The recent publication of an unprecedented number of new crystal structures predicted by AI has attracted both praise and criticism. According to Bernd Ensing, this definitely means that AI has now reached the field of (in)organic chemistry.
‘I have yet to see the first successful prediction of a chemical product or a synthesis route by an AI algorithm that I would not have come up with myself’, a colleague recently told me, somewhat sceptical about the ongoing AI revolution. Indeed, in most (in)organic laboratories, AI does not seem to have made much headway—yet. But for how long?
Last November, Google Deepmind presented their GNoME project in Nature, in 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’, raved technical reviews in Wired, Science and Time, among others.
‘The explosive increase in scale begins to sink in’
The number of man-made inorganic crystal structures to date is around 20,000. Add to that another 28,000 computationally predicted materials yet to be made, and the explosive increase in scale of the new AI algorithms begins to sink in. Deepmind has ‘only’ added the 381,000 most stable structures to the International Crystal Structure Database (ICSD).
Why can both GNoME and MatterGen generate so many more new structures and, more importantly, so much faster than before? GNoME generates new materials by making variations on existing structures, for example by substituting elements, and determining their energy using NequiP, a well-known neural network that can predict quantum chemical (DFT) energy. The lower the energy, the more stable the material. MatterGen is based on another AI algorithm, ‘denoising diffusion’, which has been used successfully in photorealistic and artistic image creation software such as Dall-E and Midjourney. An added advantage of this technique is that it can be used to create structures with specific properties, such as a desired symmetry or density.
Predicting new materials is one thing; actually making them is another challenge. Deepmind gave 58 predicted GNoME structures to a fully automated robotics lab at the Lawrence Berkeley National Laboratory, which managed to synthesize 41 of them (although some reviews raise concerns). But predicting hundreds of thousands of structures and then trying to make them is simply not practical. So the next challenge for AI is to generate a select set of usable materials with optimal properties. Nevertheless, GNoME and MatterGen represent a significant turning point in materials development that should not be underestimated. It is an indication of what we can expect from AI in chemistry in the (near) future. Still, some colleagues may remain sceptical about the usefulness of AI in their field for a long time to come…
Bernd Ensing is associate professor of computational chemistry and scientific director of the AI4Science Laboratory at the University of Amsterdam.
Want to respond? Send a message to firstname.lastname@example.org