Differentiable multiphysics for induction furnaces: Learning temperature-dependent material laws from sparse thermal videos
Ochiq kirish: yoʻq
Annotatsiya
We present a differentiable multiphysics framework for induction furnaces that jointly models eddy-current electromagnetics and transient heat transfer while learning unknown material and interface parameters from sparse thermographic videos. The forward model couples a quasi-static Maxwell formulation with anisotropic Fourier conduction and Joule heating, discretized by an unstructured finite-element scheme. End-to-end differentiability is achieved via adjoint operators and automatic differentiation, enabling gradient-based solution of the PDE-constrained inverse problem. A grey-box prior links parameters (electrical conductivity, magnetic permeability, thermal diffusivity, and coil-workpiece contact coefficients) to temperature trajectories through stability-regularized dynamics; identifiability is analyzed using sensitivity spectra and Fisher information. To exploit limited, noisy infrared measurements, we employ spatio-temporal masking with physics-consistent data assimilation and uncertainty calibration through Laplace approximations. On laboratory steel billets and production-line rings, the method recovers parameters within 5-10% of reference values and reduces the forecast error in surface temperature by 30-45% compared to PINN and black-box baselines, while preserving real-time feasibility for control. The approach enables self-calibrating digital twins of induction heating systems, improving energy efficiency, quality assurance, and safe operation under changing alloys and operating regimes.
Hali tarjima qilinmagan