A Unified Physics-Constrained Deep Reinforcement Learning Framework for Parameter Identification of Nonlinear Hysteretic Models

Abstract

Reliable nonlinear structural analysis requires hysteretic parameters that reproduce cyclic stiffness, strength, pinching, degradation, and energy dissipation. Conventional calibration is often tailored to one constitutive model and unit system, while repeated population searches become costly as dimensionality and parameter coupling increase. This study develops a unified physics-constrained deep reinforcement learning framework for OpenSees Steel02 and DowelType identification. Target responses and candidate parameters are expressed in dimensionless coordinates; bounded latent variables are decoded into admissible model parameters and mapped back to source units after calibration. A twin-delayed deep deterministic policy gradient (TD3) agent performs continuous search, with differential evolution providing local refinement when required. Validation used synthetic targets, random initial vectors, public steel records, and ten experimental hysteresis records from the authors’ research group; particle swarm optimization and a genetic algorithm served as benchmarks. The framework satisfied an NRMSE threshold of 0.02 in all 384 held-out Steel02 evaluations and achieved a mean NRMSE of 0.0166 on synthetic DowelType targets. On the ten experimental DowelType records, TD3+DE reached a mean NRMSE of 0.0654 and 30% success under the relaxed 0.05 threshold, giving accuracy comparable with tuned PSO at the same online OpenSees-call budget while retaining a reusable learned initialization step. One normalized workflow can rapidly obtain response-equivalent fits for distinct hysteretic laws and return solver-ready parameters in physical units.