Answers are grounded in this paper and cite their sources.
Calibration of advanced soil constitutive models remains challenging because multiple parameters jointly control stress–strain responses. In practice, these parameters are often determined by manual trial and error from a few laboratory tests, leaving two questions unanswered: whether the fitted parameters are identifiable and which additional test would best reduce the remaining ambiguity. A parameter set may therefore reproduce the available data yet remain non-unique, leading to divergent predictions under untested conditions. In this work, we develop a human-supervised agentic-AI calibrate–diagnose–suggest workflow for parameter identification in soil constitutive models. The workflow combines gradient-based parameter estimation, Fisher-information-based diagnosis of weak parameter directions, and a large language model (LLM) agent that translates the weakest Fisher direction into a targeted follow-up laboratory test. This translation is guided by an LLM-organized parameter–response map, which is checked against established expert guidance. We demonstrate the workflow using NorSand in eight simulator-generated triaxial scenarios spanning drainage conditions, density states, and confining pressures. A single agent-suggested follow-up test substantially reduces the broad multi-parameter ambiguity left by single-test calibration: excluding one weakly observable parameter, the median relative error is 0.17%, and Fisher information analysis indicates improved local identifiability in all eight scenarios. The workflow also exposes the parameter that fit-only calibration would leave weakly constrained. We further find that a complementary test selected retrospectively from published data reduces local ambiguity in the targeted parameter combination. These findings show that agentic AI can support auditable calibration by turning numerical identifiability diagnosis into actionable decisions.
No takes yet. Share an insight, caveat, or question.
Choi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: