Semantic segmentation-based quantification of hydrogen embrittlement fracture features and estimation of H diffusivity in GS80A martensitic steel

High-strength martensitic steels are susceptible to hydrogen embrittlement (HE) when exposed to H-containing environments, but quantitative characterization of HE-related fracture evolution and H transport remains challenging. In this study, unsaturated H pre-charging tests are conducted on GS80A steel, and the evolution of fracture surface morphology is analyzed using scanning electron microscopy (SEM). A CBAM-DeepLabv3+-based semantic segmentation model is developed to identify dimples and cleavage facets. Grad-CAM visualization is used to examine whether the model focused on physically meaningful fracture features during recognition. Based on the segmented brittle area, an area-based method is proposed to calculate the equivalent H penetration depth under unsaturated H pre-charging conditions. The obtained time-dependent equivalent H penetration depths are then combined with FEM H diffusion simulation to estimate the H diffusivity of GS80A steel. The H diffusivity obtained by the proposed deep learning-based area method–FEM approach is 3.30×10-5mm²s⁻¹, which is reasonably consistent with the Image-Pro Plus measurement–FEM method and electrochemical H permeation test. The proposed workflow enables automated quantification of HE-related fracture morphology and provides a basis for linking fracture features with H transport behavior.