Parameter Estimation and Adaptive Solution of the Leray-Burgers Equation Using Physics-Informed Neural Networks

In this paper, we employ the Physics-Informed Neural Network (PINN) to estimate the practical range of the characteristic wavelength parameter(referred to as the smoothing parameter) $\alpha$ in the Leray-Burgers equation. The Leray-Burgers equation, a regularization of the inviscid Burgers equation, incorporates a Helmholtz filter with a characteristic wavelength $\alpha$ to replace the usual convective velocity, inducing a regularized convective velocity. The filter bends the equation's characteristics slightly and makes them not intersect each other, leading to a global solution in time. By conducting computational experiments with various initial conditions, we determine the practical range of $\alpha >0$ that closely approximates the solutions of the inviscid Burgers equation. Our findings indicate that the value of $\alpha$ depends on the initial data, with the practical range of $\alpha$ being between 0.01 and 0.05 for continuous initial profiles and between 0.01 and 0.03 for discontinuous initial profiles. The Leray-Burgers equation captures shock and rarefaction waves within the temporal domain for which training data exists. However, as the temporal domain extends beyond the training interval, data-driven forward computation demonstrates that the predictions generated by the PINN start to deviate from the exact solutions. This study also highlights the effectiveness and efficiency of the Leray-Burgers equation in real practical problems, specifically Traffic State Estimation.