Numerical homogenization

Neural networks for multiscale PDEs, and approximation rates from the Kolmogorov superposition theorem.

From 2019 to 2023 I worked in the Research Engineering Laboratory for Advanced Computational Science with Matthew Knepley on scientific machine learning for high-dimensional and multiscale problems.

One project solves multiscale partial differential equations with neural networks, using the localized orthogonal decomposition method as the numerical homogenization scheme. I presented this as “Machine Learning Methods for Numerical Homogenization” at the SIAM Conference on Computational Science and Engineering in 2023.

A second thread is approximation theory. We worked on dimension-independent approximation rates for Lipschitz functions by deep ReLU networks, using a modified Kolmogorov superposition theorem, and on whether the inner function of that theorem admits a Lipschitz reparameterization. I also wrote a C implementation of the theorem for dimension reduction.