AI in science: weather forecasting, materials and physics research
Artificial intelligence is transforming science by enabling faster, more accurate weather forecasts, discovering millions of new materials, and building on physics-based neural networks. Tools such as GraphCast, GNoME, and the work honored by the Nobel Prize in Physics illustrate these advances.
AI as a tool for scientific discovery
Scientific discovery often involves solving combinatorial puzzles with countless possible answers. Traditional methods can struggle with such complexity. Artificial intelligence offers a way to systematically generate and test hypotheses at scale. As Google DeepMind's science page explains, AI can decode these vast puzzles and address the root challenges that impede entire fields of scientific knowledge.
AlphaFold, also highlighted on that page, solved the protein-folding problem and is now used by over 3 million researchers. This success has inspired applications across life sciences, earth and sustainability, energy and materials, and algorithmic foundations.
Machine learning in weather forecasting
Conventional numerical weather prediction improves accuracy by adding computing power, but it cannot learn directly from historical weather data, as noted in the GraphCast paper. GraphCast is a machine learning method that uses past reanalysis data for training. It forecasts a wide range of weather variables on a global grid at 0.25 degree spacing for ten days ahead, completing the run in under one minute.
Compared to the most accurate operational deterministic models, GraphCast shows better skill on most of the evaluation targets, and it also improves forecasts of severe events such as tropical cyclones, atmospheric rivers, and temperature extremes. Operational centers like ECMWF are also adopting machine learning. The ECMWF AIFS blog documents the development of the AIFS, a machine-learning-based forecasting system that is being integrated with physics-based models.
Accelerating materials discovery with deep learning
Discovering new materials typically involves expensive trial and error. Graph networks, a type of neural network, are well suited to explore crystalline structures. The DeepMind materials blog describes GNoME, a deep learning tool that found 2.2 million new crystals, including 380,000 stable materials that are promising for technologies like superconductors and batteries.
GNoME uses an active learning loop with density functional theory calculations. It expanded the number of known stable materials from 48,000 to 421,000. External researchers have independently synthesized 736 of these predicted structures.
Physics foundations of neural networks
Many machine learning techniques are rooted in physics. John Hopfield created an associative memory that uses the physics of atomic spin, and Geoffrey Hinton devised the Boltzmann machine based on statistical physics. Their work, recognized by the Nobel Prize in Physics press release, laid the foundation for modern artificial neural networks.
These networks are used in physics to develop new materials with specific properties, as noted by the Nobel Committee. The methods from physics have enabled machine learning to become a powerful tool across science.
Measuring AI's scientific impact
To understand the overall influence of AI on science, independent measurement is essential. The Stanford HAI AI Index provides carefully vetted and globally sourced data on AI's technical progress, economic influence, and societal impact. It is a credible resource for policymakers and researchers.
The AI Index tracks trends that help evaluate how AI is transforming scientific research and other domains, offering a comprehensive view of AI's trajectory.
| AI tool | Field | Key result |
|---|---|---|
| GraphCast | Weather forecasting | Predicts 10-day global forecasts at 0.25° resolution in under a minute |
| AIFS | Weather forecasting | Operational machine-learning forecasting system at ECMWF |
| GNoME | Materials discovery | Found 2.2 million new crystals, including 380,000 stable ones |
| Hopfield network | Physics and AI | Associative memory based on atomic spin physics |
| AlphaFold | Life sciences | Predicted over 200 million protein structures |
Frequently asked questions
How does machine learning improve weather forecasting?
Machine learning models like GraphCast are trained directly on historical reanalysis data, allowing them to learn weather patterns without relying solely on physical equations. They can produce forecasts faster and often more accurately than traditional numerical models.
What is GNoME and why is it important?
GNoME is a deep learning tool that predicts the stability of crystalline materials. It discovered 2.2 million new crystals, including 380,000 stable ones, vastly expanding the known set of materials for potential technological use.
How does the Nobel Prize in Physics relate to AI?
The Nobel Prize in Physics honored John Hopfield and Geoffrey Hinton for using physics concepts, such as atomic spin and statistical physics, to develop artificial neural networks. These networks are fundamental to modern machine learning.
What role does the AI Index play in understanding AI's impact?
The AI Index, produced by Stanford HAI, provides unbiased, globally sourced data on AI's progress, including technical capabilities, economic influence, and societal impact, helping policymakers and researchers make informed decisions.
How do AI tools complement traditional scientific methods?
AI tools like GraphCast and GNoME can generate and test hypotheses at unprecedented scale, while traditional methods rely on physical laws or experimentation. The combination often leads to faster discoveries and improved accuracy.