AI Can Predict Millions of Stable Crystals. Making Them Is the Harder Problem.

6 min read

Key Claim: AI materials discovery has demonstrated the ability to generate and screen candidate structures at unprecedented scale, but the synthesis gap — the distance between thermodynamic stability on a computer and a material that can actually be made with the predicted properties — remains the field’s defining unsolved problem.

In January 2026, the journal Nature issued a correction to one of the most widely cited papers in AI-assisted materials science: the A-Lab study from Berkeley, which had claimed to autonomously synthesise 41 novel inorganic compounds in 17 days. A critique led by University College London chemist Robert Palgrave had found the compounds were largely already catalogued in the Inorganic Crystal Structure Database. The correction did not invalidate the engineering achievement of an AI-guided robotic synthesis system. It did, precisely, invalidate the central claim of novelty.

That episode captures where the field stands in early 2026. Two landmark AI systems — Google DeepMind’s GNoME and Microsoft’s MatterGen — have demonstrated that machine learning can identify and generate candidate crystal structures at a scale and speed no human team could match. The question of whether those predictions survive contact with a furnace, a crucible, and a characterisation instrument is a different one, and the answers so far are instructive rather than conclusive.

GNoME: Scale, Stability, and What “Discovery” Means

Google DeepMind published GNoME — Graph Networks for Materials Exploration — in Nature in November 2023, alongside the A-Lab synthesis paper. The model is a graph neural network trained using active learning, which evaluates crystal candidates by predicting formation energy and comparing it against the convex hull of known stable phases. Structures that land below the convex hull are classified as thermodynamically stable.

By that measure, GNoME is productive. It identified 2.2 million new crystal structures, of which 380,000 are ranked among the most stable and viable for attempted synthesis. The model improved the hit rate of stability predictions from roughly 1% in earlier computational screening to above 80% when full structural data is provided. (Google DeepMind Blog; Nature, 2023) The 380,000 most stable structures were added to the Lawrence Berkeley National Laboratory’s Materials Project database, making them available to researchers worldwide. (Berkeley Lab News Center)

Independent laboratories validated 736 of those predicted structures experimentally in concurrent work. That figure — 736 out of 2.2 million — is not a failure rate; it reflects the pace of experimental chemistry, not a model deficiency. The harder constraint is structural: GNoME predicts formation energy, not synthesisability. It does not model kinetic barriers, the reaction conditions required to reach a target phase, or the substitutional site disorder that commonly appears when a material is actually made. (Nature, 2023)

MatterGen: Generation Rather Than Screening

Microsoft’s MatterGen, published in Nature in January 2025, takes a different approach. Where GNoME screens and ranks candidate structures against stability criteria, MatterGen generates novel crystal structures from scratch using a diffusion model — the same probabilistic framework used in image generators such as DALL-E, adapted to three-dimensional crystallography. Critically, it can be conditioned on target properties: a researcher can specify a desired bulk modulus, magnetic moment, or electronic band gap and receive candidate structures designed to meet those constraints. (Nature, 2025; Microsoft Research Blog)

Benchmarked against prior generative models, MatterGen structures are more than twice as likely to be both novel and stable, and generate 2.9 times more stable, novel, unique structures than the CDVAE baseline. (Microsoft Research Blog)

The key experimental demonstration involved TaCr2O6, synthesised by collaborators at the Shenzhen Institutes of Advanced Technology after MatterGen generated the structure with a target bulk modulus of 200 GPa. The synthesised material measured 169 GPa — within roughly 15% of specification. However, the synthesised material also exhibited compositional disorder between the Ta and Cr sites that the model had not predicted, illustrating the gap between idealised crystal design and real-world synthesis outcomes. (Nature, 2025) The result is substantive — a property-conditioned design workflow reaching a physically coherent outcome in a real lab — but the disorder finding tempers the precision the model implies.

The caveats extend further. A 2025 ChemRxiv preprint identified TaCr2O6 as structurally identical to Ta1/2Cr1/2O2, a compound reported in 1972 and present in MatterGen’s own training data. The finding points to a persistent problem: generative models trained on known crystal databases can reproduce training examples under alternative naming conventions and present them as novel. (ChemRxiv, 2025) MatterGen is also bounded by its architecture — it was trained on structures with up to 20 atoms per unit cell, which excludes large or complex phases.

The Synthesis Gap: Where Predictions Fail

The mismatch between AI prediction and experimental outcome is not primarily a model accuracy problem. It is a category problem. Stability as computed — formation energy relative to the convex hull — is a thermodynamic concept. Whether a material forms in practice depends on kinetics: the reaction pathway, precursor availability, temperature profile, and competing phases. A structure can be thermodynamically stable and practically unreachable under any realistic synthesis condition.

The problem is sharper than it initially appeared. A December 2025 study led by the Fritz Haber Institute of the Max Planck Society and the University of Bayreuth found that in at least one screened AI-predicted materials database, more than 80% of the recommended candidates were likely to exhibit crystallographic disorder in experiment — meaning the actual material would have different, often worse, properties than predicted. (Fritz Haber Institute; Phys.org, December 2025) Disorder — where multiple elements partially occupy the same crystallographic site — is common in real inorganic compounds but largely absent from the idealised structures that train most models.

Efforts to close this gap are active but still early. Work published on arXiv in late 2025 proposed a synthesisability-guided pipeline that combines composition analysis with structural screening, successfully synthesising seven of 16 targeted structures — a better ratio than naive screening produces, though it still means the majority of model-selected targets do not materialise. (arXiv, 2025) Large language models trained specifically for synthesisability prediction are also under development, with early benchmarks showing significant improvement over thermodynamic screening alone. These tools are being layered onto the generative pipeline, but they are not yet embedded in GNoME or MatterGen. Notably, disorder prediction is a distinct problem from synthesisability prediction — a material can be synthesisable yet still exhibit the site disorder that invalidates the predicted properties — and this remains largely unsolved.

Autonomous Labs: What Closed-Loop Systems Can Do

The synthesis gap has pushed investment and research toward closed-loop laboratory automation, where AI planning, robotic execution, and instrument characterisation form a continuous feedback cycle. The rationale is not just speed but feedback density: autonomous systems can detect disorder and unexpected phases in real time and update predictions accordingly — precisely the kind of correction that static screening cannot provide. A Nature overview published in early 2026 documented that institutions including Berkeley, NC State, MIT, and several Chinese universities now operate autonomous labs synthesising dozens of new material candidates per week rather than a handful per year. (Nature, 2026)

Argonne National Laboratory’s RAPID (Robotic Autonomous Platforms for Innovative Discovery) facility conducted more than 6,000 experiments on organic redox-flow battery electrolytes in five months — a volume that would have taken years with manual methods. Published in the Journal of the American Chemical Society, the study identified a molecular stability ceiling: charged organic molecules degrade at higher voltages, placing a fundamental constraint on organic redox-flow battery energy density. (Argonne National Laboratory) The value of the automated system was not that it found a solution; it was that it characterised the ceiling quickly enough to redirect research strategy.

A multi-agent AI and robotics system published in January 2026 demonstrated a fully autonomous closed-loop workflow, with separate AI agents handling synthesis planning, robotic execution, and data interpretation. (Phys.org, January 2026) Berkeley Lab has separately described an initiative to build an AI assistant for energy materials characterisation, connecting automated synthesis with real-time spectroscopic analysis. (Berkeley Lab News Center, February 2026)

On the commercial side, Lila Sciences — founded by Flagship Pioneering and with MIT’s Rafael Gomez-Bombarelli as Chief Scientific Officer of Materials — raised $200 million in committed seed capital in March 2025 to build autonomous scientific labs spanning materials, chemistry, and life sciences. (Flagship Pioneering) MIT Technology Review noted in December 2025 that the sector has not yet produced a commercial material that originated from an AI prediction and reached production at scale. (MIT Technology Review, December 2025)

What to Watch

• Synthesisability models becoming first-class components. GNoME and MatterGen in their published forms score candidates on stability, not synthesisability. The next meaningful upgrade — already demonstrated at research scale — is embedding synthesisability prediction and precursor routing directly into the generation loop. When that becomes standard, the gap between prediction count and experimentally viable count should narrow.
• The disorder problem in database construction. The Fritz Haber Institute finding is an upstream issue: if the training databases themselves contain structures that are modelling artefacts of disorder, every model trained on them inherits that bias. Cleaning and re-annotating the Materials Project and ICSD for disorder is a materials informatics project, not an ML one, and it will take years.
• Closed-loop throughput vs. insight rate. Argonne’s 6,000 battery experiments produced one key finding (the stability ceiling) that redefined the research direction. The question is whether autonomous labs generate proportional scientific return — or whether high-throughput screening amplifies noise as readily as it finds signal.
• The commercial benchmark. The field will be judged not on prediction count but on whether any AI-originated material enters commercial production in battery electrodes, semiconductor substrates, or structural applications. No such case exists as of early 2026. Lila Sciences and the growing cohort of materials AI startups are explicitly targeting that first proof point.

Further Reading

• AlphaFold’s Path from Protein Structure to Clinical Pipeline
• Silicon Quantum Computing and the Logical Operations Milestone

This article was produced with AI assistance and reviewed by the editorial team.