A Machine Learning Approach to Cold Spray Nozzle Geometry Optimization

Optimizing cold spray nozzle geometry is challenging due to the lack of direct solution functions, requiring iterative, computationally expensive simulations. This study introduces a novel approach integrating one-dimensional simulations with machine learning to optimize nozzle design efficiently. A Schlieren imaging setup under various spray conditions is employed to validate 3D Computational Fluid Dynamics (CFD) simulations. This validation is then extended to a quasi-one-dimensional compressible flow model in MATLAB, preserving empirical data fidelity. Following model validation, a Monte Carlo approach focuses on the nozzle's diverging section, with a fixed inlet (10 mm), throat (2 mm), and length (150 mm). The diverging section, divided into 50 radial-variable nodes, is analyzed under varying particle sizes (10-50 µm, 0.8 sphericity), inlet temperatures (400-1000 K), and pressures (2-7 MPa). This extensive dataset then trains a neural network capable of forward and inverse problem-solving:  deposition efficiency prediction and optimized nozzle geometry generation for given conditions—a solution that bypasses the traditionally prolonged iterative design process. Under fixed spray conditions, the current neural network achieves 96% accuracy in nozzle optimization. Future work will incorporate physics-informed neural networks (PINNs) to optimize nozzle geometry and spray conditions simultaneously, providing a robust framework for cold spray nozzle design.