Physics-Informed Cascaded Learning for Predicting and Optimizing Geometry and Quality in Laser Cladding Repair of Carburized Gear Steel

Abstract

Laser cladding is promising for use in repairing carburized gear steels, but parameter selection is challenging when clad geometry, substrate thermal disturbance, and morphological quality must be considered. Fifteen tracks of NHT.22.A01 iron-based powder were deposited on carburized-quenched 18CrNiMo7-6 steel with varying laser powers, scanning speeds, and powder feed rates. A physics-informed cascaded-learning framework predicted track width, height, Ac1-boundary depth, and quality. Leave-one-out out-of-fold width predictions were transferred to height and depth models to prevent target leakage. An auxiliary continuous quality index enabled the bi-objective optimization of quality and high-hardness layer depth, while multi-indicator process maps with local sensitivity analysis supported rapid parameter adjustment. Leave-one-out (R2) values for three geometric responses ranged from 0.934 to 0.968, and the maximum relative error at an unseen boundary condition was 4.3%. The fitted HAZ-depth/track width scaling coefficient (0.211) lay within the Rosenthal theoretical range (0.15–0.25), confirming the physical consistency of the cascade relationship. The optimization revealed a trade-off between quality and high-hardness layer depth; maximizing predicted high-hardness layer depth under (Qcont≥4.5) selected a condition already represented in the training set. Repeatability was supported by an independent batch replicate. By integrating leakage-controlled cascading, physically interpretable validation, constrained optimization, and sensitivity-resolved process mapping, the framework provides a transparent and practically applicable approach to small-sample laser cladding process design.