A Dual-Branch Neural Network for High-Load Strain Extrapolation Using Finite Element and Experimental Data

Abstract

Under high-load conditions, accurate strain prediction remains challenging due to nonlinear responses, noise, and incomplete measurements. This study proposes a pointwise dual-branch neural regression model using multisource data. A 1:8 scale-reinforced concrete frame was modeled in ABAQUS, and based on measurements up to 1.35 times the design load, three loading conditions were simulated to generate finite element (FE) strain data. The model combines an encoder-style nonlinear branch implemented using a Transformer encoder block with an MLP branch representing the dominant load–strain trend, while a dynamically weighted loss increases emphasis on large-strain samples. The best tested configurations achieve validation R² values of 0.9975 and 0.9997, with high-load MAPE values of 2.56% for experimental S7 extrapolation and 0.39% for held F5-at-1.0× cross-condition FE prediction. Under high-noise conditions, multisource data fusion increases validation R² from 0.9621 to 0.9965 while reducing high-load MAPE from 11.15% to 3.09%. These results demonstrate stable prediction performance within the tested benchmark.