Teaching AI to build scientific models

The evolution of scientific modeling

Scientific modeling has historically relied on a three-step process: building a model based on theoretical assumptions, constraining it to fit experimental data, and interpreting the results to gain insight. While human scientists like Hodgkin and Huxley historically performed these steps manually—often requiring weeks of calculation—AI systems are now increasingly capable of automating these tasks. The primary challenge remains that a model which merely replicates data is insufficient; true scientific progress requires understanding why a model works and which features are essential versus incidental.

Constraining models with simulation-based inference

Constraining models—the process of tuning parameters to match experimental measurements—is the area where AI has made the most significant progress. Traditional methods often struggle with complex systems where multiple parameter combinations can produce identical results, hiding the underlying uncertainty.

Simulation-based inference (SBI) addresses this by utilizing neural networks to map model outputs back to the range of parameters that could have produced them. By running a simulator millions of times, researchers can train a model to perform inference in seconds. This approach is particularly powerful because it provides a probability distribution of plausible parameters rather than a single "best fit," allowing scientists to identify where data is well-constrained and where further experimentation is required.

Automating model construction

Beyond parameter tuning, AI agents are now being used to design the structure of models themselves. By leveraging large language models (LLMs), these agents can propose candidate equations based on natural-language descriptions and experimental data, iteratively refining them until they accurately replicate observations.

However, this capability introduces a new risk: an agent that returns only one successful model may obscure other equally valid alternatives. To maintain scientific rigor, researchers advocate for treating model building like model constraining—generating a distribution of plausible models and weighting them based on their natural fit to the data. This allows scientists to see which features are shared across all successful models, providing a more robust foundation for discovery.

The challenge of interpretation

Interpretation remains the most difficult step to automate because it requires moving from numerical accuracy to conceptual understanding. A model is considered "intelligible" if it allows researchers to deduce consequences without running the simulation. While AI can assist by generating code that reveals underlying design principles, the ultimate goal is to provide human scientists with an account of which parts of a model are pinned down by data and which alternatives were ruled out during the process.

Scaling scientific discovery

As AI tools lower the barrier to building and testing complex models, the scale of scientific inquiry is expanding. Researchers are now applying these methods to systems previously considered too complex for manual modeling, such as the neural architecture of entire brains. In these high-dimensional spaces, the objective shifts from finding a single "correct" model to identifying the properties shared by all data-compatible models. The future of scientific AI lies in systems that do not just provide answers, but communicate the process of discovery, ensuring that the insight gained remains accessible to human researchers.