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

TL;DR

This work tackles the challenge of accurate, efficient predictions for path-dependent elasto-plastic behavior in short-fiber reinforced composites using recurrent neural networks. It introduces test-time data augmentation (TTA), which rotates inputs and back-rotates outputs to generate multiple predictions per test example, whose mean yields a more accurate and smoother time signal while providing an uncertainty estimate. Across mean-field and full-field data, TTA reduces prediction error by roughly 19% on average and demonstrates improved shape consistency; the associated prediction uncertainty correlates with actual errors, supporting practical confidence estimates. The approach offers a robust, training-free uncertainty proxy and could be extended to physics-informed architectures and other domains relying on data-driven surrogate models.

Abstract

Recurrent Neural Networks (RNNs) have emerged as an interesting alternative to conventional material modeling approaches, particularly for nonlinear path dependent materials. Remarkable computational enhancements are obtained using RNNs compared to classical approaches such as the computational homogenization method. However, RNN predictive errors accumulate, leading to issues when predicting temporal dependencies in time series data. This study aims to address and mitigate inaccuracies induced by neural networks in predicting path dependent plastic deformations of short fiber reinforced composite materials. We propose using an approach of Test Time data Augmentation (TTA), which, to the best of the authors knowledge, is 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. Our analysis also demonstrates improved shape consistency between the prediction and the target pseudo time signal. Additionally, this method provides an uncertainty estimation which correlates with the absolute prediction error. The TTA approach is reproducible with different randomly generated data augmentations, establishing a promising framework for optimizing predictions of deep learning models. We believe there are broader implications of the proposed method for various fields reliant on accurate predictive data driven modeling.

Figures (13)

• Figure 1: Micro-mechanical simulation (target) and RNN results (prediction) of von Mises stress ($\sigma_{\text{v}}$) and a near-zero shear stress component ($\sigma_{12}$).
• Figure 2: The RNN architecture, consisting of 3 hidden GRU layers each containing 500 hidden states, where ${h_{t}}$ is the hidden layer vector.
• Figure 3: A representation of the proposed TTA method, in which an input is rotated a certain number of times, and the corresponding output is rotated back to the original configuration.
• Figure 4: The corresponding probability distribution of $\text{MeRE}_i$ values for 100,000 back-rotated predictions.
• Figure 5: $\text{MeRE}_i$ of each corresponding back-rotated prediction on a unit sphere, projected on a 2D plane using Mollweide projection Feeman2000.