Science increasingly produces more data than humans can analyze — from telescopes, particle detectors, genome sequencers and climate simulations. AI is becoming a core scientific instrument, helping researchers find patterns, make predictions and even propose hypotheses. In 2024, both the Nobel Prize in Physics (for foundational work on neural networks) and the Nobel Prize in Chemistry (for protein structure prediction and design) recognized AI-related work.

Weather forecasting

Traditional weather forecasts solve physics equations on supercomputers. Machine learning models trained on decades of historical weather data now rival or beat them on many measures. Google DeepMind’s GraphCast, published in Science in 2023, produced 10-day global forecasts in under a minute on a single machine and outperformed the leading operational system on most of the variables tested. Weather agencies, including the European Centre for Medium-Range Weather Forecasts, have since developed and deployed their own AI forecasting models.

Traditional models remain essential, both to generate the data AI models learn from and to provide physical consistency.

Biology

  • Protein structures: AlphaFold predicted structures for most known proteins — see our drug discovery explainer.
  • Genomics: models predict the effects of genetic variants and gene regulation.
  • Protein design: AI designs new proteins with specific functions, such as enzymes and binders.

Materials and chemistry

AI screens and proposes new materials for batteries, solar cells, catalysts and semiconductors. Researchers have reported AI-predicted stable crystal structures numbering in the hundreds of thousands, though turning predictions into synthesized, useful materials remains the challenge.

Astronomy and space

In 2017, Christopher Shallue of Google and Andrew Vanderburg used a neural network to search Kepler space telescope data and identified two new exoplanets, including Kepler-90i, making Kepler-90 the first known system besides our own with eight planets. AI now helps classify galaxies, detect transient events like supernovae, remove noise from images and search for gravitational lenses in huge sky surveys.

Physics

Machine learning helps filter collision events at particle accelerators, control plasma in fusion experiments and accelerate simulations.

Mathematics

AI systems have produced olympiad-level proofs, and mathematicians use AI to explore conjectures and formalize proofs in proof assistants.

Automated labs

“Self-driving laboratories” combine AI with robotics to plan experiments, run them, analyze results and decide what to try next — especially in chemistry and materials science.

Language models as research assistants

Scientists use language models to search and summarize literature, write analysis code and draft papers. Experimental “AI scientist” systems attempt to generate hypotheses and run computational experiments end to end.

Pitfalls

  • Reproducibility: results can depend on undisclosed data or code.
  • Data leakage between training and test sets can inflate reported performance.
  • Hallucinated references in AI-assisted writing.
  • Interpretability: predictions without understanding may not advance theory.
  • Access: large compute budgets concentrate capabilities in a few institutions.

How scientists can use AI responsibly

  • Share code, data and model details so others can reproduce results.
  • Separate training and test data carefully, including by time or location where relevant, to avoid leakage.
  • Validate AI predictions with experiments or independent data.
  • Disclose AI use in papers according to journal policies.
  • Verify every reference and quotation produced with AI assistance.

Who benefits

Open models and datasets — such as the freely available AlphaFold database and open weather models — let researchers at smaller institutions and in lower-income countries use cutting-edge tools without large computing budgets. Keeping key scientific AI resources open is therefore an important question of fairness in research.

What to watch next

  • AI weather models running operationally at more national weather services.
  • Experimental validation of AI-predicted materials.
  • Self-driving labs moving from demonstrations to routine research.
  • Journal policies on AI use in research and peer review.

Key takeaways

  • AI is becoming a core scientific instrument across disciplines.
  • GraphCast showed machine learning could outperform leading weather forecasts on many measures.
  • A neural network helped discover Kepler-90i, an eighth planet in the Kepler-90 system.
  • AI accelerates biology, materials, physics and mathematics research.
  • Reproducibility, data leakage and interpretability are key pitfalls.

Frequently asked questions

Are AI weather forecasts used today?

Yes. Several weather agencies run AI models alongside traditional ones, and some publish AI forecast products.

Can AI make scientific discoveries on its own?

AI has found patterns and candidates that led to discoveries, but scientists still design studies, validate results and interpret them.

Why did AI work win Nobel Prizes?

In 2024, the physics prize recognized foundational work on artificial neural networks, and the chemistry prize recognized protein structure prediction and computational protein design.