AI in materials science: how machine learning predicts and designs new materials

What machine learning does in materials science

New materials used to come from trial and error at the bench. Computation changed that: databases such as the Materials Project, the Open Quantum Materials Database, AFLOW and NOMAD use density functional theory (DFT), a quantum-mechanical approximation, to compute the energies of known and hypothetical crystals. By the GNoME team's recalculation, that approach had identified about 48,000 computationally stable crystals, against about 20,000 computationally stable structures among the 200,000 entries that experiments have catalogued in the Inorganic Crystal Structure Database (ICSD).

DFT is accurate enough to act as the referee, but too slow to run on everything. Machine learning sits in front of it and does four jobs:

The first three jobs learn from computed data, mostly DFT. Their outputs are therefore predictions of what DFT would say, not of what a lab will measure. That distinction runs through everything below.

How AI predicts a material's properties from its crystal structure

The workhorse is the graph neural network (GNN). A crystal is a small cell of atoms repeated in space, and a GNN represents it as a graph: a node for each atom, an edge for each bond to a neighbour. Each node starts as a description of its element. In each layer, nodes exchange messages with their neighbours and update, so after a few layers every atom carries a summary of its local chemical environment. The network then pools the atoms into one prediction for the whole crystal.

The crystal graph convolutional neural network (CGCNN), published in Physical Review Letters in April 2018, showed that this works. Learning directly from how the atoms in a crystal connect, it predicted eight different DFT-calculated properties accurately across many structure types after training on the order of 10,000 examples, and it could show which local environments contributed most to a property. Later models such as MEGNet and ALIGNN refined the idea. Note what the labels were: DFT results. A model trained on DFT reproduces DFT's answer, systematic errors included, only much faster.

The MIT study: predicting how heat moves through a crystal

Thermal properties show why a fixed-size answer is not always enough. MIT's July 2024 announcement cites an estimate that about 70% of the energy generated worldwide ends up as waste heat. In semiconductors and insulators, heat is carried by phonons, the quantized vibrations of the crystal lattice, and several thermal properties depend on the phonon dispersion relation: how phonon energy varies with momentum. It is expensive to compute. One of the study's lead authors put the cost of calculating it for a single material at roughly 100 CPUs for a few weeks.

A standard GNN produces a fixed set of outputs, while a dispersion relation is a set of curves whose size depends on the material. The team's paper, published in Nature Computational Science on 12 July 2024, handles this with virtual nodes: extra nodes added to the crystal graph that receive messages from the atoms but send none back. The atoms never see them, so they do not change what the network learns about the crystal, and their number can be set to the size of output each material needs. Three variants predict phonons at the centre of the Brillouin zone (the Γ point) and full dispersion curves, directly from atomic coordinates. The training data came from a high-throughput database of phonons computed with density functional perturbation theory (DFPT), the standard way to calculate them.

The gain is speed at similar accuracy. Compared with machine-learned interatomic potentials, the paper reports efficiency orders of magnitude higher with comparable or better accuracy, and the authors used the model to build a database of Γ-point phonons for over 146,000 materials. MIT's announcement put the speed-up at up to 1,000 times over other AI-based methods, and possibly a million times over conventional non-AI approaches; its senior author, Mingda Li, said a few thousand materials could be done in a few seconds on a personal computer. MIT also reports slightly better accuracy for heat capacity, with errors in some cases two orders of magnitude lower.

What the study did not do is make a material. It predicts computed phonons, so its role is screening: finding candidates for thermal storage, energy conversion or better heat management in microelectronics, which DFPT and then experiments must confirm. Cheap phonons matter for discovery in another way, covered below: they can reveal crystals that a stability calculation calls stable but that are dynamically unstable.

How machine-learned interatomic potentials replace expensive simulations

A property model answers one question about a fixed structure. A machine-learned interatomic potential (MLIP) learns something more general, the potential energy surface. Given the elements and positions of the atoms, it returns the energy, the force on each atom and the stress on the cell, having been trained on those quantities from DFT. With forces, it can do the everyday work DFT is used for: relax a hypothetical structure to its lowest-energy shape, run molecular dynamics at a temperature, or compute vibrations.

Early potentials were fitted to one material at a time, each with its own training set. Large databases made universal potentials possible, covering 90 or more of the most useful elements. M3GNet, published in November 2022, was trained on ten years of structural relaxations from the Materials Project. Its authors screened 31 million hypothetical crystals with it and flagged about 1.8 million as potentially stable; of the 2,000 lowest in energy, DFT confirmed 1,578. GNoME's hundreds of millions of DFT calculations were put to the same use, training potentials whose molecular dynamics predicted ionic conductivity without ever being trained on it.

A benchmark built for exactly this question is Matbench Discovery, published in Nature Machine Intelligence in June 2025. It evaluates models the way they are used, as a filter in front of DFT, on 257,487 hypothetical structures generated by element substitution. Ranked by F1 score for sorting stable from unstable crystals, universal potentials took the top six places in the paper's initial release: EquiformerV2 + DeNS, Orb, SevenNet, MACE, CHGNet and M3GNet, ahead of energy-only GNNs such as ALIGNN, MEGNet and CGCNN. The top models scored F1 between 0.57 and 0.82. The reason is mechanical: a potential uses forces and stresses to imitate the DFT relaxation before it predicts the final energy, while an energy-only model has to judge from the unrelaxed input. The benchmark keeps a public leaderboard that grows as models are submitted, so treat this ranking as a snapshot.

The same technique, trained on molecules instead of crystals, is behind the machine-learned force fields used to simulate how drugs bind their targets.

Screening at scale: what GNoME predicted, and what was made

GNoME (graph networks for materials exploration), published by Google DeepMind in Nature on 29 November 2023, generated candidate crystals in two ways: symmetry-aware partial substitution of elements into known structures, and ab initio random structure searching. GNNs filtered the candidates, DFT computed the energy of the survivors, and every result went back into training for the next round: active learning, which the paper calls a data flywheel. Over six rounds, the share of predicted-stable structures that DFT confirmed (the hit rate) rose above 80% for models given a structure, and to 33% per 100 trials for models given only a composition, against 1% in earlier work.

The output was 2.2 million structures below the previous convex hull, the boundary of lowest-energy compounds that decides whether a crystal is computed to be stable. Because new crystals compete with each other, 381,000 of them remain as new entries on the updated hull, for 421,000 stable crystals in total. The paper reports that 736 of its stable structures had been made independently by other labs, and screens the set for uses: about 52,000 layered compounds where earlier methods found about 1,000, and 528 promising lithium-ion conductors.

The pushback came from chemists. In April 2024, Anthony Cheetham and Ram Seshadri examined a small random subset of the predictions in Chemistry of Materials and found scant evidence of compounds that are at once novel, credible and useful. Two problems recur in their examples. DFT at 0 K ignores entropy, so it places similar atoms on separate sites where, at the temperatures used for synthesis, they would mix; in several of their examples, the disordered version of a "new" compound matches a known structure. And more than 18,000 of the predictions contain promethium, actinium or protactinium, radioactive elements available only in minute quantities. They still judged the methods promising, and asked for expertise in synthesis and crystallography inside the loop.

Generative models: how MatterGen designs a crystal to order

Screening can only rank the candidates someone thought to generate. A generative model proposes structures directly from a specification. MatterGen, published by Microsoft researchers in Nature on 16 January 2025, is a diffusion model. Much like an image diffusion model turns noise into a picture, it starts from a random structure and refines the atom types, their positions and the periodic lattice step by step until a crystal emerges. Adapter modules let it be fine-tuned towards a target chemistry, symmetry or property, such as magnetic density, band gap or bulk modulus (stiffness under compression).

The base model was trained on 607,683 stable structures of up to 20 atoms, recomputed from the Materials Project and Alexandria databases. Compared with earlier generative models such as CDVAE and DiffCSP, its structures were more than twice as likely to be stable, unique and new by DFT, and more than ten times closer to their relaxed energy minimum. Microsoft released the code under the MIT licence, together with the training data.

The lab test is where precision matters. For each of four bulk modulus targets (50, 100, 150 and 200 GPa), the team generated 8,192 candidates. Filters for novelty, stability (with Microsoft's MatterSim potential, then DFT), vibrational stability and oxygen content narrowed them to 75. Experts chose four to synthesize, and one worked: TaCr2O6, generated for the 200 GPa target. The sample was a compositionally disordered version of the ordered structure MatterGen proposed, with tantalum and chromium sharing sites. DFT gave the ordered structure 222 GPa; the bulk modulus estimated from nanoindentation was up to 169 GPa, within 20% of the target.

In June 2026, a paper in Materials Horizons argued from a detailed crystallographic analysis that the compound is not new: it is identical to a tantalum chromium oxide first described in 1971, which was in MatterGen's training data. The MatterGen team had tried to guard against exactly this. Its novelty check compares generated structures with known ones, including 117,652 disordered ICSD structures, using a matcher built to treat ordered and disordered versions as the same. The critique's broader point is the one Cheetham and Seshadri made about GNoME: tools that give each element its own site keep mistaking known disordered phases for new ordered compounds. As of September 2026, the MatterGen paper's Nature page shows no correction.

Self-driving labs: the A-Lab and the dispute over its results

Making the material is the slow step, so the next move was to automate it. The A-Lab at Lawrence Berkeley National Laboratory, described in Nature on the same day as GNoME, is a self-driving laboratory for solid-state synthesis of inorganic powders. Its loop:

1. Targets are chosen from large-scale DFT phase-stability data from the Materials Project and Google DeepMind.
2. Natural-language models trained on the synthesis literature propose up to five starting recipes for each target.
3. Robots dose the precursor powders, heat the samples and measure the products by X-ray diffraction (XRD), with robotic arms moving samples between the stations.
4. Two ML models, trained on structures from the Materials Project and the ICSD, identify the phases in each XRD pattern, and automated Rietveld refinement confirms them.
5. If the target yield is below 50%, an active-learning algorithm grounded in thermodynamics proposes a better recipe, and the loop runs again.

Berkeley Lab announced that over 17 days the A-Lab produced 41 new compounds from 58 attempted. In March 2024, chemists from Princeton University and University College London re-examined the products in PRX Energy. They found four recurring shortfalls in the analysis, concluded that no new materials had been discovered, and judged that two thirds of the claimed successes were likely known, compositionally disordered versions of the predicted ordered compounds. Their main technical finding was that automated Rietveld analysis of powder XRD data is not yet reliable.

The authors published a correction in January 2026. "Novel", they wrote, had meant new to the prediction platform, not necessarily new to science. A manual re-analysis of the diffraction patterns, peer-reviewed after publication, confirmed 36 of the 40 reported successes and found 4 inconclusive, and one compound that had mistakenly been included in the training data was removed from the discussion. The paper now reports 36 compounds from 57 targets.

What survived is still useful: a robotic loop that ran for 17 days and made 36 of its targets, as confirmed by people reading the data, and a record of its failures. Seventeen of the 57 targets never formed, and the paper sorts them into four failure modes: slow reaction kinetics, which affected 11 of the 17; precursors that evaporated; products that came out amorphous; and inaccurate computations.

What limits AI in materials discovery

Four limits recur in the cases above.

Training data sets the ceiling

A model is only as good as its labels. The Matbench Discovery authors write that a universal potential's quality is circumscribed by the quality and level of theory of its training data. Their benchmark uses energies from PBE, one standard DFT approximation, and they call for large amounts of higher-fidelity data. Data can also be skewed: they suspect that the Materials Project's unusually high share of stable materials biases models towards low energies, so every model tested was more inclined to call an unstable crystal stable than the reverse, and all of them overestimated how many stable structures the test set held. And a model can "discover" what it was shown: both the MatterGen compound and the compound removed from the A-Lab paper had been in training data.

Disorder and temperature

As the Matbench Discovery authors note, high-throughput DFT often uses small unit cells, which can impose an artificial ordering of atoms that the real material, disordered by entropy at synthesis temperatures, does not have. The same calculations can settle on structures that are dynamically unstable, which a phonon calculation would reveal, but phonons computed with DFT are often too costly for screening. This is where fast phonon models such as MIT's can earn their place.

Disorder is the common thread in the critiques of GNoME, the A-Lab and MatterGen. The stability of an ordered cell at 0 K, which is what the models learn to predict, is not the question a crystallographer asks of a real powder.

Synthesis is the bottleneck

Computational screening proposes candidates far faster than labs can make them; the A-Lab was built to close that gap, and its record shows why the gap persists. Of its 16 failed targets that were predicted stable, 13 were stopped by experimental barriers rather than by the calculations. When people reground the products by hand and heated them hotter, steps outside the lab's active-learning algorithm, two more targets formed. How far machine learning can help predict synthesis routes is, in the Matbench authors' words, a key open question.

Validation: good numbers can mislead

Global error metrics such as mean absolute error, RMSE and R² can give misleading confidence. Matbench Discovery showed that an accurate model still produces many false positives when its predictions fall close to the stability line at 0 eV per atom above the hull, and each false positive costs lab time. The benchmark's answer is to test prospectively, on data generated by the discovery workflow the model will serve, and to score it on the decision it supports. Human checking still matters too: A-Lab's automated XRD analysis and MatterGen's novelty check both passed peer review before crystallographers challenged them.

Predicted, made or measured: how to read a claim about an AI-discovered material

Most confusion about AI-discovered materials comes from mixing levels of evidence. Place any claim on this ladder:

1. Predicted by a model. A score from a GNN or a potential. Fast and cheap, and prone to false positives near the stability line.
2. Computed stable by DFT. On or below the convex hull in a calculation. GNoME's 381,000 new crystals are at this level.
3. Checked against what is known. Compared with experimental databases such as the ICSD, disordered entries included. MatterGen's synthesized compound failed this step, according to the 2026 critique.
4. Made in a lab. The target phase formed, confirmed by XRD and a person's Rietveld analysis. The A-Lab's 36 compounds are here, and so are GNoME's 736 structures, made by other labs in work of their own.
5. Measured. The property itself was measured on the sample, as with MatterGen's 169 GPa against a 200 GPa target.
6. Useful. The material does a job in a device or process. This is the third leg of Cheetham and Seshadri's test, and the one a stability calculation cannot supply.

Then ask three more questions. Was the model tested prospectively, on data it never saw? Are the code and data public, as MatterGen's are? Does the article page list a correction, as the A-Lab's does?

The same discipline applies when a company trains models on its own test or production data: the model can only be as good as the measurements it learns from. If yours sit in spreadsheets and instrument exports, fix that first. Our data platform service builds tested pipelines and governed datasets, with lineage, that serve dashboards and machine learning alike. Drug discovery follows the same pattern of cheap predictions and expensive proof; see the role of AI in drug discovery.