Catalog / Science, Engineering & R&D / SCI-003

Materials discovery & property prediction

Use graph neural networks, machine-learned interatomic potentials and generative models to screen, propose and rank new materials (batteries, catalysts, alloys, semiconductors, polymers) before costly synthesis and testing.

Parts needed to run it yourselfAgents & automationCoding assistantLanguage modelOptimizationScience & Earth dataPrediction & time series

Finding a better battery electrolyte, catalyst, magnet or alloy traditionally means years of trial and error plus expensive quantum-mechanical (DFT) calculations. AI shortcuts this in three ways: fast surrogate models predict properties of millions of candidates, machine-learned interatomic potentials (MLIPs) run near-DFT-quality atomistic simulations at a fraction of the cost, and generative models propose new structures for a target profile. The result is a short, ranked list for the lab instead of a blind search. It remains a funnel: AI narrows, experiments decide.

Problem & value

  • Today: materials R&D combines chemical intuition, literature, design of experiments and DFT. DFT scales poorly with system size (typically cubic in electrons), so large screening campaigns, finite-temperature dynamics or reactive systems are often out of reach.
  • AI value: universal MLIPs and property predictors make screening of very large candidate spaces feasible; generative models search compositions that humans would not try; active learning chooses the next experiment. Published landmark results: DeepMind's GNoME reported about 2.2 million new crystal structures predicted stable, of which a few hundred thousand lie on the updated convex hull [3]; Microsoft's MatterGen generated structures for target properties and one was synthesised and measured within about 20% of the target (bulk modulus) [4]. Both are research demonstrations; treat counts as candidates, not discoveries.
  • For SMEs: the realistic gain is faster iteration on one problem using open pretrained models plus a few hundred own data points.

Variants

  • Inorganic crystals / energy materials: battery cathodes, solid electrolytes, magnets, thermoelectrics, photovoltaics, hydrogen-storage materials.
  • Catalysis and surfaces: adsorption energies, reaction barriers (Open Catalyst-style tasks) for electrolysers, fuel cells, CO2 conversion.
  • Alloys, polymers, semiconductors: CALPHAD-assisted ML for alloys, monomer-graph property prediction (see SCI-014 for formulation), defects and dielectrics.
  • Atomistic simulation with MLIPs (method and building block): MACE, UMA and MatterSim serve as the simulation path inside this use case, for catalysis, batteries and semiconductors and also for pharma solid forms and polymorph ranking (molecular crystals; see SCI-002 for the drug-discovery side). MLIPs are a reusable component, not a separate business case.

How it works

target profile (properties, constraints, cost/toxicity/supply)
  -> candidate generation (substitution, enumeration, generative model, literature mining)
  -> fast filters (charge neutrality, composition rules, ML stability/property surrogates)
  -> structure relaxation + properties with MLIP (energy, forces, phonons, MD, barriers)
  -> uncertainty check (ensembles/model disagreement) -> selective DFT verification
  -> ranking (stability, property, cost, REACH/supply-risk screen)
  -> human expert review -> synthesis planning / lab queue (SCI-007) -> measurements
  -> feedback into data store and retraining / active learning
  • Property predictors: graph neural networks on crystal graphs (CGCNN lineage, MEGNet, M3GNet/CHGNet, equivariant models) predict formation energy, band gap, elastic moduli.
  • Machine-learned interatomic potentials: equivariant networks trained on DFT energies/forces. Universal ("foundation") MLIPs cover most of the periodic table: MACE-MP-0 and MACE-OMAT, Meta's UMA family (trained on OMat24, OMol25, OC20 and related datasets, Mixture-of-Linear-Experts design) [2], Microsoft's MatterSim, plus SevenNet, ORB, Equiformer-type and CHGNet models. Fine-tuning on a few hundred system-specific DFT calculations is the standard way to close accuracy gaps.
  • Generative models: diffusion models over atom types, positions and lattice (MatterGen, DiffCSP-style) and LLM-based CIF generators; output must always be relaxed and verified.
  • LLM/agent layer: LLMs mine literature, write ASE/pymatgen scripts and orchestrate workflows (SCI-011); they never replace the physics check.
  • Human-in-the-loop: domain experts define targets, judge synthesisability, decide which candidates go to the lab, and validate any model used outside its training domain.

Data preparation

Inputs are crystal structures, compositions, measured properties, simulation results (DFT) and process parameters. In a materials lab or manufacturer they live in lab notebooks, Excel sheets, instrument exports, simulation output folders, ERP/MES for process data and open databases.

Typical gaps: properties measured under undocumented conditions, inconsistent units and composition notation, results from failed syntheses never recorded, mixed DFT settings that make energies incomparable, and process data scattered across machines and paper protocols.

Preparation steps: 1. Inventory sources, owners and formats (PREP-001, PREP-002, PREP-005). 2. Digitise paper protocols and logbooks (PREP-006, PREP-008). 3. Restructure spreadsheets into tidy tables with units and measurement conditions (PREP-011). 4. Fix a notation standard and glossary for compositions and properties (PREP-012, PREP-015). 5. Harmonise DFT settings or label them clearly, and keep negative results (PREP-020). 6. Build a hold-out set with domain experts (PREP-018).

Effort share: high - often more than half of the effort, since pre-trained potentials need little setup but property models depend on consistent in-house data. See GUIDE-003.

Implementation paths

Which path for whom

  • Enterprise: Path C - material compositions and process data are trade secrets, and open universal interatomic potentials (MACE, CHGNet, MatterSim-class) run on an own GPU cluster (HW3) with open databases such as Materials Project.
  • Developer: Path C - the open models and ASE-based tooling are mature and easy to integrate; cloud APIs (Path B) are only a fallback for short peaks.
  • Public: Path C - universities and research institutes can run the open stack on HPC or sovereign cloud and keep data and results reproducible.

Path A — Buy (off-the-shelf)

  • Commercial platforms: Schrodinger Materials Science suite (DFT, MD, ML), Dassault Systemes BIOVIA Materials Studio and Pipeline Pilot, Citrine Informatics (materials data platform and Sequential Learning), Microsoft Azure Quantum Elements / Generative Chemistry, Rowan (cloud atomistic workflows with open MLIPs), SandboxAQ AQChemSim, Matlantis (Preferred Networks / ENEOS, universal MLIP as a service), Orbital Materials, CuspAI (UK/NL, EU-linked), Radical AI. Mostly US/UK vendors; CuspAI has EU ties; check EU hosting and data terms per contract.
  • Choose A when: one well-defined problem, small computational team, need for support and validated workflows.

Path B — Build on cloud APIs

  • Compute and models: run open MLIPs (MACE, UMA, MatterSim) on EU-region GPU cloud (OVHcloud, Scaleway, IONOS, STACKIT, AWS/Azure EU regions); Azure Quantum Elements and Rowan expose hosted workflows.
  • Data/APIs: Materials Project API, OPTIMADE (standard API across databases), NOMAD (Germany, FAIR repository).
  • Orchestration: AiiDA, atomate2; ASE and pymatgen; LLM APIs (Claude, Mistral, Gemini, GPT) for literature mining and code.
  • Choose B when: you want burst compute without owning GPUs and have in-house scientific-Python skills.

Path C — Self-host (open source)

  • Models/tools: MACE (MIT licence), UMA via fairchem (Meta; check model licence terms before commercial use), MatterSim, CHGNet, SevenNet, ORB; MatterGen (MIT-licensed code and weights from Microsoft); Matbench Discovery to compare MLIPs; DFT with Quantum ESPRESSO, GPAW, CP2K, ABINIT, VASP (commercial licence), LAMMPS and ASE for dynamics.
  • Hardware: HW2 runs inference and small MD with universal MLIPs; HW3 (48-141 GB GPU) is the practical minimum for fine-tuning, large cells and generative sampling; large DFT campaigns use CPU clusters or EuroHPC/national HPC.

Path D — Custom training

Worth it when the chemistry is outside the training distribution (e.g. molten salts, specific interfaces, rare-earth-free magnets) or when you own proprietary measurement data. Needs hundreds to a few thousand DFT-labelled structures chosen by active learning, or experimental datasets with consistent protocols. Fine-tuning a foundation MLIP is preferred over training from scratch.

State of the art (2026)

  • Universal MLIPs are the biggest practical change: MACE-MP-0, UMA, MatterSim and peers reproduce DFT energies and forces across much of the periodic table; the Matbench Discovery benchmark tracks how well they find stable crystals [5]. Meta's OMat24/OMol25 and the UMA paper (up to about 1.4B parameters, roughly 50M active per structure) pushed training data into the hundreds of millions of structures [2].
  • Generative design (MatterGen, GNoME-style active learning loops) produces stable-looking candidates for target properties; synthesis validation remains sparse [3][4].
  • Limits and failure modes: (1) MLIPs often soften (underestimate) potential-energy curvature and can fail on out-of-distribution chemistry, high-energy configurations, magnetism, charged or long-range systems; (2) low DFT-level error does not guarantee correct finite-temperature or kinetic properties; (3) energy above hull ignores kinetics, so "stable" compounds may be unsynthesisable; (4) training data are mostly DFT with functional errors (PBE band gaps, vdW); (5) database bias towards well-studied chemistries; benchmarks vs experiment are scarcer than vs DFT.

Shared resources & combinations

Effort, cost & KPIs

  • Effort: A weeks; B 2-4 months; C 3-6 months (1-3 computational scientists); D 6-12+ months, DFT labelling dominates cost.
  • KPIs: experimental hit rate of predicted candidates; MAE vs measured property on held-out data; experiments or DFT hours saved per qualified material; time to first validated candidate; share of out-of-distribution flags.

People & change

Involve early: management and the R&D department lead define target properties and the budget. Data owners (lab, simulation group, production data from MES/ERP) must agree on data release. IT provisions GPU and storage; IT security and legal/compliance handle trade secrets, research contracts and export control for dual-use materials. Materials scientists as end users should co-design the pilot.

Works council: depends. The models evaluate materials, not people. §87(1) no. 6 BetrVG only applies if the platform records individual user activity (job logs, usage statistics) in a way suited to monitor performance; then regulate it by works agreement. Inform the works council in good time (§90(1) no. 3 BetrVG); universities and public bodies involve the Personalrat. See GUIDE-004.

Data protection: normally no personal data; check user accounts and logs with the DPO.

AI literacy: support the team with training on model limits, out-of-distribution predictions and uncertainty, as Art. 4 AI Act expects (GUIDE-006).

Communication and pilot: start with one material class and a small volunteer group; compare predictions with measurements and report openly (GUIDE-008, GUIDE-009).

Typical concerns: fear that simulation replaces experiments or experience. Make clear that candidates are ranked, not decided, and that lab validation stays mandatory.

Risks, regulation & compliance

  • EU AI Act: typically minimal risk; scientific research models are largely outside the Act when developed solely for scientific research (Art. 2(6)), and materials property prediction is not an Annex III use. No Art. 5 prohibited practice applies. AI literacy (Art. 4) applies to staff. If a general-purpose model is integrated, GPAI provider obligations sit with the model provider.
  • GDPR: chemical/physical data, no personal data beyond minor lab-notebook author metadata.
  • Chemicals and product law: candidate materials must be screened under REACH/CLP, RoHS, the EU Batteries Regulation (2023/1542), and PFAS restriction proposals; use SCI-015 early to avoid dead ends.
  • Dual use: generative models can propose hazardous substances; use access control and blocklists in shared platforms.
  • IP and licences: check dataset and model-weight licences (Materials Project, OMat24/OMol25, UMA) before commercial use; document inventorship.
  • Scientific risk: over-trust in predictions; mitigate with uncertainty estimates, DFT spot checks, experimental validation and recorded model versions.

Getting started

  1. Define one concrete target property and constraints (for example Li-ion conductivity above a threshold, no cobalt, REACH-compliant).
  2. Install ASE, pymatgen and a universal MLIP (MACE or UMA); reproduce 3-5 known properties of a material you already understand and compare with your lab data.
  3. Pull candidates through the Materials Project API or OPTIMADE, screen with the MLIP and rank by stability and property.
  4. Verify the top 10-20 with DFT and send 3-5 to the lab; log all results in a workflow manager such as AiiDA.
  5. If errors are systematic, fine-tune on a few hundred targeted DFT calculations and repeat with active learning.

Sources

  1. Batatia et al. — MACE-MP-0: a foundation model for atomistic materials chemistry — https://arxiv.org/abs/2401.00096
  2. Wood et al. (Meta FAIR) — UMA: A Family of Universal Models for Atoms — https://arxiv.org/abs/2506.23971
  3. Merchant et al. (DeepMind) — Scaling deep learning for materials discovery, Nature 2023 — https://www.nature.com/articles/s41586-023-06735-9
  4. Zeni et al. (Microsoft) — A generative model for inorganic materials design, Nature 2025 — https://www.nature.com/articles/s41586-025-08628-5
  5. Riebesell et al. — Matbench Discovery — https://arxiv.org/abs/2308.14920
  6. Materials Project API and documentation — https://docs.materialsproject.org/
  7. OPTIMADE consortium — https://www.optimade.org/
  8. AiiDA workflow and provenance framework — https://www.aiida.net/
  9. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
  10. Regulation (EU) 2023/1542 (Batteries Regulation) — https://eur-lex.europa.eu/eli/reg/2023/1542/oj