A Structure-Informed DNN Model for Welding Flux Viscosity Prediction

Abstract

Viscosity plays a pivotal role in the performance of welding fluxes by directly affecting the weld pool and the resultant weld metal, both of which are largely determined by heat and mass transfer during the welding process. Accurate viscosity prediction is essential for developing high-quality welding fluxes. However, existing models struggle to capture the complex nonlinear relationships between viscosity, composition, and temperature, which may significantly compromise robustness and generalization. To address this challenge, we incorporate NBO/Si (the average number of non-bridging oxygens per network former Si4⁺) as an intrinsic structural descriptor and propose a structure-informed deep neural network (S-DNN) that hierarchically models the composition-structure-viscosity relationship, effectively capturing nonlinear viscosity features. The S-DNN demonstrates excellent predictive performance (coefficient of determination [R2] = 0.9053, mean absolute error [MAE] = 0.0633 Pa·s, and root mean square error [RMSE] = 0.1101 Pa·s). Cross-validation employing multiple experimental datasets further validates the generalization and engineering applicability. Incorporating structural descriptors improves model performance by providing physical constraints. Compared to a basic deep neural network that relies solely on composition and temperature, S-DNN offers significantly higher accuracy. This work establishes a reliable, data-driven framework for the intelligent design of welding fluxes.