Data-driven convolutional neural network model with multi-channel input framework for predicting structural displacement response

Abstract

Predicting the displacement response of structures accurately is essential for assessing their performance. This research introduces a data-driven model utilizing an encoder-decoder Convolutional Neural Network (CNN) to efficiently and precisely analyze the complete displacement response of structures. A multi-channel input framework is devised to convert detailed physical attributes, such as geometry, boundary conditions, and loads, into data that can be processed by the CNN. Within this framework, boundary conditions and loads are treated as variables and mapped using Approximation Distance Functions (ADFs), enabling the pre-trained model to accurately predict responses under diverse conditions without repetitive retraining. The effectiveness of the proposed model is tested on various typical structures, including simply supported plates, beams, and walls with variable openings. Remarkably, results for walls with variable openings show a Root Mean Square Error (RMSE) of less than 0.072 and a coefficient of determination greater than 0.9992 for full-field displacement predictions. Comparisons of Floating-Point Operations (FLOPs) and execution time indicate that the proposed model can be significantly faster—up to hundreds of times—than traditional Finite Element Methods (FEM), especially when dealing with models discretized using a large number of grid points (over 40,000).