Accelerating heat exchanger design by combining physics-informed deep learning and transfer learning

Abstract

Recently developed physics-informed deep learning is regarded as a transformative learning philosophy that has been applied in many scientific domains, but such applications are often limited to simulating relatively simple equations and well-defined physics. Here, we propose a systematic framework that can leverage the capabilities of space decomposition, physics-informed deep learning, and transfer learning to accelerate the multi-objective stochastic optimization of a heat exchanger system. In particular, this method seamlessly integrates the strengths of the modified Fourier network for capturing steep gradient variation, the point density adjustment strategy to identify the appropriate size of residual points, as well as the accelerated linear algebra to allow for kernel fusion and just-in-time compilation that enables an acceptable computational expense. The performance is verified by discovering the best-performing geometric design and the corresponding optimal operating conditions of an air cooler system under uncertainty.