Surface roughness prediction in milling process using an interpretable physics-informed Swin Transformer framework

Surface roughness is a core indicator for evaluating machining quality, thereby necessitating accurate prediction and control within milling operations. Nevertheless, conventional physical models often rely on idealized assumptions to simplify the modeling process, resulting in insufficient predictive accuracy. Conversely, data-driven methods, while offering high precision, lack interpretability due to their black-box nature, thereby hindering widespread industrial adoption. To bridge this gap, this study presents a Physics-Informed Swin Transformer (PIST) framework for milling surface roughness prediction. The core innovation of the framework lies in the deep coupling of machining mechanism prior knowledge with the Swin Transformer architecture. First, an analytical physical model for surface roughness is derived from milling dynamics to extract prior physical knowledge. Subsequently, this knowledge is integrated with multi-source sensor signals through Cross Physics-Data Fusion (CPDF) and fed into the Swin Transformer model. Furthermore, a physics-informed loss function incorporating physical consistency constraints is constructed. This allows the prior physical knowledge to effectively correct data-driven prediction biases, thereby enhancing both predictive performance and interpretability. The effectiveness and performance advantage of the proposed method are confirmed through comprehensive comparative tests. Experimental results demonstrate that the proposed method achieves a MAPE of only 0.70 %, an RMSE of 0.0877 μ m , an MAE of 0.0143 μ m , and an R 2 of 0.9822, outperforming both purely physical and data-driven approaches. This study provides a mechanism-guided monitoring framework for the prediction of surface roughness in digital manufacturing environments.