Energy-derivative attention enhanced deep learning for multi-phase segmentation of mesoscale heterogeneous material using X-ray computed tomography images

The precise and autonomous segmentation of mesoscale phases in heterogeneous materials remains challenging due to the similarity in characterization between holes and fractures. To address this issue, a neural network for computed tomography of multi-phase composites (MCCTNet) is established with an innovative architecture of adaptable configurations to recognize pixel-level mesoscale phases using X-ray Computed Tomography (X-CT) images of composites. Based on the original U-Net-like encoder-decoder structure, M to N layers are integrated with a novel attention module, termed the energy-derivative attention module (EDAM), which is designed to learn explicit feature representations for regional energy and boundary geometry. A pixel-level labeled dataset with 600 X-CT images covering diverse phases was established. The effectiveness of the proposed method and its superiority over existing methods were validated through comparative studies and ablation tests. EDAM significantly improves the recognition of small region-of-interests (RoIs), achieving in an improvement of 2.29 %, 1.16 %, 0.92 %, and 0.66 % for fracture, hole, cement paste, and aggregate, respectively. In addition, the proposed MCCTNet-1-4 embedded with EDAM demonstrated robust and consistent segmentation accuracy under Gaussian, salt-and-pepper, and speckle noise. Finally, practical applications including two-dimensional analysis, uniaxial compression simulations, and three-dimensional reconstruction based on the multi-phase segmentation results were conducted to verify the proposed method.