Abstract
This study critically examines the performance of physics-informed machinelearning (PIML) approaches for traffic flow modeling, defining the failure of aPIML model as the scenario where it underperforms both its purely data-drivenand purely physics-based counterparts. We analyze the loss landscape byperturbing trained models along the principal eigenvectors of the Hessianmatrix and evaluating corresponding loss values. Our results suggest thatphysics residuals in PIML do not inherently hinder optimization, contrary to acommonly assumed failure cause. Instead, successful parameter updates requireboth ML and physics gradients to form acute angles with the quasi-true gradientand lie within a conical region. Given inaccuracies in both the physics modelsand the training data, satisfying this condition is often difficult.Experiments reveal that physical residuals can degrade the performance of LWR-and ARZ-based PIML models, especially under highly physics-driven settings.Moreover, sparse sampling and the use of temporally averaged traffic data canproduce misleadingly small physics residuals that fail to capture actualphysical dynamics, contributing to model failure. We also identify theCourant-Friedrichs-Lewy (CFL) condition as a key indicator of datasetsuitability for PIML, where successful applications consistently adhere to thiscriterion. Lastly, we observe that higher-order models like ARZ tend to havelarger error lower bounds than lower-order models like LWR, which is consistentwith the experimental findings of existing studies.