Date:
Monday, October 27, 2025 - 12:00pm
Location:
Elings Hall 1601 | Zoom: https://ucsb.zoom.us/j/8685211234
Speaker:
Colby Fronk
Abstract:
Scientists devote years to the model development cycle, which is the process of finding a model that describes a process, using data to fit parameters to the model, analyzing uncertainties in the fitted parameters, and performing additional experiments to refine and validate the model. This dissertation addresses these challenges by introducing machine learning–based tools that accelerate and automate key steps in the model development cycle, enabling the experimentalist to focus on what they do best: scientific discovery. First, we show how Neural Differential Equations can be used for data-driven modeling of time-series data and dynamical systems found in science and engineering. Building on this foundation, we introduce our state-of-the-art symbolic neural ordinary differential equation (neural ODE) tools for the symbolic regression of dynamical systems. To account for noise and uncertainty, we further develop Bayesian inference techniques for symbolic neural ODEs, enabling principled uncertainty quantification and more robust modeling. Finally, we propose new numerical methods for training stiff neural differential equations, addressing a critical challenge that has limited their wider adoption. Our novel stiff integration methods, built for compatibility with automatic differentiation, remove a major computational bottleneck and open new avenues for the use of neural ODEs in complex scientific problems.
Event Type:
General Event
