AI in Materials Science​

Materials science spans the discovery, design, characterisation and manufacture of materials. This review covers the full pipeline, from modelling and design through characterisation and laboratory automation to production and scale-up, linked by a cross-cutting PSPP layer (processing–microstructure–property–performance; CALPHAD, phase-field, ICME).

• Infrastructure, not method, is the bottleneck: no shared layer lets models be trained, compared, reused and linked to the instruments producing the data
• Predicting what a fabricated material does requires learning the electronic Hamiltonian, not only the interatomic potential, and solving it at device scale without O(N³) diagonalisation
• Models trained on simulated or idealised data degrade on real specimens; no benchmark with independent ground truth measures by how much
• A foundation-scale multi-modal model must reason jointly over the experimental path: synthesis, processing, microscopy, spectroscopy, diffraction
• One semantic layer of ontologies and graphs describing materials, properties and processes together, where different formalisms can coexist
• Autonomous laboratories need safeguards built into the models and agents, not only external controls

• Scarce experimental and industrial data – far worse for surfaces and interfaces, so batteries and catalysis cannot yet be reliably predicted from atomic-scale data, despite claims
• Negative results rarely reported; independent reproduction not yet established
• Interoperability gaps across workflows, instruments and closed vendor formats
• High automation, robotics and compute costs

• Operational: Property prediction, high-throughput screening, machine-learning interatomic potentials and materials database analytics​
• Emerging: Generative materials design, AI-enhanced characterisation, materials foundation models and workflow optimisation​
• Frontier: Self-driving laboratories, autonomous discovery systems, AI-enabled manufacturing optimisation and integrated lab-to-fab workflows​

• Data and modelling: NOMAD, FAIRmat, MatWerk NFDI, Materials Cloud, AiiDA, OPTIMADE, MaterialDigital, EMMC.
• Federation: MaterialsCommons, ELAI under RAISE.
• Experimental data: PaNOSC, ExPaNDS, DAPHNE4NFDI, LEAPS-INNOV at photon and neutron sources, already issuing DOIs for raw data.
• Programmes: BATTERY 2030+ (incl. BIG-MAP), CAPeX, GenAI4EU, EuroHPC
• Policy: GCSA Scientific Opinion No. 18 on Advanced Materials; forthcoming Advanced Materials Act

• Federated materials data
• Multiscale PSPP-aware AI
• Agentic labs with one computational–experimental loop
• Safe and Sustainable by Design
• Model safety

Rather than replicate the compute-heavy large-model paths of the US and China, Europe should build on its lead – first-principles simulation (DFT in particular), open data, large experimental facilities – and supply benchmarks, independent validation and models of real, imperfect materials.