Abstract
Differential equations are involved in modeling many engineering problems.Many efforts have been devoted to solving differential equations. Due to theflexibility of neural networks, Physics Informed Neural Networks (PINNs) haverecently been proposed to solve complex differential equations and havedemonstrated superior performance in many applications. While the L2 lossfunction is usually a default choice in PINNs, it has been shown that thecorresponding numerical solution is incorrect and unstable for some complexequations. In this work, we propose a new PINNs framework named Kernel Packetaccelerated PINNs (KP-PINNs), which gives a new expression of the loss functionusing the reproducing kernel Hilbert space (RKHS) norm and uses the KernelPacket (KP) method to accelerate the computation. Theoretical results show thatKP-PINNs can be stable across various differential equations. Numericalexperiments illustrate that KP-PINNs can solve differential equationseffectively and efficiently. This framework provides a promising direction forimproving the stability and accuracy of PINNs-based solvers in scientificcomputing.