Deep Learning Approaches for Learning Inverse Mappings
Simulation-based workflows form an integral part of design optimization and scientific analysis. As simulation models become progressively expressive, the computational complexity also increases. Consequently, solving inverse problems (e.g., design optimization, simulation-based inference, model exploration, sensitivity analysis) in a computation-efficient way becomes challenging. Deep learning has recently emerged as an effective way of learning the inverse mappings in a data-driven manner. Training a deep learning model as a surrogate model to learn the inverse mapping can be seen as a one-time investment. Once trained, the process of utilizing the model to predict what-if scenarios, plugging in design optimization workflows, etc. can be accelerated substantially. This project explores data-efficient deep architectures, in particular Bayesian neural network architectures for learning inverse mappings for several applications - simulation-based inference of gene regulatory networks, design optimization within Photonics, and inverse problems within particle physics.