AI for Science

AI for science is the use of machine learning as an instrument of scientific discovery rather than as a consumer product: predicting molecular structures, screening candidate materials, emulating expensive simulations, fitting equations to data and mining the literature for ideas. The common thread is that a learned model stands in for something slow or intractable, an experiment, a physics simulation or a human literature search, and the scientific question is then how far that stand-in can be trusted outside the data it learned from.

The clearest success is protein structure prediction: AlphaFold2, described by Jumper and colleagues in 2021, predicted three-dimensional protein structures from amino-acid sequences at accuracy often comparable to experiment, and its predictions are now routinely used by biologists. The achievement has its own entry in the glossary, but it set the template for the field: a well-defined prediction task, decades of curated public data (the Protein Data Bank), a community benchmark and a model with the right inductive structure. Not every science problem offers all of those, which is why AlphaFold is better regarded as a milestone than as a forecast that every field will be transformed at the same rate.

Materials discovery shows both the promise and the caution. In November 2023 researchers at Google DeepMind published GNoME (graph networks for materials exploration) in Nature, reporting that active-learning loops combining graph neural networks with density functional theory calculations had predicted about 2.2 million new crystal structures, of which roughly 380,000 were judged the most stable. A related autonomous laboratory at Lawrence Berkeley National Laboratory, called A-Lab, reported synthesizing a number of the predicted compounds. Independent materials scientists then questioned how many of the claimed compounds were truly novel, and how useful they were, as some were disordered or chemically unusual variants of known materials. The episode illustrates that a predicted stable structure is a computed claim about energy, and that synthesis, characterization and a useful property are separate steps.

Weather modeling is another success, with models like GraphCast learning to forecast from decades of reanalysis data at a tiny fraction of the computing cost of physics-based simulation. Climate and fluid dynamics more generally use learned surrogates for expensive solvers, and physics-informed neural networks add known equations to the training loss, though their advantage over conventional numerical methods is problem-dependent and sometimes disappointing.

Symbolic regression searches for a mathematical expression that fits data, aiming for a formula a scientist can read rather than a black-box network. Schmidt and Lipson published an influential genetic-programming approach in Science in 2009 that recovered conservation laws from motion data; Udrescu and Tegmark's AI Feynman (2020) combined neural fitting with physics-inspired simplifications to rediscover equations from the Feynman Lectures, and open-source tools such as PySR are widely used. Hypothesis generation is the fuzziest cousin: literature-based discovery began with Don Swanson's 1986 proposal linking fish oil and Raynaud's syndrome through two disjoint bodies of papers, and modern language models can suggest connections across millions of articles. The pitfall is that suggestions are cheap and testing is expensive, so a system that generates many plausible ideas shifts the bottleneck to validation rather than removing it.

Reproducibility is the persistent caveat. A 2023 survey by Sayash Kapoor and Arvind Narayanan in the journal Patterns found data leakage, where information from the test set contaminates training, in papers across many fields that use machine learning, and that in many cases it inflated reported performance. Other recurring problems include benchmarks on narrow distributions, tuned baselines for the new method but untuned ones for competitors, unreleased code and data, and models that learn artifacts of how the data were collected. A common misconception is that high accuracy on a held-out split shows scientific understanding; a model can predict well without any mechanism, and a surrogate fails silently outside its training domain.

Read AI for science as a way of making slow scientific steps cheap enough to search at scale, from folding proteins to screening crystals to emulating the atmosphere, with the strict condition that outputs are hypotheses and surrogates until experiments, leakage-free evaluation and independent replication say otherwise.