Zhao research group

Data-Driven Discovery

Learning physics from image data

A flowchart starting from image data, to optimization and Bayesian inference, to the discovery of constitutive laws

Microscopy generates vast amounts of data that are crucial for advancing basic science. However, a significant gap remains between these large datasets and the discovery of underlying physical laws in complex systems like biology. With the development of quantitative, high-throughput microscopy and simulations, alongside rapid advances in data science, we are at a unique moment in time. This convergence allows us to envision learning biophysics directly from image data and enabling multiscale modeling of soft and living matter.

A schematic showing the process of learning equations from experimental and simulation data

To achieve this goal, we use PDE-constrained optimization, Bayesian inference, and machine learning to discover continuum equations and constitutive laws (e.g. hydrodynamic equations of active matter, free energy landscape, reaction kinetics, etc.) from experimental images and molecular simulations.

Publication:

Inverse design and optimal control

A schematic going from engineering design space, predictive model, to target design

Integrating differentiable physical models of soft and living matter with PDE-constrained optimization opens new avenues for the translation of basic science into practical applications in biotechnology and adaptive materials. Examples include designing microstructure via multiphase and multi-material soft matter, controlling pattern formation in active matter by engineering microscopic interactions or applying external fields like light via optogenetic control.