Interpretable subsurface damage prediction for optical components via physics-guided neural network adaptive symbolic layer

Subsurface damage (SSD) depth prediction is a critical challenge in the ultraprecision grinding of optical components because it directly affects subsequent material removal, surface integrity, and the laser-induced damage threshold. Accurate prediction remains difficult due to the nonlinear coupling between machining parameters and crack evolution, as well as the limited interpretability of purely data-driven models. To address this problem, we propose a physics-guided neural network with an adaptive symbolic layer (PGNN-ASL) for SSD depth prediction in brittle optical materials. By integrating an attention-based symbolic engine, the framework identifies physically meaningful features governing crack propagation and reconstructs them into an interpretable symbolic surrogate relation, which is incorporated into the network as a fracture-mechanics-based constraint. Comparative results show that PGNN-ASL achieves higher predictive accuracy than representative baseline methods, reducing the root-mean-square error by 7.5%. Experimental validation on previously unseen glass-ceramic specimens using the magnetorheological finishing spot method further demonstrates good agreement between predicted and measured SSD depths, together with reliable uncertainty estimates. The proposed approach provides an interpretable and practically useful tool for predictive quality assessment, process optimization, and low-damage manufacturing of precision optical components.