17.6 Materials science
Materials discovery has historically been slow: synthesise candidates suggested by chemical intuition or high-throughput experimentation, characterise them, iterate. Computational screening using density functional theory (DFT) accelerated this in the 2010s by allowing properties to be computed before synthesis, but DFT is expensive and limited in the size of system it can handle.
GNoME (Graph Networks for Materials Exploration; Merchant and colleagues, Nature 2023, DeepMind) used an active-learning loop with a graph neural network to propose novel stable inorganic crystal structures. Starting from the Materials Project (a publicly available database of around 150,000 known and predicted stable materials), GNoME ran iterative cycles of structure generation, GNN-based stability filtering, and DFT verification. The result was 2.2 million predicted stable materials, of which roughly 380,000 were considered new and added to the Materials Project. Subsequent experimental validation by Lawrence Berkeley National Laboratory's A-Lab autonomous laboratory (Szymanski and colleagues, Nature 2023) synthesised 41 of 58 attempted candidates from this set, demonstrating that the predicted structures were physically realisable.
The numbers should be qualified. A 2024 critique by Cheetham and Seshadri argued that many of the GNoME-proposed materials are minor compositional variants of known structures, and that the rate of genuinely novel functionality is much lower than the headline figure suggests. ML-augmented materials discovery is producing a step change in the rate at which candidate materials are proposed and screened, with the bottleneck shifting back to synthesis and characterisation.
Battery materials are an area of particular interest. Microsoft and Pacific Northwest National Laboratory's January 2024 announcement of a candidate solid-state electrolyte found through ML screening attracted attention; the candidate, a sodium-lithium superionic conductor, was synthesised and tested within weeks of the prediction. Other active areas include catalysts for hydrogen production, magnetocaloric materials and superconductors. The MACE family of equivariant neural-network potentials (mentioned in Section 17.2) is increasingly used to drive molecular dynamics for these applications.