Abstract
Physics Informed Neural Networks (PINNs) merge neural networks with underlying physics laws on solving differential equations and other numerical problems which is used in scientific computing and engineering fields. However, numerous applications employing numerical methods have been deployed on edge devices, necessitating dedicated computational units for efficient execution of high-performance PINNs.
This research shows the execution of a PINN on a hardware device for the real-time application on the resource-sensitive device. The approach optimizes hardware utilization through a piece-wise non-linear approximation. [1] To demonstrate the effectiveness of this method, the PINN is implemented on a Field Programmable Gate Array (FPGA) and designed for working out a non-linear ordinary differential equation (ODE). The results indicate that the PINN on hardware can achieve a minimum accuracy of 95% by comparing to the actual solution. The method using High Level Synthesis is also shown in the paper, which highlights the parallel and pipeline structure achieved on hardware platform. The proposed methods suggest that an edge device can utilize the PINN with the appropriate hardware unit.