Scientific Discovery
Science advances when new instruments let us see what was previously out of reach. Our work on scientific discovery asks how machine learning can serve as such an instrument — and what it takes for its outputs to meet scientific standards of evidence.
- Scientific machine learning
- Computational biology
- Physics-informed learning
- Chemistry and materials modeling
- Simulation and surrogate models
- Scientific reasoning
- AI-assisted hypothesis generation
Why it matters
Many scientific problems are bottlenecked by computation: simulations that take weeks, search spaces too large to enumerate, or data too high-dimensional to inspect. Learned models can sometimes approximate these computations orders of magnitude faster, opening questions that were previously impractical to ask.
But scientific use imposes standards that ordinary benchmarks do not. Predictions must respect conservation laws and symmetries, come with honest error bars, and remain valid outside the regime they were trained on — or say clearly when they are not. Meeting those standards is a research problem in itself, and one that feeds back into the rest of machine learning.
Key research questions
How can known physical laws and symmetries be built into learned models without limiting what they can discover?
When can a learned surrogate safely replace a simulator, and how do we detect when it cannot?
What role can machine learning play in proposing hypotheses, as opposed to fitting data?
How should models represent biological and chemical systems that span many scales of space and time?
What does reproducibility mean for discoveries made with the assistance of learned models?
Primary subfields
- Physics-informed learning
- Models constrained by differential equations, symmetries, and conservation laws.
- Surrogate modeling
- Learned approximations of expensive simulations with quantified error.
- Computational biology
- Sequence, structure, and systems-level modeling of living systems.
- Chemistry and materials
- Representations of molecules and materials for property prediction and design.
- Hypothesis generation
- Using models to propose, rank, and design experiments for scientific claims.
Representative methodologies
Equivariant architectures
Networks whose outputs transform correctly under physical symmetries.
Simulation-based inference
Inferring parameters of scientific models from observed data.
Active learning
Choosing the most informative next experiment or simulation.
Benchmarking against established methods
Comparing learned approaches with domain-standard numerical and experimental baselines.
Relationships to other domains
- Machine Learning
Scientific data tests generalization far outside the training distribution.
- AI Safety
Uncertainty quantification is essential when predictions inform experiments.
- Artificial Intelligence
Experimental design is a planning problem under uncertainty.
Working on questions in Scientific Discovery?
We welcome conversations with researchers and organizations working on related questions — whether to compare notes, discuss a collaboration, or share work in progress.