Machine Learning for Materials Discovery

The rate-limiting step in finding a new inorganic material has never been imagination; it has been arithmetic. To know whether a hypothetical crystal will actually hold together, you compute its energy with density functional theory (DFT), and a single DFT relaxation of a modestly sized unit cell costs hours of CPU time. Searching a chemical space of millions of candidate compositions and structures that way is a supercomputer-decade problem. In 2023, DeepMind's GNoME project reported 2.2 million candidate crystals screened and 381,000 predicted to be stable, an order-of-magnitude expansion of the known stable inorganic set, achieved by putting a cheap neural surrogate in front of the expensive physics. That is the whole game: learn to predict what DFT would say, then only run DFT where the prediction is promising.

Two learned surrogates for two different costs

Materials machine learning replaces two distinct expensive calculations, and it helps to keep them separate.

The first is stability screening: given a proposed crystal structure, will it sit at or below the convex hull of competing phases (the energy floor of already-known compounds at that composition), meaning it is thermodynamically favoured to form rather than decompose? This is a per-structure scalar prediction. GNoME uses graph neural networks (GNNs) over the crystal graph, atoms as nodes, bonds or near-neighbour contacts as edges, to predict formation energy and hence distance to the hull. The network is trained on the Materials Project corpus of DFT-computed energies, then used to rank enormous candidate sets so that only the plausible few percent are handed to real DFT.

The second is the energy and forces of a specific atomic configuration, which is what molecular dynamics and structure relaxation actually need at every timestep. This is the job of a machine-learning interatomic potential (MLIP), also called a machine-learned force field. An MLIP such as CHGNet or MACE learns a function from atomic positions and species to total energy (and its gradient, the forces on each atom), trained to reproduce DFT energies and forces. Once trained, it evaluates in milliseconds where DFT takes hours, making molecular-dynamics-scale simulation of materials, thousands of atoms over nanoseconds, feasible on a workstation.

The two are complementary. A stability model tells you which compositions are worth investigating; a potential lets you simulate the winning ones without invoking a quantum chemistry code every step. GNoME in fact trained both a stability filter and a general-purpose potential, and used the potential to pre-relax candidate structures before the final DFT check, compounding the saving.

The discovery pipeline

The standard workflow is a funnel, wide and cheap at the top, narrow and expensive at the bottom.

1. Generate candidates. Substitute elements into known prototype structures (swap the cation in a known oxide framework), or sample structures more aggressively. This produces millions to hundreds of millions of hypothetical crystals.
2. Screen with the learned surrogate. Run the GNN stability model, and often an MLIP relaxation, over the whole set. Keep only structures predicted to be at or very near the convex hull. This step is where the compute leverage lives; it cuts the set by one to two orders of magnitude.
3. Validate the survivors with DFT. Run real DFT relaxations on the shortlist to confirm the surrogate's verdict. This is the ground truth for a computational claim of stability, and the new DFT results are folded back into the training set, an active-learning loop that widened GNoME's coverage into chemistries the initial model had never seen.
4. Synthesise and characterise experimentally. Hand the best candidates to a lab (or an autonomous lab). This is the only step that produces a real material, and it is by far the hardest to scale.

The leverage is enormous and real. GNoME's authors reported that external laboratories, working independently, had already synthesised 736 of the predicted structures around the time of publication, evidence that the stability predictions correlate with reality for a meaningful slice of the set. The honest reading of that number is also its limit: 736 confirmations against 381,000 predictions is roughly two in a thousand, and every one of those was chosen by a human with synthesis expertise, not sampled at random from the database.

When it falls down

This is where the field's real subtlety lives, and where the 2023-2024 debate around GNoME is genuinely instructive.

• "Stable in silico" is not "synthesised". A prediction that a structure sits on the DFT convex hull is a thermodynamic statement about a perfect crystal at zero kelvin. It says nothing about whether a viable kinetic pathway exists to make the thing, whether it survives at room temperature, or whether it does anything useful. Cheetham and Seshadri, materials scientists at UC Santa Barbara, argued in Chemistry of Materials (2024) that among GNoME's stable set there was "scant evidence for compounds that fulfil the trifecta of novelty, credibility, and utility", and that many entries are chemical compounds, not materials, with no demonstrated function. A separate critique led by Robert Palgrave (UCL) noted that many database entries assume specific orderings of metal ions on shared sites that are unlikely to order that way in a real solid, and that some proposed compounds already exist in experimental databases. Treat a database of "new stable materials" as a database of candidates, not discoveries.

• A model of DFT inherits DFT's errors. The surrogate is trained to reproduce DFT, so its ceiling is DFT's accuracy, not nature's. DFT with standard functionals systematically mis-estimates certain energies (strongly correlated electrons, van der Waals interactions, some transition-metal oxides). The neural network cannot correct an error it was trained to imitate; it can only propagate it faster and to more compounds.

• Distribution shift to novel chemistries. The models are strongest near their training distribution. Push a stability GNN or an MLIP into an element combination or a structural motif poorly represented in the Materials Project and confidence degrades silently, the model returns a smooth, plausible-looking energy that is simply wrong. This is exactly why the active-learning loop (validate with DFT, retrain) matters, and why extrapolation claims deserve scepticism.

• The synthesis gap. Even a genuinely novel, credible, useful prediction still has to be made. Precursor selection, temperature and pressure schedules, and reaction kinetics are not captured by a zero-kelvin hull calculation. Autonomous laboratories (DeepMind's A-Lab collaboration among them) attack this, but the throughput of real synthesis remains orders of magnitude below the throughput of prediction, and independent chemists have questioned how many autonomous-lab "successes" were truly new versus known phases.

The correct mental model: machine learning has made the computational funnel thousands of times wider, so the bottleneck has moved decisively downstream to DFT validation and, above all, to experimental synthesis. Prediction is cheap now. Verification is not.

Further reading

• Scaling deep learning for materials discovery - Merchant et al, Nature 2023; the GNoME paper (2.2M screened, 381k predicted stable).
• Millions of new materials discovered with deep learning - DeepMind's accessible companion, including the 736 externally synthesised structures.
• CHGNet: a pretrained universal neural network potential for charge-informed atomistic modelling - Deng et al 2023; a universal MLIP that carries charge/magnetic-moment information.
• MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields - Batatia et al 2022; the equivariant MLIP architecture now widely used for materials MD.
• Boffins deem Google DeepMind's material discoveries shallow - readable summary of the Cheetham and Seshadri and Palgrave critiques of the GNoME stability claims.