacts. Modern equivariant GNNs handle PBCs natively; older architectures sometimes had subtle bugs at the periodic boundary.

3D point-cloud and continuous-position representations

Some methods bypass discrete graph structures entirely and treat the crystal as a continuous point cloud. SchNet (Schütt et al. 2017) uses continuous-filter convolutions on atom positions, with no explicit graph but with effective neighbourhoods determined by distance-based filters. DimeNet, PaiNN, and the various 2021–2022 successors use related continuous-position approaches with increasingly sophisticated handling of angular information. The point-cloud paradigm has substantial overlap with molecular ML methodology (Ch 03 §11–12 for proteins) and is increasingly common for molecular-scale crystals (defects, surface adsorption, low-dimensional materials).

Choosing a representation

The choice of representation depends on the task. High-throughput initial screening: composition-only descriptors (Magpie, Roost) are fast and adequate. Standard property prediction: crystal graphs (CGCNN, MEGNet, ALIGNN) work well. Force-field and dynamics: equivariant graph nets (NequIP, MACE) are essential for accurate forces. Generative design: continuous representations (CDVAE, DiffCSP) are increasingly used, though discrete-graph variants exist. The methodology has been substantially organised around these choices, and the empirical landscape is mature enough that representation selection is an established design decision.

Property prediction — given a crystal structure, predict a target property — is the most-mature AI-for-materials application. The methodology has been deployed in high-throughput screening pipelines at major industrial labs and academic groups, with substantial empirical wins on standard benchmarks.

The benchmark landscape

The community-standard benchmarks anchor empirical evaluation. Matbench (Dunn et al. 2020) provides 13 standardised tasks spanning formation energy, band gap, elastic moduli, dielectric constants, refractive indices, perovskite stability, and various other endpoints, with explicit train-test splits and evaluation metrics. Materials Project benchmarks are widely used for individual-property comparisons. JARVIS-Leaderboard aggregates results across methods and tasks. The 2022–2026 generation of methods is routinely benchmarked across multiple of these standards, with the empirical landscape substantially organised. Modern methods reach mean-absolute-errors of ~25 meV/atom on formation energy and ~0.3 eV on band gap (vs ~0.4–0.7 eV typical DFT-vs-experiment uncertainty), making them competitive with the underlying DFT calculations.

The CGCNN-MEGNet-ALIGNN trajectory

CGCNN (Xie & Grossman 2018) established the crystal-graph-convolutional methodology and remains the most-cited materials-ML paper. MEGNet (Chen et al. 2019) extended with global state variables and improved empirical performance. ALIGNN (Atomistic Line Graph Neural Network, Choudhary & DeCost 2021) added explicit angular information through a line graph (where edges of the original graph become nodes of a "bond graph"), substantially improving performance for properties sensitive to local geometry. Each generation reduced errors by ~10–20% on standard benchmarks, with the trajectory plateauing around 2022–2023 as architectural improvements yielded diminishing returns. The 2024–2026 generation has shifted attention from architecture refinement to data-quality improvements, multi-task training, and equivariance.

Multi-task and transfer learning

A specific methodological win is multi-task training: train a single network to predict multiple properties simultaneously, sharing representations across tasks. The empirical result is that auxiliary tasks (predicting density, predicting elastic moduli) often improve the primary task (predicting formation energy) by inducing better internal representations. Transfer learning from large databases to specific application domains is similarly productive: pretrain on the Materials Project, fine-tune on the smaller specialised datasets relevant for batteries, magnets, or photovoltaics. The methodology has been substantially developed and is increasingly the default approach for production materials-AI pipelines.

The DFT-vs-experiment problem

A persistent challenge is the gap between DFT-computed properties and experimental measurements. Band gaps: standard GGA-DFT underestimates band gaps by ~50%; ML methods trained on GGA data inherit this systematic error and predict GGA-style band gaps, not experimental ones. Formation energies: DFT typically agrees with experiment to ~0.1 eV/atom but with systematic biases for specific chemistries (transition-metal oxides, van der Waals systems). Elastic moduli: DFT typically overestimates elastic constants by ~10–20%. The methodology of bridging the gap includes multifidelity learning (training on both DFT and experimental data with explicit fidelity labels), Δ-learning (training to predict corrections to DFT rather than absolute properties), and post-DFT functionals (training on hybrid-functional or GW-corrected data when available, even if expensive). The empirical case for each approach is mixed; production deployments increasingly combine multiple strategies.

Uncertainty quantification

Production materials-AI pipelines need calibrated uncertainty estimates: a candidate material with predicted band gap 1.5 ± 0.5 eV is much more useful than one with predicted band gap 1.5 eV (no uncertainty). Standard ML uncertainty methods — ensembles, Monte Carlo dropout, mean-variance networks, evidential deep learning — have been adapted for materials applications with substantial empirical refinement. The methodology of selecting candidates for experimental validation depends sensitively on uncertainty: high-mean candidates with small uncertainty are obvious successes; high-mean with large uncertainty are exploration targets; low-mean with small uncertainty are obvious rejects. Modern active-learning pipelines use the uncertainty estimates explicitly, and Section 7's autonomous-laboratory methodology depends on them.

Beyond simple property prediction

Property prediction has expanded substantially beyond simple scalar targets. Phonon dispersions: predicting the full vibrational spectrum from crystal structure, useful for thermal conductivity and lattice dynamics. Optical absorption spectra: predicting wavelength-dependent absorption for photovoltaic and photonic applications. Phase diagrams: predicting which compositions are stable across temperature and pressure ranges. Defect formation energies: predicting the cost of creating specific defects in specific structures. Surface energies: predicting energy costs for various surface terminations. Each extension has its own training-data substrate (typically smaller and more-specialised than the bulk-property databases) and its own active research community. The methodology continues to expand the scope of what materials AI can predict.

Machine-learned interatomic potentials (MLIPs) are neural networks trained to predict atomic energies and forces from local atomic environments. They provide fast surrogates for expensive DFT calculations — orders-of-magnitude faster while maintaining DFT-level accuracy — and have substantially expanded the accessible system size for atomic-scale simulation.

The interatomic-potential problem

Traditional interatomic potentials (also called force fields) are analytical functional forms — Lennard-Jones, embedded-atom method (EAM), Tersoff, ReaxFF, and many others — that approximate the dependence of atomic energies and forces on positions. Each is hand-designed for specific applications, with parameters fit to either experimental data or DFT calculations. The methodology has been the substrate of molecular dynamics for decades, but the analytical functional forms are limited: they cannot easily capture complex multi-body interactions, transferability across diverse chemistries is poor, and improving accuracy means manually proposing better functional forms. Machine-learned interatomic potentials replace the hand-designed functional form with a neural network trained on DFT (or higher-level) calculations. The flexibility is substantially greater; the challenge is to ensure stability, transferability, and respect for physical constraints.

The Behler-Parrinello foundation

The foundational paper is Behler & Parrinello 2007 (PRL): represent each atomic environment by a fixed-length feature vector (computed from radial and angular symmetry functions), train a separate neural network for each element to predict atomic energies from these features, and sum atomic contributions to get total energy. The methodology demonstrated that neural-network-based potentials could achieve DFT-level accuracy on bulk silicon. Subsequent work expanded the framework: Gaussian Approximation Potentials (GAP, Bartók et al. 2010) used Gaussian process regression rather than neural networks; Spectral Neighbor Analysis Potentials (SNAP) used bispectrum coefficients as descriptors; the various other variants (ANI, AENET, RuNNer, sGDML) explored related ideas. The 2010–2020 generation established MLIPs as a credible alternative to traditional force fields for specific applications.

Equivariant message-passing MLIPs

The 2022 watershed was NequIP (Batzner et al. 2022, Nature Comms): an E(3)-equivariant graph neural network for MLIPs, with intermediate representations that include not just scalars but also vectors and higher-order tensors transforming appropriately under rotations. The empirical case was substantial: NequIP achieved DFT-level accuracy with ~10× less training data than non-equivariant baselines, and produced more-stable molecular dynamics over long simulations. The paper triggered the modern MLIP wave. MACE (Batatia et al. 2022, NeurIPS) extended with higher-order equivariant message passing for further accuracy improvements at comparable computational cost. Allegro (Musaelian et al. 2023) introduced a strictly-local architecture (no global graph operations) for substantial speed improvements on large systems. Each generation has been deployed in increasingly-production-grade workflows.

Training data and active learning

MLIP training requires DFT-computed energies and forces on representative atomic configurations. Static training uses a pre-computed dataset (typically a few thousand configurations covering the relevant chemistry and thermodynamic conditions). Active learning iteratively expands the training set: train an initial model, run molecular dynamics with it, identify configurations where the model has high uncertainty, run DFT on those, retrain. The methodology has substantially improved data efficiency — modern active-learning pipelines can train production-grade MLIPs with O(1,000) DFT calculations rather than O(100,000). The empirical case for active learning is now well-established, and most modern MLIP training pipelines incorporate it explicitly.

What MLIPs enable

MLIPs have substantially expanded what's possible in atomic-scale simulation. Long timescales: DFT-quality molecular dynamics for nanoseconds rather than picoseconds, enabling the study of slow processes (diffusion, phase transformations, defect dynamics). Large systems: simulations of 10⁵–10⁶ atoms rather than 10²–10³, enabling realistic models of grain boundaries, interfaces, defect clusters. Free-energy calculations: free energies of formation, of transformation, and of reaction pathways at temperatures and pressures far from equilibrium. Reactive simulations: bond-breaking and bond-forming events that traditional force fields cannot capture (with appropriate training data including reactive configurations). The combined methodology has substantially expanded the accessible problem space for atomic-scale materials science.

Production deployment

MLIP workflows are increasingly mature. Software ecosystems include LAMMPS (the dominant classical MD code) with MLIP plugins, ASE (Atomic Simulation Environment) integrating multiple MLIP backends, JAX-MD for JAX-based differentiable MD, and the various Python-based packages for specific MLIP families. Computational performance on GPUs has reached the point where MLIPs are operational tools rather than research curiosities — modern equivariant MLIPs run at ~10⁻⁵ s/atom-step on GPU, sufficient for nanosecond-scale MD on systems of 10⁵ atoms. The methodology has substantially professionalised since 2022 and is now the practical default for atomic-scale simulation in many materials applications.

The 2022–2026 frontier of machine-learned interatomic potentials is universal force fields: single models trained on the entire Materials Project (or comparable substrate) that work across the periodic table without retraining. The methodology represents the materials-AI equivalent of foundation models in language and protein structure.

The case for universality

Traditional MLIPs are trained for specific chemistries: a Cu-Si MLIP for that binary, a NiTi MLIP for that shape-memory alloy, etc. Each requires its own training data and validation, and transferring across chemistries requires retraining. Universal force fields aspire to do for materials what AlphaFold did for proteins: train a single sufficiently-large model on a sufficiently-broad data substrate, and it works for arbitrary new chemistries without retraining. The empirical case has been gradually accumulating since 2022: pretraining on the entire Materials Project produces MLIPs that generalise to new compositions far better than chemistry-specific models, and the methodology has been substantially refined by the 2024–2026 generation.

M3GNet and the first generation

M3GNet (Materials 3-body Graph Network, Chen & Ong 2022, Nature Computational Science) was the foundational universal-MLIP paper: train a graph neural network on the entire Materials Project, providing a single model that works across the periodic table. The empirical results demonstrated substantial transferability — M3GNet predicts energies and forces for new chemistries with accuracy approaching chemistry-specific MLIPs trained from scratch, while requiring no specialised training. The methodology established the credibility of the universal-MLIP framing and triggered substantial subsequent work.

CHGNet and charge handling

CHGNet (Crystal Hamiltonian Graph Network, Deng et al. 2023, Nature Machine Intelligence) extended M3GNet with explicit handling of magnetic moments and atomic charges. The methodology is particularly important for transition-metal oxides, where charge-state changes (e.g., Mn²⁺ vs Mn³⁺ vs Mn⁴⁺) substantially affect energetics and structure. CHGNet has become the dominant universal MLIP for battery-cathode applications and various other transition-metal-rich materials, with substantial industrial adoption since 2024.

MatterSim and the foundation-model wave

The 2024–2026 generation of universal MLIPs is increasingly foundation-model-style. MatterSim (Microsoft Research, Yang et al. 2024) trains a 64-million-parameter equivariant model on a substantially-expanded dataset including high-temperature and high-pressure configurations beyond the Materials Project's typical range. The empirical case is that MatterSim handles thermodynamic conditions where M3GNet and CHGNet fail. Aurora (Microsoft Research, 2024) is a multi-domain foundation model that includes a materials-science backbone alongside its atmospheric-and-ocean components. SevenNet, EquiformerV2, and the various 2024–2026 successors continue the trajectory toward larger, more-broadly-trained models.

Empirical performance and limitations

The empirical performance of universal MLIPs is substantial but not perfect. In-distribution performance: for materials similar to the training set (typical bulk inorganic crystals near equilibrium), universal MLIPs achieve accuracy approaching chemistry-specific models. Out-of-distribution behaviour: for unusual chemistries (rare-earth-heavy, exotic intermetallics), high-temperature configurations, or substantial defect concentrations, universal MLIPs can fail substantially. The empirical evidence is mixed, with some studies showing OOD failures and others showing surprising robustness. The methodology of test-time fine-tuning — fine-tune a pretrained universal MLIP on a small task-specific dataset before deployment — is increasingly the production-grade approach, combining the universality benefit with task-specific accuracy.

The compute-vs-accuracy tradeoff

Universal MLIPs are computationally heavier than chemistry-specific models — more parameters, more attention computations, more inference cost per step. The tradeoff vs accuracy is favourable for many applications (the universality saves the cost of training a chemistry-specific model from scratch) but unfavourable for production high-throughput simulations where a chemistry-specific model trained for the application can be ~10× faster at fixed accuracy. The methodology has settled into a hierarchy: universal MLIPs for exploration and initial screening, fine-tuned or chemistry-specific models for production-scale simulations within established chemistries. The empirical landscape continues to evolve, and the long-term role of universal MLIPs in production materials-science workflows remains an open question.

Open frontiers

Several methodological frontiers shape the field. Going beyond DFT-level accuracy: universal MLIPs trained on DFT inherit DFT's systematic errors; methods that train on higher-level theory (CCSD(T), GW) for specific applications are emerging. Electron-explicit MLIPs: most current MLIPs predict total energy and forces but not electronic-structure properties (band gaps, magnetic moments); methods that predict these jointly are an active research direction. Reactive chemistry: extending universal MLIPs to bond-breaking and bond-forming events relevant for catalysis and synthesis is increasingly addressed by the 2024–2026 generation. The frontier is substantial, and the next several years will likely see continued advance.

Property prediction (Section 3) and force fields (Sections 4–5) evaluate existing materials. Generative crystal design proposes new ones. The methodology has matured substantially since 2022, with current methods routinely producing diverse, thermodynamically-stable, synthesisable candidates conditioned on target properties.

The problem and early approaches

Generative crystal design is the inverse of property prediction: given target properties, propose crystal structures that achieve them. The methodology is fundamentally harder than property prediction because the output space (the space of possible crystal structures) is vast and structured (must respect periodic boundary conditions, must have physically-reasonable bond lengths and angles, must satisfy charge balance for ionic compounds). Early approaches included enumeration over substituted prototypes (start with known structures, substitute elements in specific positions, evaluate stability), VAE-based generation (variational autoencoders trained on crystal structures), and GAN-based generation. The empirical performance was limited until the 2022 introduction of diffusion-based methods.

CDVAE and diffusion-based generation

CDVAE (Crystal Diffusion Variational Autoencoder, Xie et al. 2022, ICLR) was the watershed paper: a periodic-aware diffusion model for crystal-structure generation, demonstrating substantially better quality than prior VAE and GAN methods. The methodology adapts the diffusion-model framework to handle crystal-specific constraints: lattice parameters, atomic species, fractional coordinates, and the various symmetries are all represented appropriately. Generated candidates are routinely thermodynamically stable (within ~50 meV/atom of the convex hull) and have correct bond lengths, angles, and stoichiometry. The paper triggered substantial subsequent work.

DiffCSP and pure-diffusion methods

DiffCSP (Jiao et al. 2023) refined the diffusion methodology with substantial improvements in generation quality and training stability. The methodology operates directly on fractional coordinates with periodic-aware noise schedules, producing candidates with better lattice geometry than CDVAE. DiffCSP++ (Jiao et al. 2024) added space-group-aware generation, producing candidates that respect specified symmetry groups. The methodology has matured to the point of competitive empirical performance with the most-developed alternatives.

MatterGen and property-conditioned generation

MatterGen (Microsoft Research, Zeni et al. 2024, Nature) is the most-developed modern generative model for crystal structures. The methodology trains a unified diffusion model that generates crystal structures conditioned on target properties — composition, space group, magnetic ordering, mechanical moduli, chemical stability. The empirical case is substantial: MatterGen produces candidates that are simultaneously thermodynamically stable, satisfy specified target properties, and look novel relative to existing databases. The paper demonstrated experimental synthesis of MatterGen-generated candidates, providing concrete validation that the methodology produces synthesisable materials. The 2024 release triggered substantial industrial adoption.

Symmetry-aware generation

A specific architectural challenge is generating structures that respect crystallographic symmetries. Standard diffusion models generate continuous structures that may or may not have meaningful symmetries; physical crystals typically have specific space groups with corresponding restrictions on atomic positions. The 2024 generation increasingly handles this explicitly: generate the symmetric Wyckoff positions (the symmetry-equivalent classes of atomic sites), then expand to the full unit cell. Methods like DiffCSP++, SymCD, and various others substantially improve generation quality for specific space groups by working in the symmetric representation directly. The methodology connects to the broader equivariance machinery and is increasingly seen as essential for production-grade generative design.

Inverse design and property-conditioned workflows

Inverse design — start with a target property profile and propose materials that achieve it — is the natural use case for generative methods. Production workflows typically combine: (1) a generative model that proposes candidates conditioned on target properties; (2) a property-prediction model that evaluates the candidates; (3) a stability filter (typically using DFT for initial validation); (4) a synthesisability score (Section 10). The methodology has been substantially developed for batteries (cathode candidates with high specific capacity, low cost, good stability), magnets (hard ferromagnets without rare earths), and various other applications. The empirical record is increasingly substantial — multiple AI-discovered materials have entered the experimental-synthesis pipeline at major industrial labs since 2023.

Open challenges

Several open problems shape the methodology. Synthesisability: many generated candidates are thermodynamically stable but cannot be synthesised by realistic methods. Ranking the unknown: of many candidate materials, which deserve experimental investigation? Methods for combining property predictions with synthesisability estimates and uncertainty are still maturing. Beyond bulk crystals: generative methods for surfaces, interfaces, defected materials, and 2D materials are less mature. Multi-property optimisation: real applications need balanced property profiles (battery cathodes need high voltage AND high capacity AND good stability AND low cost), and joint optimisation across multiple targets remains technically challenging. The frontier is substantial, and the methodology continues to evolve rapidly.

The 2022–2026 wave of autonomous laboratories closes the design-make-test loop: ML-driven candidate selection plus robotic synthesis plus automated characterisation, all running with minimal human intervention. The methodology represents one of the most-direct examples of AI-for-Science methodology producing operational deployment.

The closed-loop concept

Traditional materials discovery involves substantial human labour at every stage: selecting which candidates to synthesise (chemist judgment), preparing precursors and running synthesis reactions (graduate-student labour), characterising results (XRD, microscopy, etc.), interpreting data, and deciding what to make next. Each cycle takes weeks to months. Autonomous laboratories automate this loop: ML-driven candidate selection from large databases, robotic synthesis platforms that prepare reagents and execute reactions, automated characterisation instruments, and ML-driven analysis that updates the candidate-selection model for the next cycle. The cycle time can shrink to days, and the human labour is reallocated from routine execution to high-level scientific direction.

A-Lab and the watershed result

The watershed paper is Szymanski et al. 2023 (LBNL, Nature): the A-Lab autonomous-laboratory platform demonstrated synthesis of 41 novel oxides over a 17-day operating window. The methodology combined: a ML-driven candidate-selection pipeline that proposed targets from the Materials Project; a robotic synthesis platform that prepared precursors, mixed them, fired in furnaces, and produced solid-state products; automated XRD characterisation that identified successful syntheses; and a feedback loop that updated the synthesisability predictor based on observed successes and failures. The paper demonstrated that autonomous materials discovery had crossed from research demonstration into operational viability, and triggered substantial subsequent work at academic and industrial sites.

Self-Driving Labs and the broader landscape

Several major efforts have produced increasingly-capable autonomous-lab platforms. Acceleration Consortium (University of Toronto): the largest academic centre, developing platforms across multiple chemistries. CMU's CAMD (Computational Autonomy in Materials Discovery): substantial methodological work on Bayesian optimisation for materials experiments. NIST's autonomous-lab efforts: substantial industrial-scale deployment of automated experimentation. Industry deployments: BASF, Dow, Toyota, and various other corporations have deployed autonomous-lab platforms for specific applications. The 2024–2026 generation has substantially expanded coverage from solid-state inorganic chemistry to organic synthesis, polymer formulation, and various other domains.

Bayesian optimisation and active learning

The mathematical core of autonomous-lab candidate selection is Bayesian optimisation: maintain a probabilistic model of how target properties depend on synthesis conditions or material composition, choose the next experiment to maximally reduce model uncertainty (or maximally improve expected outcomes), update the model based on the experiment, repeat. The methodology has been substantially developed for materials applications — multi-objective Bayesian optimisation for balancing competing properties, batch Bayesian optimisation for proposing multiple experiments simultaneously when synthesis platforms can run in parallel, cost-aware Bayesian optimisation for prioritising cheaper experiments, and the various active-learning variants. Modern implementations integrate these methods with specific synthesis-platform capabilities and characterisation-instrument constraints.

Specific application successes

Autonomous labs have produced concrete material-discovery wins across multiple domains. Solid-state battery electrolytes: discovery of new lithium-ion conductors with higher conductivity than state-of-the-art commercial materials. Catalyst optimisation: faster discovery of improved catalysts for various reactions. Photovoltaic materials: identification of new perovskite and emerging-photovoltaic chemistries. Permanent magnets: candidate magnet compositions without rare earths. Alloy optimisation: improved compositions for high-entropy alloys, biomedical implants, and various structural applications. The combined empirical record is increasingly substantial, and the autonomous-lab methodology is rapidly becoming a standard tool in industrial materials R&D.

Limitations and open problems

Autonomous labs have substantial limitations. Synthesis-method coverage: most platforms handle specific synthesis methods (solid-state, hydrothermal, certain wet-chemistry approaches) but cannot handle the full range of methods a human chemist might use. Characterisation depth: automated characterisation produces limited data per sample compared to detailed expert analysis; some material properties are simply not assessable in autonomous workflows. Scale-up: synthesising a few milligrams of a material is different from producing kilograms; autonomous labs handle the discovery scale but not the manufacturing scale. Novelty: autonomous platforms work best within established chemistries where the synthesis methodology is understood; truly-novel chemistries remain hard. The empirical case is that autonomous labs are substantially accelerating discovery within established frameworks, while not fundamentally changing what's discoverable.

The cultural and economic implications

A non-technical observation worth flagging: autonomous laboratories substantially change the economics and labour structure of materials research. The capital cost of an autonomous-lab platform is substantial (typically $1–10M for a meaningful capability), but the labour cost per experiment is much lower than traditional graduate-student-driven research. The methodology favours large institutions with substantial capital budgets, and the implications for academic research culture are still being worked through. The 2024–2026 trajectory suggests that autonomous-lab capability will increasingly be the differentiator between materials-research groups, with substantial implications for how materials science is conducted as a discipline.

AI-for-materials methodology is increasingly mature across multiple application domains. This section surveys the major ones — batteries, magnets, catalysts, photovoltaics, superconductors — with attention to which methods work where and what the empirical record looks like.

Battery materials

Battery materials are arguably the most-active AI-for-materials application area. The challenge: lithium-ion batteries have specific performance limits (energy density, power density, cycle life, cost, safety) and improvements require new cathode materials, new anode materials, new electrolytes (liquid or solid), and various other components. Cathode discovery: ML methods screen Materials Project candidates for high-voltage, high-capacity cathodes; the methodology has produced multiple candidates that have entered experimental synthesis at major battery companies. Solid-state electrolyte discovery: lithium-ion conductors with high conductivity, low electronic conductivity, and good mechanical stability — the methodology of high-throughput screening combined with autonomous synthesis has been particularly productive here. Beyond-lithium chemistries: sodium-ion, magnesium-ion, lithium-sulfur, and various other alternatives have their own substantial AI-driven discovery work. The empirical landscape is rapidly evolving, with new ML-discovered candidates appearing in commercial pipelines at multi-year cadence.

Permanent magnets

High-performance permanent magnets currently rely on rare-earth elements (Nd-Fe-B is the dominant chemistry), with substantial supply-chain and national-security concerns about rare-earth dependence. The materials-AI challenge is to discover rare-earth-free or rare-earth-light magnets with comparable performance. The methodology combines: ML-driven screening for candidates with high magnetisation and high magnetic anisotropy energy; specialised property prediction (magnetic properties are harder to compute than thermodynamic ones); and autonomous-lab synthesis of candidates. The empirical record is mixed — multiple candidates have been proposed and synthesised, but matching Nd-Fe-B's combined properties has proven hard. The methodology continues to develop, and rare-earth-light alternatives have produced incremental wins even where full rare-earth-free replacement remains elusive.

Catalysts

Catalyst discovery has substantial implications for energy transition (hydrogen production, CO₂ reduction, ammonia synthesis without natural-gas feedstocks) and chemical manufacturing. The materials-AI methodology for catalysts is mature: scaling relations (correlations between catalyst properties and reaction rates that simplify screening), DFT-based microkinetic modelling (predicting full reaction networks from first-principles calculations), and increasingly ML-augmented catalyst discovery (training networks to predict catalytic activity directly from catalyst structure). The methodology has produced concrete advances for hydrogen evolution, oxygen reduction, and various other reactions. Single-atom catalysts — catalysts where the active sites are individual metal atoms in specific environments — have been a particularly active AI-driven discovery target.

Photovoltaics and solar absorbers

The photovoltaic industry is dominated by silicon, but next-generation cells use perovskite solar absorbers (organic-inorganic hybrid perovskites achieving >25% efficiency in laboratory cells), tandem cells (silicon plus perovskite, achieving >30%), and various emerging chemistries. AI methods drive candidate-screening for perovskite compositions (the parameter space is enormous: ABX₃ structure with multiple options for each site, and substantial mixing). The methodology has been substantially active since ~2018; multiple AI-screened compositions have been synthesised and characterised, with substantial empirical refinement of the property-prediction methodology. The 2024–2026 generation increasingly addresses stability (a major remaining challenge for perovskite photovoltaics) alongside efficiency.

Superconductors and quantum materials

High-temperature superconductors and other quantum materials are particularly hard targets for materials AI: the underlying physics is incompletely understood, training data is sparse (we know which compositions are superconductors empirically but predicting from structure remains challenging), and the property prediction is inherently more difficult than for simpler endpoints. The methodology has nonetheless produced substantial work: candidate-superconductor screening using neural networks trained on the SuperCon database; topological-material discovery using ML methods to identify candidates with topological band structures; quantum-spin-liquid candidates identified through ML-augmented searches. The empirical record is more mixed than for less-exotic materials — claims of AI-discovered high-T_c superconductors have not always reproduced — but the methodology continues to develop.

Structural alloys and high-entropy alloys

Structural materials (alloys for aerospace, automotive, infrastructure) have a long history of empirical optimisation; AI methods are increasingly augmenting this. High-entropy alloys (HEAs) — alloys with five or more components in roughly equal proportions, opening enormous compositional spaces — are a particularly active AI target. The methodology screens HEA compositions for desired property combinations (strength, ductility, corrosion resistance, high-temperature performance) at speeds the traditional Edisonian approach cannot match. Aerospace nickel-base superalloys, titanium alloys for biomedical applications, and various other structural-alloy classes have substantial AI-driven optimisation work. The empirical record is solid for incremental improvements; producing fundamentally-new alloy concepts remains the harder problem.

Beyond traditional materials

AI-for-materials methodology is increasingly applied to non-traditional targets. Metal-organic frameworks (MOFs, with applications in gas storage, catalysis, and separations): generative MOF design has been substantially active. Polymers: while the methodology is less mature than for crystalline materials, polymer-property prediction and generative polymer design are active research areas. Nanomaterials: nanoparticles, 2D materials, nanowires each have specific applications and specific AI methodologies. Functional ceramics: piezoelectrics, ferroelectrics, thermoelectrics each have ongoing AI-driven discovery efforts. The application landscape continues to expand, and the methodology continues to mature across the broader materials space.

The equivariance methodology has been substantially developed for materials applications, and equivariant graph neural networks dominate the modern MLIP and property-prediction landscape. This section surveys the architectural framework that recurs across the chapter.

The symmetry-and-equivariance framework

Crystals have specific symmetries: the 3D Euclidean group E(3) (rotations, reflections, translations) acts on atomic positions and on physical observables. Equivariant neural networks are architectures whose internal representations transform predictably under group operations: a rotated input produces an appropriately-rotated output (for rotation-equivariance) or an unchanged output (for rotation-invariance). The methodology connects to the broader geometric deep learning framework (Bronstein et al. 2021, Ch 08 §9 develops the synthesis) and has been most-thoroughly tested in materials and protein applications.

Tensor-product representations

Modern equivariant architectures use tensor-product representations: internal features are not just scalars but include vectors (rank-1 tensors), tensors (rank-2), and higher-order tensors transforming according to specific irreducible representations of the rotation group. Operations on these representations use Clebsch-Gordan coefficients to combine tensors of various ranks while preserving equivariance. The methodology is mathematically demanding but produces architectures with substantially-better empirical performance than non-equivariant alternatives. e3nn (the dominant open-source library, Geiger & Smidt) provides the mathematical infrastructure for E(3)-equivariant networks and is used by NequIP, MACE, and most modern equivariant MLIPs.

Spherical harmonics and angular basis sets

A key implementation detail is that angular features are typically expanded in spherical harmonics: the natural basis for functions on the sphere with explicit rotation properties. The expansion order ℓ_max controls expressiveness: ℓ_max = 1 captures vector features (like force directions), ℓ_max = 2 captures tensor features (like quadrupole moments), and higher orders capture progressively-finer angular detail. The methodology trades expressiveness against computational cost — higher ℓ_max means more parameters and slower inference. Most production MLIPs use ℓ_max = 2 or 3, which empirically captures most of the relevant chemistry while remaining computationally tractable.

Message passing with equivariance

Standard graph neural networks compute messages between neighbouring atoms using scalar-only operations; equivariant GNNs compute messages that respect the geometric structure. The methodology requires specific architectural choices: messages are tensors transforming appropriately under rotation, aggregation operations preserve equivariance, and update functions combine equivariant features without breaking the symmetry. The various modern equivariant MLIPs (NequIP, MACE, Allegro, EquiformerV2) differ in specific architectural details — message construction, attention mechanisms, basis-function choices — but share the underlying equivariance machinery.

Empirical case for equivariance

The empirical evidence for equivariant architectures in materials is substantial. Data efficiency: equivariant networks require ~10× less training data than non-equivariant baselines for comparable accuracy on force prediction. OOD generalisation: equivariant networks generalise better to compositions and configurations outside the training distribution. Force-prediction accuracy: equivariant networks produce systematically more-accurate force predictions, particularly for non-equilibrium configurations relevant for molecular dynamics. MD stability: equivariant MLIPs produce more-stable molecular-dynamics simulations over long times. The combined empirical case has made equivariance the default for production-grade MLIPs as of 2026.

Beyond E(3): material-specific equivariances

Beyond standard E(3) equivariance, materials have additional symmetries that can be exploited. Space-group equivariance respects the discrete symmetries of specific crystal structures (the 230 space groups). Time-reversal symmetry matters for magnetic-structure prediction. Permutation equivariance handles the indistinguishability of atoms of the same element. The methodology of building these symmetries into neural networks is increasingly developed, with the 2024–2026 generation substantially expanding the equivariance toolkit.

The methodology synthesis

Geometric deep learning for materials has substantially matured since 2022. The empirical case is now sufficient that equivariance is essentially required for state-of-the-art MLIPs, and the methodology is well-supported by open-source infrastructure (e3nn, the various MLIP packages, the JAX-based reimplementations). The frontier increasingly involves combining equivariance with other architectural advances (transformers, foundation-model-style pretraining, multi-task learning) rather than developing equivariance per se. The methodology has reached a maturity comparable to convolutional networks for vision or transformers for language — the architectural foundations are settled and the field is iterating on applications and refinements.

AI for materials has matured substantially over the past five years. This final section surveys the frontier — the open methodological problems, the active research directions, and the operational questions that will shape the field over the next several years.

Out-of-distribution generalisation

The most-substantive open question is out-of-distribution generalisation: do ML methods trained on existing materials reliably extrapolate to fundamentally-new chemistries? The empirical evidence is mixed. Prediction-end methods (property prediction for materials similar to training data) work well; discovery-end methods (proposing fundamentally-new materials) face genuine OOD challenges that current methodology cannot fully address. The methodology of detecting when a model is operating out of distribution, providing calibrated uncertainty in those regimes, and combining ML predictions with physics-based grounding for OOD reliability is an active research direction. Whether AI methods can produce qualitatively-new materials beyond the training distribution, or remain primarily a methodology accelerator within human-led discovery, is an open question.

Defect-aware methodology

A persistent challenge (§4 of this chapter) is that most ML training data is perfect-crystal DFT, but real-world materials properties depend substantially on defects (point defects, dislocations, grain boundaries, surfaces, interfaces). The 2024–2026 generation has begun to address this systematically. Large-cell DFT calculations including representative defects produce training data that captures defect energetics. Defect-energy MLIPs are trained specifically on defect-rich configurations. Multi-scale frameworks combine perfect-crystal predictions with defect-energetics models. The methodology is rapidly maturing, and defect-aware materials AI is increasingly expected for production deployments. The frontier remains substantial: bridging from atomic-scale defect predictions to mesoscale microstructure to macroscale engineering performance is an unsolved problem with substantial implications.

Synthesisability and the discovery-to-deployment gap

A specific operational challenge is synthesisability: many ML-proposed materials are thermodynamically stable but cannot be synthesised by realistic methods. The methodology of predicting synthesisability is developing — SynthCNN and the various ML-based synthesisability predictors, retrosynthetic-style methods for solid-state synthesis (analogous to the organic-chemistry retrosynthesis of Ch 05 §16), autonomous-lab feedback that updates synthesisability scores based on experimental success rates. The methodology is substantially less mature than property prediction, and synthesisability remains a major bottleneck in materials-discovery pipelines. The 2024–2026 generation increasingly integrates synthesisability into candidate selection, rather than treating it as a downstream filter.

Multi-scale bridging

A persistent methodological challenge is multi-scale bridging: connecting atomic-scale ML predictions to mesoscale dynamics to macroscale engineering performance. The traditional methodology uses different theories at different scales (DFT at atomic, MD at nanometre, phase-field at mesoscale, finite-element at macroscale), with hand-coded interfaces between scales. ML methods are increasingly central to bridging these scales — neural-network surrogates for expensive lower-scale calculations, ML-derived constitutive relations for mesoscale models, and the various other multi-scale-AI applications. The methodology represents the substantial frontier for engineering-relevant materials AI.

The data-quality and curation frontier

A non-architectural frontier is data quality and curation. The major materials databases (Materials Project, OQMD, AFLOW) were assembled with various methodological choices that affect ML training: which DFT functional, what k-point density, what convergence criteria. Discrepancies across databases (the same material with different reported properties in different databases) are common and confusing. The methodology of multi-database harmonisation, DFT-functional unification, and systematic-error correction is gradually maturing. The 2024–2026 wave of large unified materials databases (the various "Big 5" consolidation efforts) addresses some of these issues; more remains to be done.

Industrial deployment and the productivity question

An operational question shaping the field is the actual productivity of AI-driven materials discovery in industrial settings. The empirical record is substantial: multiple AI-discovered materials have entered industrial pipelines, autonomous labs are deployed at major chemical and materials companies, and the methodology is increasingly central to industrial R&D. Whether the methodology produces the dramatic acceleration claimed by some (10× faster discovery, 100× more candidates evaluated) or merely incremental improvements over traditional methods is an empirical question that the next several years will resolve. The 2026–2030 timescale is likely to produce substantial empirical evidence on the productivity question.

The connection to AI-for-physics and AI-for-chemistry

A specific cross-domain observation: AI for materials connects to AI for physics (Ch 08, particularly NQS and density-functional ML methods) and AI for chemistry (Ch 02 §10's representations, Ch 05's drug-discovery methods that share the molecular-graph representation). The methodology of equivariant networks transfers across all three; force-field methods originally developed for materials are increasingly used for organic molecules and biomolecules; generative methods first developed for molecules transferred to crystals with appropriate adaptations. The cross-domain methodology is substantially shared, and developments in one area routinely propagate to others. The 2026–2030 trajectory will likely see continued integration across these adjacent fields.

What this chapter does not cover

Several adjacent areas are out of scope. The substantial materials-modelling literature on classical force fields (LAMMPS, GROMACS, AMBER, the various others) is acknowledged but not developed in detail. Specific subdomains — soft matter and polymers in production, biological materials, nuclear materials — are touched only briefly. Finite-element analysis and macroscale engineering simulation appears only in passing. Manufacturing methods (additive manufacturing, semiconductor fabrication, the various processing technologies) are out of scope despite their substantial AI applications. The mathematical foundations of continuum mechanics, plasticity theory, and fracture mechanics are not developed. The chapter aimed at the methodological core of AI for materials science; the broader landscape of materials AI is genuinely vast.