H.B. Keller Colloquium

Building models for biological, chemical, and physical systems has traditionally relied on domain-specific intuition about which interactions and features most strongly influence a system. Alternatively, machine-learning methods are adept at finding novel patterns in large data sets and building predictive models but can be challenging to interpret in terms of or integrate with existing knowledge. Our group balances traditional modeling with data-driven methods and optimization to get the best of both worlds. Recently developed for and applied to dynamical systems, sparse optimization strategies can select a subset of terms from a library that best describes data, automatically interfering potential model structures from a broad but well-defined class. I will discuss my group's development and application of data-driven methods for model selection to 1) recover chaotic systems models from data with hidden variables, 2) discover models for metabolic and temperature regulation in hibernating mammals, and 3) model selection for differential-algebraic-equations including chemical reaction networks and electrical grids. I'll briefly discuss current preliminary work and roadblocks on these topics.