Physics Informed Neural Network (PINN) and Differential Physics Simulator for Controlling Fluid Flow

The Compute System Architecture Unit (CSA) at imec is researching next-generation large-scale heterogeneous computer architectures for advanced HPC and AI applications. The team is responsible for algorithm research, performance modeling, architecture simulations, prototyping for future applications and the future systems to execute them on.

Specific focus of this internship will be applied AI/ML. The goal is to develop AI assisted simulations of fluid flow that is well-described by nonlinear partial differential equations (PDEs). Simulations of fluids are key to a wide range of areas and applications starting from weather prediction, engineering design of vehicles to cosmology and plasma physics. Numerical simulations of such systems can be a daunting task due to the huge computational cost of representing the finer details of fluid flow. Deep Neural Networks (DNNs) have emerged as one of the potential candidates in this context. However, traditional DNNs does not really guarantee the laws of physics. In this context, we would like to use Physics Informed Neural Networks (PINNs) that obey the underlying physical laws to achieve faster and accurate fluid simulations in presence of external control parameters.  Further, a combination of differentiable numerical simulation and neural networks will be investigated in the context of controlling fluid flow.

Image courtesy: github.com/benmoseley/harmonic-oscillator-pinn