AI is rapidly transforming science

In 2021, Google DeepMind published results for AlphaFold, an AI model that essentially "solved" the protein folding problem, and then they went on to publish the predicted 3-D structures for the 200M or so proteins so far identified by science. This is arguably the most important AI contribution to date, with immediate impact on how biological and pharmacological research is now conducted.

This week, DeepMind reported another AI-based scientific advancement which may turn out similarly impactful, this time for the field of material science. In an article in Nature, "Scaling deep learning for materials discovery", they report on contributions from their GNoME system (Graph Networks for Material Exploration) which has so far proposed 2.2 million new stable inorganic new crystal structures, a substantially increase from the previously know 48,000 such structures. They filtered these to 381,000 new inorganic compounds to add to the Materials Project database of known compounds, corresponding to the compounds living on the "convex hull" of known stable compounds. As GNoME explores the immense space of describable chemical structures and explores paths that don't work out, it actively learns from its mistakes to get better and better at finding stable compounds.

Many of these new materials are likely to be useful for rechargeable batteries, photovoltaics and computer chips -- fields that could see big improvements from better materials.

Both these examples highlight a shift in what we consider to be a scientific advancement. These systems discover complex regularities in nature that are encoded in large neural networks, but are not reduced to simple, understandable rules that mere humans can understand. Yet, like traditional scientific knowledge, this new form of scientific discovery can still be leveraged to engineer new real-world solutions.