Input-Space Scientific Machine Learning for PDE-Constrained Optimization of Geometries

Explore input-space scientific machine learning approaches for PDE-constrained optimization through this technical seminar from Lawrence Livermore National Laboratory's DDPS series. Learn about Physics-Enhanced Deep Surrogates (PEDS), an innovative input-space SciML model that accelerates simulation of PDE-driven properties by modeling the input space of PDE solvers rather than learning the PDE solutions themselves. Discover the HiLAB optimization technique that enables efficient exploration of high-dimensional design spaces for PDE-constrained problems. Examine how these methods differ from traditional scientific ML approaches that attempt to relearn physical laws already encoded in PDEs, and understand the advantages of leaving physics computations to numerical solvers while using machine learning to enhance the input modeling process. See practical applications through benchmark PDE problems and experimental validation in nanophotonic metasurface design, demonstrating how input-space scientific machine learning can improve both simulation and optimization in engineering design contexts governed by partial differential equations.