Bayesian differential programming for robust systems identification under uncertainty.

Area of Science:

• Computational Science
• Machine Learning
• Dynamical Systems Theory

Background:

• Accurate modeling of nonlinear dynamical systems is crucial across scientific disciplines.
• Traditional methods struggle with noisy, sparse, and irregular observational data.
• Discovering interpretable and parsimonious models from complex data remains a challenge.

Purpose of the Study:

• To develop a machine learning framework for Bayesian systems identification.
• To enable efficient inference of model parameters and quantify uncertainty.
• To facilitate the discovery of interpretable dynamical system models from limited data.

Main Methods:

• Utilizes differentiable programming to integrate ordinary differential equation solvers with gradient propagation.
• Employs Bayesian inference with Hamiltonian Monte Carlo sampling for parameter estimation.
• Incorporates sparsity-promoting priors to discover parsimonious latent dynamics.

Main Results:

• Demonstrates effective Bayesian systems identification from noisy, sparse, and irregular data.
• Achieves efficient inference of posterior distributions with quantified uncertainty.
• Successfully discovers interpretable and parsimonious models for various nonlinear systems.

Conclusions:

• Presents a flexible and robust workflow for data-driven model discovery under uncertainty.
• Highlights the power of differentiable programming and Bayesian inference for complex systems.
• Provides a valuable tool for advancing scientific understanding through data-driven modeling.