Interfacial conditioning in physics informed neural networks

Physics informed neural networks (PINNs) have effectively demonstrated the ability to approximate the solutions of a system of partial differential equations (PDEs) by embedding the governing equations and auxiliary conditions directly into the loss function using automatic differentiation. Despite demonstrating potential across diverse applications, PINNs have encountered challenges in accurately predicting solutions for time-dependent problems. In response, this study presents a novel methodology aimed at enhancing the predictive capability of PINNs for time-dependent scenarios. Our approach involves dividing the temporal domain into multiple subdomains and employing an adaptive weighting strategy at the initial condition and at the interfaces between these subdomains. By employing such interfacial conditioning in physics informed neural networks (IcPINN), we have solved several unsteady PDEs (e.g., Allen–Cahn equation, advection equation, Korteweg–De Vries equation, Cahn–Hilliard equation, and Navier–Stokes equations) and conducted a comparative analysis with numerical results. The results have demonstrated that IcPINN was successful in obtaining highly accurate results in each case without the need for using any labeled data.