Digital tvilling-modellering av värmebehandling av gjutna metaller med OpenFOAM och fysikinformerad AI

Sammanfattning

Heat treatment is a critical processing step in the manufacturing of cast metal components. During this process, workpieces are held at elevated temperatures (typically 900–1100 °C) for a prescribed duration to drive metallurgical transformations that deliver the desired hardness, tensile strength, and fatigue resistance. Most foundries still rely on empirical trial-and-error to optimize temperature and time, whereas the complementary route of digital-twin modeling remains underdeveloped. This thesis establishes an open-source digital twin of an electrically heated laboratory furnace for industrial applications by coupling an OpenFOAM ground-truth solver with three physics-informed neural-network surrogate architectures and benchmarks their accuracy and computational efficiency. A conjugate heat-transfer model coupling conduction, convection, and surface-to-surface radiation across 12 material regions of a reference furnace was constructed and validated against thermocouple measurements provided by the Research Institutes of Sweden (RISE), achieving agreement within 2 °C at the end of the heating phase. Latin-Hypercube sampling was used to generate 89 candidate cases by varying the furnace setpoint and cylinder position, of which 78 (87.6%) passed the convergence filter and were retained as the training dataset. Three surrogate architectures (a graph neural network based on MeshGraphNet, a Fourier neural operator, and a deep operator network) were trained in the NVIDIA PhysicsNeMo framework and evaluated through a 326-step autoregressive rollout on seven held-out test cases. The graph-based surrogate achieved a steel-cylinder mean absolute error of 2.06 K with a coefficient of determination of 0.9996 in the in-distribution regime, outperforming the Fourier and deep-operator baselines by one and two orders of magnitude, respectively, despite having the smallest parameter count of the three architectures. Inference was completed within seconds rather than the approximately ten hours of CPU time required per case by the ground-truth solver, yielding a deployment speed-up of three to four orders of magnitude. These findings indicate that a graph-based surrogate, aligned with the finite-volume discretization of the underlying solver, can replace high-fidelity simulations within the ±10 K industrial control resolution. This provides a credible path toward surrogate-accelerated digital twins for industrial heat treatment. The dataset-generation pipeline and the training and evaluation code for the three surrogate architectures are publicly available at https://github.com/Dhruhil/RISE_furnace.