Optimizing material and process parameters in laser engineered net shaping using liquid neural networks

This study introduces a physics-based Liquid Neural Network (LNN) framework for optimization of material and process parameters in Laser Engineered Net Shaping (LENS). The model combines neural differential equations with dynamic liquid weight mechanism, and enables adaptive learning under thermal stress. Physical constraints derived from Fourier law and thermo-mechanical relationships are introduced during training to improve physical interpretability and generalization. We train the model on 90 samples (30 per material) from three representative alloys: Ni60A, Stellite6, and In625. Experimental results show high prediction accuracy within the explored parameter space (R²> 0.9 for key quality measures on the test set) and effective multi-objective optimization. The optimized parameters yield improved microhardness (+ 12.5%), reduced dilution (18%), and improved cladding consistency (CUI = 0.97). Results demonstrate that this approach can achieve defect-minimized, energy-efficient additive manufacturing. We lay a foundation for closed loop control and digital twin integration for advanced laser cladding applications.