Laser-assisted robotic roller forming (LRRF) significantly improves the formability of ultra-high-strength steels (UHSS, ≥1.5 GPa) through localised thermal softening, yet predicting its transient thermo-mechanical response and springback remains fundamentally challenging. A critical domain gap exists between standardised isothermal calibrations and the highly dynamic, gradient-dominated in-process conditions, rendering conventional constitutive models and purely data-driven surrogates unreliable under distribution shift. This work presents a physics-informed hybrid machine learning framework for MS1700 martensitic steel (1700MPa grade) that explicitly bridges this gap. A two-level experimental dataset is constructed to explicitly link material-level thermo-mechanical behaviour with process-level state variables. Methodologically, a stacking ensemble leveraging Random Forest and Gradient Boosting is integrated with a physics-informed neural network (PINN). The PINN enforces experimentally validated gradient-sign constraints-thermal softening, early-strain hardening and energy-driven springback reduction-via an adaptive loss function that progressively balances data fidelity with physical consistency. The stacked model achieves R2 = 0.9996 (RMSE = 4.56 MPa) for stress prediction and R2 = 0.9539 (RMSE = 1.24°) for pass-wise springback prediction. Crucially, compared to purely data-driven baselines, the physics-regularised framework successfully suppresses unphysical extrapolations in out-of-distribution evaluations, providing a robust, physically consistent digital-twin pathway for process optimisation and real-time compensation.
Jimmy et al. (Thu,) studied this question.