“AI for science” names a recurring pattern: instead of computing answers by simulating physics step by step or searching a space by brute force, a deep-learning model learns to predict the answer directly. The model is trained on existing data - solved structures, decades of weather records, known stable materials - and then generalizes to cases nobody has computed before, usually far faster than the conventional method.
The clearest examples cluster around DeepMind. AlphaFold predicts protein structure from sequence, a problem that previously needed slow lab experiments (AlphaFold 2, 2020), and AlphaFold 3 extends this to how molecules interact (2024). GraphCast forecasts weather and beats the leading conventional system on most metrics while running in under a minute (2023). GNoME predicts millions of new candidate materials (2023). AlphaGeometry and AlphaProof produce machine-checkable mathematical proofs near medal level at the olympiad (2024).
The common shape is the same in each case. There is a hard target - a structure, a forecast, a stable crystal, a proof - and a conventional method that is slow, expensive, or both. A learned model is trained on examples, then either replaces the slow method outright or narrows its search to the most promising candidates. The payoff is speed and scale: predictions that once took months or supercomputer-hours arrive in seconds, so researchers can screen orders of magnitude more possibilities.
What makes the pattern significant is that these are not games or text tasks. They are problems with direct real-world stakes in medicine, climate, energy and mathematics, and in several cases the AI matched or beat methods that scientists had spent decades refining.