Teaching AI to build scientific models (www.anthropic.com)

Researchers at the ELLIS Institute Tübingen and the University of Tübingen are leveraging AI to enhance scientific modeling, particularly in neuroscience. In a recent discussion, they reflected on the historical challenges faced by scientists—like Andrew Huxley and Alan Hodgkin, who relied on manual calculations to derive complex models from observations. Their Hodgkin-Huxley model, which describes the electrical activity of neurons, illustrates how scientific models can evolve through iterative testing and theoretical insight. The introduction of AI in this realm represents a significant advancement, as it now aids in the creation, constraint, and interpretation of models far beyond previous human capabilities. The core idea is to employ AI systems, particularly those based on large language models (LLMs), to automate and optimize this modeling process. For instance, simulation-based inference (SBI) allows researchers to explore a broad space of model parameters through extensive simulations, ultimately identifying plausible combinations that align with the data. This allows scientists to not only fit a model but also gain a deeper understanding of the underlying biological phenomena. Moreover, AI's ability to generate multiple candidate models increases the comprehensiveness of scientific exploration, moving beyond the limitations of singular explanatory frameworks. This paradigm shift in scientific modeling could lead to more reliable predictions and innovative insights across various disciplines.