Abstract
Conference Title: 2023 IEEE 16th International Symposium on Embedded Multicore/Many-core Systems-on-Chip (MCSoC)Conference Start Date: 2023, Dec. 18 Conference End Date: 2023, Dec. 21 Conference Location: SingaporePhysics-Informed Neural Networks (PINNs) are integrated with fundamental physical principles to solve complex differential equations relevant to scientific computation and engineering disciplines. As edge computing platforms increasingly deploy applications reliant on numerical equations, a growing necessity emerges for specialized computational modules that execute PINNs efficiently and with high performance. In this study, the effectiveness of various approaches is demonstrated through the implementation of a PINN on a Field Programmable Gate Array (FPGA) to address a nonlinear Ordinary Differential Equation (ODE) corresponding to the Reynolds equation. High-Level Synthesis (HLS) is investigated for real-time applications on resource-sensitive devices. Both parallel and pipeline computing techniques are employed in the approach. An alternative method of implementation involves the direct use of Hardware Description Language (HDL) on hardware platforms, optimizing hardware utilization via piece-wise nonlinear approximation. Experimental results indicate that the hardware-implemented PINN achieves an accuracy of 95% in comparison to the actual solution. It is suggested that edge devices can efficiently employ PINNs when paired with appropriate hardware units.