High-Velocity Air-Fuel (HVAF) spraying of Fe-based amorphous coatings involves strong nonlinear coupling among multiple process parameters, while practical optimization is severely constrained by limited experimental data and poor model interpretability. To address these challenges, a systematic data-driven optimization framework integrating the Denoising Diffusion Probabilistic Model (DDPM)-based data augmentation with explainable machine learning is proposed. Coating microhardness and hardness uniformity were jointly selected as target properties to capture both performance level and spatial reliability. Three generative models—Generative Adversarial Network (GAN), Variational Autoencoder (VAE), and DDPM—were comparatively evaluated using statistical matching and distribution-consistency metrics, revealing that DDPM most faithfully reproduces the intrinsic statistical characteristics of real HVAF process data. We benchmarked ten representative regression algorithms covering classical statistical learning, ensemble methods, and deep learning paradigms, with GBR demonstrating the highest predictive accuracy and stability. The inclusion of 10% DDPM-generated samples further improved the predictive precision of the GBR model. SHapley Additive exPlanations (SHAP) quantitatively identified spraying distance as the dominant parameter governing coating hardness, while elucidating the coupled effects of multiple parameters on hardness uniformity. By interpolatively expanding the process parameter space, a two-stage screening strategy identified 98 high-performance parameter combinations. Experimental validation confirmed that the optimal parameter set simultaneously achieved higher hardness and improved uniformity compared with the original best condition, resulting in a 13.6% reduction in wear rate.
Zhang et al. (Thu,) studied this question.