AI for Inorganic Materials: What Google and Berkeley Show
Yes—AI can help predict promising inorganic materials and guide researchers toward making them, but it cannot guarantee that a predicted substance will be stable, practical, or easy to manufacture. The Google DeepMind–UC Berkeley collaboration illustrates this potential: its GNoME system predicted many candidate crystal structures, while laboratory work reported in related research confirmed a subset of newly made materials.
What the Google–Berkeley work demonstrated
GNoME, short for Graph Networks for Materials Exploration, is a machine-learning system developed by Google DeepMind to identify potentially stable crystal structures. Its results were described in a 2023 Nature paper. The team reported 2.2 million candidate structures, including about 380,000 predictions judged to be stable relative to competing known phases under the computational criteria used.
Those numbers describe predictions, not millions of materials that were synthesized or put to use. “Stable” in this context generally means that a structure is predicted to be thermodynamically favorable against relevant alternatives, often expressed through its position on a calculated convex hull. It does not automatically mean the material can be produced reliably, will remain unchanged in air or water, or has a useful property.
A separate experimental effort led by researchers at UC Berkeley and Lawrence Berkeley National Laboratory reported making 41 of 58 selected candidate materials. This is an important proof that AI-generated predictions can help direct laboratory discovery. It is not a universal 71% success rate: the researchers selected candidates for experiments, and results depend on the materials, methods, and definition of success.
How AI predicts crystal structures
Inorganic solids are often made of atoms arranged in repeating three-dimensional patterns. A crystal structure specifies which elements are present and how their atoms are positioned. Finding a new composition is not enough: researchers also need a plausible arrangement that is energetically favorable and experimentally accessible.
GNoME uses graph neural networks. In a graph representation, atoms are nodes and relationships between nearby atoms are edges. The model learns patterns from structures and calculated energies in existing materials databases, then estimates the energy of proposed structures. Lower-energy candidates are generally more promising, although the exact comparison depends on competing phases with the same overall composition.
The system used two complementary search approaches. One explored compositions related to known materials by substituting elements, while another searched more broadly for structures. Predictions were fed through computational checks and further refinement. This is a search-and-screening pipeline, not an oracle that simply names a useful material.
From a prediction to a real material
A computationally promising structure has to pass several further tests. Scientists may calculate its stability more accurately, examine its electronic, magnetic, or mechanical properties, and decide whether a plausible synthesis route exists. Then chemists attempt to make it, often adjusting temperature, pressure, starting compounds, and reaction time.
- Generate candidates: A model proposes compositions and crystal arrangements worth investigating.
- Check calculations: Researchers compare energies and assess whether the structure remains plausible under more detailed computational methods.
- Plan synthesis: Experimentalists select precursors and conditions suited to the elements and structure.
- Make and identify the sample: Laboratory measurements, such as diffraction, help determine whether the expected crystal formed.
- Measure useful properties: A confirmed structure still needs testing for the property that motivated interest in it.
Failure at one stage does not necessarily mean the model is useless. A candidate may be hard to synthesize with the chosen recipe, form only as a mixture, or require conditions not attempted. Conversely, making a predicted phase does not prove it is commercially viable or suitable for a device.
Why the predictions matter—and what they do not mean
Traditional materials discovery can be slow because the number of possible element combinations and atomic arrangements is enormous. AI can prioritize a manageable set of candidates, reduce some repetitive computational work, and reveal patterns researchers may not have searched by hand. The value is in focusing experiments, not eliminating them.
The headline count also needs careful interpretation. A predicted structure is a theoretical candidate; a synthesized material is a sample produced in a laboratory; a useful material has a measured property that meets a real need; and a deployable material must also be reproducible, safe, durable, and manufacturable. These are distinct milestones.
GNoME’s results also build on earlier materials science rather than replacing it. The model drew on prior structural and computational knowledge, and its predictions were assessed using established energetic calculations. Experimental confirmation remains the crucial independent test. Follow-up work can still revise stability estimates as methods, databases, or measurements improve.
Common limitations and misconceptions
- Predicted stability is not practical stability: A material may be thermodynamically favorable in a calculation yet degrade, react, or be difficult to isolate in ordinary conditions.
- Stability is not performance: A stable crystal is not necessarily a better battery material, catalyst, magnet, or semiconductor. Each application needs different measurements.
- Prediction is not synthesis: A model does not supply a working recipe automatically. Chemical kinetics, impurities, and experimental conditions matter.
- Large candidate totals are not product counts: The millions reported by GNoME are computational structures, not millions of verified, useful materials.
- Training data shape results: Models learn from the information available to them. Unusual chemistries or structures may be harder to predict accurately.
There is also a difference between a material being “new” in a model’s database and being entirely unknown to science. Database coverage is incomplete, and a candidate may have appeared in older literature under different conditions or description. Researchers therefore compare predictions with existing records and experimental evidence.
Where AI-guided materials research could go next
The approach is relevant wherever composition and structure can be searched systematically, including possible materials for energy storage, electronics, catalysts, and specialized magnets. But the GNoME demonstration is not evidence that every predicted candidate will serve one of these uses. Property-specific predictions and experiments are needed for each application.
A likely direction is tighter integration among models, high-throughput computation, and automated or semi-automated laboratories. Such systems could propose candidates, run calculations, guide synthesis, and return measurements that improve later predictions. Even then, human researchers remain important for choosing meaningful targets, interpreting unexpected results, and judging whether a candidate solves a real engineering problem.
Frequently Asked Questions
Can AI actually create a new inorganic material?
AI can propose a composition and crystal structure and help select experiments. A material counts as experimentally made only after researchers synthesize and characterize a sample; the algorithm itself does not manufacture it.
Did Google and Berkeley make all the materials GNoME predicted?
No. GNoME generated millions of computational candidates. The associated experimental work reported synthesis of 41 selected candidates from 58 attempted, not all predicted structures.
What does “stable” mean in the GNoME results?
It refers mainly to calculated thermodynamic stability compared with competing phases. It does not guarantee resistance to moisture, heat, or chemical reactions, and it does not establish useful performance.
Can these AI predictions lead to better batteries?
Possibly, but a candidate must first be shown to have relevant properties and work under battery operating conditions. The GNoME results establish a discovery method, not a finished battery technology.
Why do scientists still need laboratory experiments?
Calculations simplify real materials and rely on models and assumptions. Experiments determine whether a candidate can actually be made, identify what formed, and measure its properties in the real world.
Related
Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.