Article: Predicting settlement in reinforced subgrades on highways using physics-embedded sequential large models Journal: International Journal of Reasoning-based Intelligent Systems (IJRIS) 2026 Vol.18 No.21 pp.14 - 31 Abstract: The physics-embedded sequential large model integrates Boussinesq stress distribution, Mohr-Coulomb interface mechanics, and permanent deformation accumulation theory as differentiable loss constraints within a temporal-spatial attention architecture for predicting settlement in highway reinforced subgrades. Existing analytical approaches cannot capture nonlinear deformation accumulation under cyclic loading, data-driven models lack physical consistency for extrapolation, and current physics-informed methods use simplified constraints without interface mechanics or temporal attention. This model fills these gaps through temporal-spatial attention modules capturing progressive settlement dependencies and adaptive weighted loss training with heteroscedastic uncertainty modelling providing calibrated prediction intervals. Evaluation on 1250 cyclic triaxial tests and 45 field sections yields a root mean square error of 2.34 millimetres and a coefficient of determination of 0.947, with deformation and resilient modulus errors reduced by 27.1% and 27.4% over physics-informed baselines. Inderscience Publishers - linking academia, business and industry through research

Title: Predicting settlement in reinforced subgrades on highways using physics-embedded sequential large models

Authors: Ang Gao

Addresses: School of Architecture and Engineering, Zhejiang Industry Polytechnic College, Shaoxing, 312000, China

Abstract: The physics-embedded sequential large model integrates Boussinesq stress distribution, Mohr-Coulomb interface mechanics, and permanent deformation accumulation theory as differentiable loss constraints within a temporal-spatial attention architecture for predicting settlement in highway reinforced subgrades. Existing analytical approaches cannot capture nonlinear deformation accumulation under cyclic loading, data-driven models lack physical consistency for extrapolation, and current physics-informed methods use simplified constraints without interface mechanics or temporal attention. This model fills these gaps through temporal-spatial attention modules capturing progressive settlement dependencies and adaptive weighted loss training with heteroscedastic uncertainty modelling providing calibrated prediction intervals. Evaluation on 1250 cyclic triaxial tests and 45 field sections yields a root mean square error of 2.34 millimetres and a coefficient of determination of 0.947, with deformation and resilient modulus errors reduced by 27.1% and 27.4% over physics-informed baselines.

Keywords: reinforced subgrade settlement; sequential large model; physics-embedded learning; cyclic traffic loading; uncertainty quantification.

DOI: 10.1504/IJRIS.2026.155785

International Journal of Reasoning-based Intelligent Systems, 2026 Vol.18 No.21, pp.14 - 31

Received: 25 May 2026
Accepted: 29 Jun 2026
Published online: 13 Aug 2026 *

Title: Predicting settlement in reinforced subgrades on highways using physics-embedded sequential large models

Keep up-to-date