Automated Construction of Thermodynamically Consistent Digital Twins: A Neuro-Symbolic Framework using Bond Graph Constraints

The development of physiological digital twins is currently constrained by the manual effort required to derive thermodynamically consistent systems of differential-algebraic equations (DAEs). While large language models (LLMs) offer a potential solution for code generation, they suffer from structure hallucination—the generation of models that violate fundamental conservation laws. To address this, we introduce a neuro-symbolic framework that enforces thermodynamic consistency via semantic-constraint projection (SCP). Unlike end-to-end generation, our approach utilises a multi-module architecture where the LLM is constrained to output structured bond graphs. This strictly separates biological topology from physical constitutive laws, ensuring that all generated models satisfy the first law of thermodynamics by construction via the skew-symmetry of the junction structure matrix ($J = -J^T$). We validate this framework using a canonical model of epithelial glucose transport, demonstrating that the system can autonomously synthesise a spatially resolved topology from a library of 32 protein models, derive missing constitutive equations, and utilise an adversarial verification module to detect and correct parameter inconsistencies against literature data.