UC Berkeley

In many industrial applications, numerical simulations allow us to perform virtual experiments through computers by solving differential equations. However, it often requires us to put large amounts of computational resources, because solving differential equations is computationally expensive and time-consuming in general. This could be more critical when we need to run heavy simulations in real-time applications.

This work introduces a hybrid approach on how to accelerate numerical simulations by applying the fundamental idea of deep learning to the numerical simulations. Deep learning and numerical simulation have proposed two different ways for engineers and scientists to predict and understand the complex behavior of systems. While numerical simulation is a traditional technology that relies on the fundamental laws of nature, deep learning is an emerging technology that is highly data-driven.

In the first half of this dissertation, I will review a basic background on deep learning and nonconvex optimization to help readers easily understand the fundamental concepts. In the second half, I will introduce the two engineering problems on accelerating numerical simulations with both reduced simulation costs and desirable accuracy. The first problem is in rapid process control of multiphase flowing foods. The second problem is to optimize the tool path in the Selective Laser Sintering process.