Prediction of the compressive strength and carpet plot for cross-material CFRP laminate based on deep transfer learning

Abstract

The correlation mechanism between different Carbon fiber-reinforced polymer (CFRP) materials is unclear, and mechanical modeling cannot be rapidly promoted through knowledge sharing, which increases the time and cost of new material development and reduces the efficiency of accumulated data. In this paper, a Bi-Stage Optimize Deep Neural Networks (BSO-DNN) with Transfer Learning(TL) machine is proposed as a mechanics modeling method, which is ‘tailor-made’ for different materials, improving the accuracy of modeling and using efficiency of data. A compressive strength prediction model for FRP laminates was constructed by combining the components and process. TL-BSO-DNN significantly improves the robustness of the model, the predicted values are closer to the real sample distributions, and the accurate distributions provide reliable design and allowable values for the further use of the materials, which reduces the MRE of the model by 6.9 % and 8.3 %, and the RMSE by 58 % and 64 % in test set 1 and test set 2, respectively. Based on the predicted value and the prediction model, the relationship between the ply ratio and the compressive strength is reasonably extrapolated by data-driven, and the carpet plots are designed. The combination of data-driven, deep neural networks and transfer learning has brought direct benefits to the rapid construction of mechanical models, the effective improvement of modeling accuracy, the reasonable extrapolation of performance plots, and the rapid exploration of new materials.