If you wanted to make the case that AI is going to be a net good for humanity, this is the page you would point to. Of all the deployments covered in this section, scientific research is the one where the wins are clearest, the harms are smallest, and the future possibilities are most exciting. It is also the one most readers know least about, because a tool that helps a structural biologist understand a protein gets less press than a chatbot that helps a teenager write an essay.

AlphaFold — the tipping point

The story of AI in science has a clear before and after, and the after begins with AlphaFold.

For fifty years, predicting the three-dimensional shape of a protein from its amino acid sequence — the "protein folding problem" — was one of the most-discussed unsolved problems in biology. Proteins are biology's workhorses; what they do is determined by how they fold; and folding is governed by physics so complicated that no analytical solution had ever been found. The community-wide test for protein-structure prediction was the biennial CASP competition, and for two decades nobody had cracked it. In CASP14, in 2020, DeepMind's AlphaFold 2 achieved accuracy levels that experimental crystallographers described as comparable to laboratory measurements.

The implications were immediate and enormous. DeepMind, in partnership with EMBL-EBI, released the AlphaFold Protein Structure Database in 2021. By 2022 the database contained predicted structures for over 200 million proteins — essentially every protein known to science. Before AlphaFold, getting a single protein structure required either X-ray crystallography (months of work, often unsuccessful) or cryo-electron microscopy (faster, still difficult). Now you look it up. The 2024 Nobel Prize in Chemistry went jointly to David Baker, Demis Hassabis and John Jumper for protein structure work, and the citation explicitly named AlphaFold.

AlphaFold 3, released in 2024, extended the model to predict interactions between proteins and other molecules — the actual mechanism of how drugs work, how enzymes catalyse reactions, how viruses bind to cells. The model is meaningfully less reliable for these interactions than for solo structures, but it is good enough to be useful, and the field is iterating fast.

Outside DeepMind, Meta's ESMFold and the open-source RoseTTAFold from David Baker's lab cover the same territory and have largely democratised the technology. A graduate student in 2026 with a laptop and an internet connection can do structural biology that would have required a major lab in 2018.

Materials discovery

The next big AI4Science deployment after AlphaFold has been in materials science. The conceptual problem is the same: there is an enormous space of possible compounds (combinations of elements, crystal structures, dopants, processing conditions), most of which have never been synthesised, some of which would have useful properties (semiconductors, superconductors, battery cathodes, catalysts, photovoltaics). Searching the space experimentally is prohibitively slow.

DeepMind's GNoME (Graph Networks for Materials Exploration), published in Nature in late 2023, predicted around 2.2 million novel stable inorganic crystal structures, of which about 380,000 met the criteria for likely synthesisability. In one paper, the catalogue of known stable inorganic materials roughly tripled. Lawrence Berkeley National Laboratory's A-Lab, an autonomous synthesis facility, has been working through the high-priority candidates, demonstrating that some of the GNoME predictions can in fact be synthesised in the lab.

The translation from "novel material predicted" to "battery you can buy" is a long road that still mostly involves traditional materials science. But the search-space pruning that AI provides has materially shortened the front end of that road. For the energy transition specifically — better batteries, better catalysts for hydrogen production, better photovoltaics — the next decade of progress is going to be substantially shaped by AI-assisted materials discovery.

Weather and climate

Numerical weather prediction has been one of the great success stories of supercomputing. The European Centre for Medium-Range Weather Forecasts (ECMWF) and the major national meteorological services have spent decades refining physics-based models that simulate the atmosphere on grids of millions of cells.

In 2023 DeepMind's GraphCast demonstrated that a machine-learning model trained on historical reanalysis data could match or beat the ECMWF's flagship physics-based model on most forecast skill metrics, while running about a thousand times faster on commodity hardware. NVIDIA's FourCastNet, Huawei's Pangu-Weather and Microsoft's Aurora have produced similar results. ECMWF itself has moved to a hybrid approach where ML models contribute alongside the physics-based system.

The implications go beyond efficient forecasting. The ML models can be ensembled cheaply, generating much larger probabilistic forecasts than physics-based ensembles. They can also be fine-tuned for specific applications — flood forecasting in particular catchments, fire weather in particular regions — in ways that would be infeasible with a physics-based model.

For climate (longer-term, the multi-decade trajectory rather than next week's weather), ML is so far playing a complementary role rather than replacing the global climate models. The physics-based GCMs are still authoritative for projection. ML is being used for downscaling — taking coarse GCM outputs and producing high-resolution local projections — and for parameterising small-scale processes (clouds, turbulence) that GCMs handle crudely. CSIRO and the Bureau of Meteorology in Australia are both active in this work.

Mathematics and theorem proving

Of all the AI4Science applications, the one that surprised mathematicians most was AI at mathematics. The Lean proof assistant — which formalises mathematical proofs as machine-checkable code — has been around for a decade and was largely a project of the formal-mathematics community. In 2024 Google DeepMind's AlphaProof and AlphaGeometry 2 achieved a silver-medal performance on the International Mathematical Olympiad problems, working in Lean.

The applications are more pedestrian than the headlines suggest. Most of mathematics is not olympiad problems; it is the slow accumulation of structure within established theories. AI assistance is now in regular use for routine theorem-proving steps in formal verification of software (especially in industries like aerospace and finance where formal correctness matters), in tactic suggestions for working mathematicians using Lean or Coq, and in generating conjectures from large datasets of known mathematical objects.

Terence Tao — Fields Medal winner, prolific blogger about working mathematics with AI — has documented in detail how he uses LLMs as a research collaborator: not to prove anything itself, but as a sounding board, a notation cleaner, a literature search tool, and a generator of plausible (sometimes wrong) intermediate steps. The pattern of "AI as research assistant" is now standard in many maths departments, and increasingly in physics and theoretical computer science too.

Astronomy and exoplanets

Astronomy was an early adopter of ML for the same reason as medical imaging: the data is visual, abundant, and labelled enough to train on. Kepler and TESS, the NASA exoplanet-hunting space telescopes, both now have ML pipelines that run on the raw photometric data and surface candidate planets for human follow-up. The 2017 paper by Shallue and Vanderburg used a CNN trained on Kepler data to identify two previously-missed exoplanets in known systems, and the technique has been refined considerably since.

Gravitational-wave astronomy has been similarly transformed. LIGO and Virgo's detection pipelines incorporate ML for both signal detection (filtering out instrumental noise) and parameter estimation (inferring properties of the merging compact objects). The cadence of confirmed gravitational-wave events has grown with each observing run, and the analysis pipelines that make that possible are heavily ML.

The Vera C. Rubin Observatory, scheduled to come online with full survey operations in 2025-2026, will produce so much data — about 20 terabytes per night, with millions of transient detections — that ML processing is not a bonus but a precondition for the science to happen at all. Australia is a participant in Rubin and the Square Kilometre Array, both of which have ML at the centre of their operational design.

Where it gets contested

Most of this page is unapologetically positive, because the use of AI in fundamental research is the place where the technology is most clearly net good. The contested points are smaller in scale:

Reproducibility. ML models published as part of scientific results are often hard to reproduce — training code, weights, and data are not always shared, and even when they are, the computational requirements may be prohibitive. The community has been pushing hard on reproducibility standards but the practice is uneven.

Hype and overclaiming. Every research paper with "AI" in the title gets more press than it deserves. The actual rate of progress in most subfields is slower and more incremental than the headlines suggest. AlphaFold is a genuine paradigm shift; many of its imitators are not.

Concentration of capability. The largest models can only be trained by organisations with access to the right hardware. That has consolidated leadership in a small number of labs (DeepMind, Meta FAIR, OpenAI, the major US universities), and the gap between what those labs can do and what a regional university can do has grown. Counter-pressure exists in the form of open-source releases, but the centre of gravity has shifted.

The honest summary

If you want to understand why some of the smartest people working in AI are also the most optimistic about it, this is the page to read. The technology has, in less than a decade, transformed structural biology, opened up materials discovery, given working mathematicians a new collaborator, and put a substantial fraction of the world's photometric astronomy through ML pipelines. The next decade looks likely to do the same for chemistry, neuroscience, ecology, and several disciplines we have not anticipated. None of this gets the press that ChatGPT does, but a hundred years from now it will probably be the part of the AI story that mattered most.