Reflections on the Development Stages and Engineering Deployment of Physics-AI Integration: Challenges, Solutions, and Case Studies of AI Agents in System Modeling and Simulation

Web-based modeling platforms and automated modelchecking workflows in CI/CD pipelines require remotely callable Modelica frontend services. This paper presents SMC (Simtek Modelica Compiler), a service-oriented Modelica frontend implemented in Go. SMC exposes model loading, structured queries, diagnostics, and flattening through a service API backed by a pipeline for parsing, instantiation, semantic analysis, connection processing, and flat-model generation. Drawing on the SMC implementation, we also discuss the engineering implications of Go for single-binary deployment, concurrent processing, representation of Modelica language constructs, and memory allocation. We demonstrate the prototype through integration with a web modeling platform. A common set of 606 comparable classes from the Modelica Standard Library (MSL) 4.0.0 is used both to assess frontend-result consistency against the OpenModelica Compiler (OMC) and to measure flattening time, memory allocation, and garbage-collection behavior. The evaluation covers frontend artifacts and in-process processing costs; backend processing, code generation, simulation, and FMU export are outside its scope.

Deep peak-shaving in coal-fired boilers causes non-linear dynamics, rendering traditional PID control insufficient for NOx emission management. This paper proposes a Modelica-based digital twin framework for a 600 MW boiler. The CFD-PINN model is coupled with a physicsbased steam-side dynamic model and an adaptive correction layer driven by DCS (Distributed Control System) data to ensure high-precision synchronization under operating condition drift. An intelligent reinforcement learning agent using the Recurrent Proximal Policy Optimization (Recurrent PPO) algorithm is developed to inject residual bias signals into the main loop, optimizing NOx emissions without altering the original architecture. Validated in seven scenarios spanning 300–600 MW, this strategy reduces the 95thpercentile NOx by 19%–23% and emission fluctuations by 32.7%–43.2% across four dynamic scenarios relative to both manually tuned and gain-scheduled PID baselines, while maintaining combustion efficiency. This approach provides a deployable, low-risk pathway for achieving stricter NOx compliance under deep peak shaving operation.