A Deep-Learning-Based Constitutive Method for Geomaterials Using a Neural Cutting Plane Algorithm

Constitutive modeling for geomaterials remains challenging because of limited data availability, strong nonlinearity, pressure sensitivity, and the non-smooth characteristics of commonly used yield surfaces. This study presents a deep-learning-based constitutive method for geomaterials that incorporates a neural stress-integration procedure based on the cutting plane algorithm (CPA). Two compact fully connected networks are trained to learn the yield function and its stress gradient from an augmented stress-state dataset. The trained networks are then incorporated into a cutting plane return-mapping procedure, in which only first-order information is required for the plastic stress return. This avoids explicit analytical yield expressions and second-derivative evaluations and is therefore more naturally compatible with non-smooth Mohr–Coulomb-type yield-surface representations in a first-order return-mapping sense. Numerical results show that the proposed method reproduces the reference Mohr–Coulomb response along the examined monotonic triaxial compression paths. Compared with the finite-difference closest-point projection method (CPPM) implementation considered in this study, the CPA-based neural stress-update procedure requires fewer network calls per update, indicating a more economical implementation for the present learned constitutive framework.