PHYSICS INFORMED NEURAL NETWORKS IN MANUFACTURING

This thesis investigates Physics-Informed Neural Networks (PINNs) for predicting flank wear (VB) in milling, targeting scenarios where data are scarce and purely data-driven models struggle to generalize. The proposed PINN embeds an extended Taylor tool-life relationship through a learnable time scaling T_f and re-parameterized exponents (n, m, p) with a positive scale K. The network ingests dimensionless time τ = t/T_f together with normalized process settings (feed, depth of cut, material) and is trained with a composite objective that combines data fitting on the available VB labels, an initial-condition penalty to enforce near-zero wear at the start, and a physics anchor that nudges the prediction at τ = 1 toward a robust reference wear level. Evaluation follows a leave-one-case-out protocol on the NASA milling dataset (16 operating conditions spanning feed, depth, and material), and results are contrasted against a matched deep neural network (DNN) baseline. The PINN achieves lower average MAE (0.1008 vs. 0.1074), markedly lower error variance (0.0146 vs. 0.0202), and reduced shares of samples exceeding 0.05 mm (61.9% vs. 69.7%) and 0.10 mm (34.3% vs. 38.3%), with degradations largely attributable to isolated first-sample spikes in two cases. Overall, the physics-guided time scaling improves trajectory tracking and stabilizes residuals relative to the purely data-driven baseline, demonstrating a practical route to robust tool-wear monitoring under limited labeled data.