Automated Modelica Model Repair Using Generative AI

Abstract

Cyber-physical systems (CPS) design and validation rely heavily on large-scale multidisciplinary models. Their growing size and complexity make them difficult to maintain: subtle faults such as singularities or unit inconsistencies cause compile- or simulation-time failures, or misleading results. Existing debugging practices rely on error diagnostics reported by compilers. Although compilers back-annotate these diagnostics onto the model source, the reports describe errors at the equation level rather than in the modeler’s conceptual view, leaving modelers without effective support for reasoning about root causes. Agentic large language models (LLMs) have shown strong repair capabilities on imperative code but remain largely unexplored for declarative, equation-based CPS models such as Modelica. To address this gap, this PhD contributes an agentic LLM-based framework for diagnosing and repairing Modelica models, together with the empirical foundations that make it feasible and reproducible. More specifically, its contributions are (1) an industry-needs study of generative AI in simulation-based test environments for CPS that scopes the rest; (2) an agentic technique that diagnoses faults by detecting, localizing, and explaining them in model-level terms; (3) an agentic technique that repairs faults by synthesizing patches under a compile/simulate feedback loop; (4) a benchmark dataset of mined Modelica model snapshots; (5) an empirical analysis of how model changes propagate; (6) a fault/fix taxonomy; and (7) a synthetic repair benchmark built by taxonomy-driven fault injection.