A physics-informed computer vision framework for graded erosion processes of hydraulic concrete under submerged sediment-laden jets

Erosion of hydraulic concrete induced by submerged sediment-laden jets constitutes a representative surface damage problem that is strongly governed by physical processes while exhibiting limited textural contrast, representing a multiphase sediment-structure interaction process relevant to sediment management and operation-maintenance of hydropower infrastructure. Its spatial heterogeneity and graded erosion patterns arise from the coupled effects of sediment momentum transfer and erosion evolution. Conventional erosion assessments predominantly rely on integral metrics such as mass or volume loss, which are insufficient to describe the two-dimensional spatial structure and graded characteristics of erosion damage. These erosion patterns represent a localized surface-morphological response of hydraulic concrete surfaces in sediment-laden jet environments. Although computer vision techniques have recently been applied to erosion detection, existing approaches remain largely texture-driven and data-centric, typically focusing on binary segmentation between damaged and undamaged regions. Such models are therefore inadequate for resolving multiple erosion grades and lack explicit incorporation of erosion mechanisms, leading to limited robustness and interpretability across varying hydraulic and sediment conditions. In this work, a physics-informed computer vision (PICV) framework is developed for intelligent segmentation of hydraulic concrete erosion, bridging mechanism-based sediment action modelling with data-driven image segmentation. The framework is built upon a structured physical-visual representation that explicitly links erosion morphology with sediment-induced physical actions. Controlled submerged sediment-laden jet experiments are conducted under systematically varied jet velocities, impingement angles, sediment concentrations, particle sizes, and exposure durations to acquire high-resolution erosion surface images. Based on particle impact and cutting mechanisms, spatially distributed sediment-phase momentum fields, including normal and tangential components, are derived to characterize the intensity of particle-wall interactions, serving as modelling-informed multiphase sediment action descriptors. These momentum fields are spatially registered to the corresponding erosion images, forming a coupled two-dimensional representation in which erosion surface images serve as the visual carrier and are associated with aligned physical descriptors. This representation provides a physics-vision integrated dataset suitable for mechanism-aware visual learning. On the basis of this coupled representation, a PICV-oriented multi-modal segmentation framework is established, in which erosion images and sediment momentum fields are jointly exploited to enable concurrent learning of textural features and physically meaningful action intensity. Furthermore, a dimensionless erosion intensity indicator derived from experimentally measured mass loss rates is incorporated into the loss function as a soft-consistency regularization term, providing sample-wise adaptive guidance during model optimization. Rather than imposing strict physical constraints on the solution space, physical information is used to guide the learning process toward physically plausible spatial patterns. Compared with image-only baselines (U-Net and DeepLab), the proposed PICV model improves multi-class graded-segmentation performance (Pixel-wise precision: +10-15 percentage points) and notably reduces grade confusion in transition regions. Under cross-condition evaluation, PICV demonstrates enhanced stability and interpretability, linking predicted grade distributions to aligned momentum patterns. This framework provides a transferable pathway for robust, mechanism-aware erosion assessment under complex submerged sediment-laden jet environments, supporting erosion-risk evaluation and sediment-management decision-making for hydraulic infrastructure.