To address the challenge of accurately evaluating the relationship between roller burnishing parameters and surface deformation in ultra-high strength steel torsion bars

this paper develops a Bayesian-optimized neural network model incorporating physical constraints. The model features a dual-output channel architecture that simultaneously predicts the rolling diameter variation and outputs a variance reliability metric

enabling quantitative assessment of prediction uncertainty. By integrating Bayesian optimization for global hyperparameter tuning

along with the SiLU activation function and a composite physical constraint loss function containing second derivatives

the model significantly enhances generalization capability and prediction accuracy through a triple-control mechanism. Experimental results demonstrate that the model reduces the mean absolute error of rolling diameter variation prediction to 0.051 μm

while its variance warning mechanism effectively identifies anomalous inputs beyond processing parameter ranges (e.g.

9000 N rolling force)

enabling dynamic evaluation of prediction reliability. This approach provides valuable technical support for high-performance and high-precision intelligent manufacturing of spring torsion bars.