Test-time data augmentation: Improving predictions of recurrent neural network models of composites

Short fiber reinforced composites (SFRCs) are widely used in variety of industries for their high stiffness- and strength-to-weight ratios, yet accurate elasto-plastic modeling via high-fidelity simulations remains computationally costly. Recurrent Neural Networks (RNNs) have emerged as an interesting alternative to conventional material modeling approaches, particularly for non-linear path-dependent materials. This study aims to address and mitigate inaccuracies induced by neural networks in predicting path-dependent plastic deformations of SFRCs. In this paper, we have developed a Test-Time data Augmentation (TTA), which, to the best of the authors’ knowledge, is a method previously untested in the context of RNNs. The method is based on augmenting the input test data using random rotations and subsequently rotating back the predicted output signal. By aggregating the back-rotated predictions, a more accurate prediction compared to individual predictions is obtained. Using the method on a micro-mechanical full-field dataset, reduces mean relative error by 19% and yields a standard-deviation-based uncertainty measure correlated with an absolute error of r=0.73. TTA incurs only 1.85 s per sample which is more than 1600 times faster than full-field simulations. Our analysis also demonstrates improved shape consistency between the predictions and the target pseudo time-signals. The TTA approach requires no retraining and is reproducible with different randomly generated data augmentations, establishing a promising and robust framework for optimizing predictions of deep learning models. We also believe there are broader implications of the proposed method for other artificial neural network architectures (such as transformers) and various data-driven predictive modeling applications.