Dual Stream Regression Neural Network (DSR-Net) for shape-based parameter estimation
TL;DR
Evaluation studies on the DSR-Net confirmed the superiority of the DSR-Net over standard unimodal deep learning baseline models and the Cross Attention-based dual stream architectures as well.
Abstract
Standard deep learning models used for inverse problems in engineering design commonly experience difficulty in obtaining both temporal and spatial features simultaneously. In this study, a Dual-Stream Regression Network (DSR-Net) is proposed. The DSR-Net model includes a concatenation-based dual-stream architecture, including a Temporal Stream utilizing a dilated 1D CNN to extract sequential features, and a Spatial Stream employing a 2D CNN to process trajectory images, thereby extracting global spatial invariants. The DSR-Net model was trained and evaluated on a simulated dataset containing 59,966 coupler curves for the respective dimensionless link length ratios of four-bar mechanisms. The DSR-Net model achieves a coefficient of determination ( R 2 Score) of 0.9832 and a Mean Absolute Error (MAE) of 0.0479, which considerably outshines the standard models, including 1D CNN, Temporal CNN, and 2D CNN. Additional improvements in the DSR-Net were tested by replacing simple concatenation-based fusion with Cross Attention and using a physics- informed loss function. The model was evaluated with multiple shape comparison metrics, including Chamfer Distance, Hausdorff Distance, and Fréchet Distance. Evaluation studies on the model confirmed the superiority of the DSR-Net over standard unimodal deep learning baseline models and the Cross Attention-based dual stream architectures as well.