Potential failures of physics-informed machine learning in traffic flow modeling: theoretical and experimental analysis
Authors:
Yuan-Zheng Lei,
Yaobang Gong,
Dianwei Chen,
Yao Cheng,
Xianfeng Terry Yang
Abstract:
This study critically examines the performance of physics-informed machine learning (PIML) approaches for traffic flow modeling, defining the failure of a PIML model as the scenario where it underperforms both its purely data-driven and purely physics-based counterparts. We analyze the loss landscape by perturbing trained models along the principal eigenvectors of the Hessian matrix and evaluating…
▽ More
This study critically examines the performance of physics-informed machine learning (PIML) approaches for traffic flow modeling, defining the failure of a PIML model as the scenario where it underperforms both its purely data-driven and purely physics-based counterparts. We analyze the loss landscape by perturbing trained models along the principal eigenvectors of the Hessian matrix and evaluating corresponding loss values. Our results suggest that physics residuals in PIML do not inherently hinder optimization, contrary to a commonly assumed failure cause. Instead, successful parameter updates require both ML and physics gradients to form acute angles with the quasi-true gradient and lie within a conical region. Given inaccuracies in both the physics models and 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 can produce misleadingly small physics residuals that fail to capture actual physical dynamics, contributing to model failure. We also identify the Courant-Friedrichs-Lewy (CFL) condition as a key indicator of dataset suitability for PIML, where successful applications consistently adhere to this criterion. Lastly, we observe that higher-order models like ARZ tend to have larger error lower bounds than lower-order models like LWR, which is consistent with the experimental findings of existing studies.
△ Less
Submitted 16 May, 2025;
originally announced May 2025.
Higher-order-modes enhanced phase-matched dispersive-wave generation in the deep-blue and UV spectral region
Authors:
X. T. Yang,
Z. Z. Luo,
J. P. Huang,
W. Y. Sun,
Y. Zheng,
R. C. Yin,
H. H. Yu,
M. Pang,
X. Jiang
Abstract:
During the last few decades, solid-core photonic crystal fibers (PCFs) have been extensively explored to generate broadband, high-coherence supercontinua (SC). Limited by the material absorption and relatively low nonlinearity of fused silica, spectral broadening in silica PCF-based SCs is usually restricted to the blue to near-infrared spectral regions, even in developed commercial sources. The o…
▽ More
During the last few decades, solid-core photonic crystal fibers (PCFs) have been extensively explored to generate broadband, high-coherence supercontinua (SC). Limited by the material absorption and relatively low nonlinearity of fused silica, spectral broadening in silica PCF-based SCs is usually restricted to the blue to near-infrared spectral regions, even in developed commercial sources. The output spectra of these sources are missing short wavelengths of the full range. Many efforts have been spent to break the limitation. Among them, dispersive-wave (DW) generation has been investigated for triggering new frequencies in short wavelengths. With satisfied phase-matching conditions, excessive energy can be directly transferred from solitons of the anomalous dispersion region to DWs of the short wavelengths. However, a systematical study of factors, including phase-matched DWs, strongly related to the dispersion tailoring of higher-order modes (HOMs), has rarely been shown. This study reports the experimental observations of HOM-enhanced phase-matchings for the DW generation in the deep-blue and ultraviolet regions. A solid-core PCF-based, UV-extended SC source spanning a 2.8-octave-wide (350 nm to 2500 nm) is demonstrated. Meanwhile, we carefully verify our findings via numerical calculations.
△ Less
Submitted 5 November, 2022;
originally announced November 2022.
