<p>This paper adopts a data-driven approach, using the key process parameters such as laser power, scanning speed, and hatch spacing for the SLM forming of 316L stainless steel as the input and the surface roughness as the output. An exponential model and an improved particle swarm optimization algorithm are employed to optimize the neural network, thereby constructing an improved BP neural network model with stronger generalization ability. The research results show that the improved BP neural network exhibits superior predictive performance: The coefficient of determination (R²) of the improved BP neural network is significantly increased by 59.13%, 18.52%, and 8.22% compared to the BP, PSO-BP, and GA-BP models, respectively; the root mean square error (RMSE) is reduced by 57.53%, 39.88%, and 24.56%, respectively; and the mean absolute error (MAE) is decreased by 55.96%, 45.73%, and 25.36%, respectively. This indicates that the BP neural network optimized through the collaborative prediction of the exponential model and the improved particle swarm optimization algorithm can not only achieve high-precision prediction of the surface roughness of the sample, but also provide theoretical guidance for optimizing processing quality by quantifying the influence weights of process parameters, and has significant engineering application value.</p>