Determining the burial depth for offshore pipelines to resist impact load is challenging owing to the spatial variability of soil strengths, which proves to significantly affect failure behaviours of soils and pipelines. To facilitate the design, accurate and fast evaluation on pipeline damage is required. Here, an integrated surrogate model was developed to forecast impact damage of pipelines buried in spatially varied soils. Through coupling the random field and numerical simulation, a stochastic finite element analysis framework was derived and verified to yield the datasets; Based on the scheme of feature extraction - integration from convolution neural network, the surrogate model was established, which mapped the three-dimensional soil spatial field to the structural response. Prediction mechanism of the developed model was explored, where correlations among soil spatial distribution patterns, failure mechanisms and feature recognitions were discussed. The models enabled to capture the key features representing the failure mechanisms under random soil conditions, including the local failure mode of soil and pipe-soil interactions, which theoretically explained its feasibility in damage estimation. Further, model performance was comprehensively evaluated with regard to prediction accuracy, uncertainty quantification, and transfer learning, and the corresponding causes were investigated. Satisfactory performance and high computation efficiency were demonstrated.

The abandoned carbonaceous mudstone has caused severe environmental problems such as land occupation and landslides. For the consideration of economic and ecological factors, carbonaceous mudstone soil-rock mixture (CMSRM) is used as an embankment material assessed by California bearing ratio (CBR) and unconfined compression strength (UCS). A series of experiments were conducted to measure the CBR and UCS of the CMSRM with different wet-dry cycles (0, 2, 4, 6 and 8) and different rock contents (0, 20, 40, 60 and 80%). The experimental results were predicted and analysed by a convolutional neural network (CNN). The experiment results show that the CBR and UCS of CMSRM increased at first and then decreased with the increase of rock content and were negatively correlated with wet-dry cycles. The CNN predicted values were highly correlated with the measured values. The CNN model enables variable parameter analysis of the experiment results via deep learning, which provides a new method to the CMSRM embankment road performance prediction.