Autonomous Digital Twins with multiPhysicsFoam: A Physics-aware, Non-intrusive Reduced-order Modeling Pipeline
Please login to view abstract download link
Two converging trends currently shape computational engineering: (i) open-source simulation software is maturing toward industrial-grade multiphysics capability, and (ii) digital twins are transitioning from conceptual prototypes to operational assets. On the solver side, OpenFOAM-derived multiphysics frameworks now provide robust multi-region coupling for conjugate heat transfer, radiation, and thermo-fluids, with multiPhysicsFoam representing a promising realization of this momentum. On the deployment side, digital twins are increasingly interpreted in a closed-loop sense: a high-fidelity virtual representation synchronized with the physical system via real-time, bi-directional data exchange to enable monitoring, prediction, and model-based decision making. A primary bottleneck for closed-loop twins is computational latency. In particular, model predictive control (MPC) and other optimization-based autonomy concepts require repeated rollouts over prediction horizons, which for realistic multiphysics models translates into required speedups far beyond real time, typically $\mathcal{O}(10^{4})$, while maintaining accuracy and physical consistency. Classical intrusive reduced-order modeling is, however, often impractical for complex industrial solvers and rapidly evolving open-source code bases. This contribution proposes a physics-aware, non-intrusive reduced-order modeling (ROM) pipeline tailored to multiPhysicsFoam-based digital twins. The workflow combines (i) simulation-driven, data-efficient training set selection with TwinLab \cite{TwinLab}, (ii) latent-state compression of coupled multiphysics fields, and (iii) stable temporal surrogates based on Stable Port-Hamiltonian Neural Networks (sPHNNs). The resulting surrogate is exported as a deployable component (e.g., FMI) to enable low-latency inference on edge hardware and integration into MPC loops. We demonstrate the end-to-end pipeline -- from multiphysics modeling and data generation to ROM training and closed-loop operation -- on a representative thermo-fluid digital-twin scenario, providing a practical blueprint for autonomous digital twins with open-source multiphysics CFD without modifying solver internals.
