‘AI represents a sea change in materials development that should not be underestimated’

‘I have yet to see a successful prediction of a chemical product or synthesis route by an AI algorithm that I wouldn’t have thought of myself’, a colleague remarked to me sceptically recently about the rampaging AI revolution. In most (in)organic laboratories, AI does indeed still appear to be gaining little traction. But for how much longer?

Last November, Google DeepMind presented their GNoME project in Nature, with which the developers discovered the crystal structures of 2.2 million new inorganic materials. Less than a week later, Microsoft AI4Science followed on ArXiv with MatterGen, generating 1 million new materials. ‘A true breakthrough and a revolution for the development of new materials’, gushed technical reviews in, amongst others, Wired, Science and Time.

‘The explosive scaling is beginning to sink in’

The number of inorganic crystal structures made by humans to date is approximately 20,000. Add to that a further 28,000 computationally predicted but not yet synthesised materials, and the explosive scaling of the new expansion by AI algorithms begins to sink in. DeepMind has ‘merely’ added the 381,000 most stable structures to the International Crystal Structure Database ICSD.

More stable

Why can GNoME and MatterGen produce so many more, and above all so much faster, new structures than was previously possible? GNoME generates new materials by creating variations on existing structures — for instance by substituting elements — and determining their energy using NequiP, a well-known neural network capable of predicting quantum chemical (DFT) energy. The lower the energy, the more stable the material. MatterGen is based on a different AI algorithm, ‘denoising diffusion’, which found success in programmes for generating photorealistic and artistic images such as Dall-E and Midjourney. An additional advantage of this technique is that it can create structures with specific properties, for example a desired symmetry or density.

Robot laboratory

Predicting new materials is one thing; actually making them remains a considerable challenge. DeepMind passed 58 predicted GNoME structures to a fully automated robot laboratory at Lawrence Berkeley National Laboratory, which managed to synthesise 41 of them (though some caveats have emerged regarding those results). But predicting hundreds of thousands of structures and then attempting to make them is simply not practical. The next AI challenge is therefore to generate a select set of viable materials with optimal properties. Nevertheless, GNoME and MatterGen represent a shift in materials development that should not be underestimated — one that is moreover indicative of what we can expect further from AI in chemistry in the (near) future. That said, some colleagues may remain sceptical for quite some time yet about the value of AI in their field…