Learning Uncertainty-Aware Composite Constitutive Laws via Bayesian Recurrent Neural Network
Date:
Learning-based constitutive laws are increasingly used to accelerate the design and analysis of new composite materials. While they have shown merits in the cases of data-rich scenarios, they can only converge to a point estimation even with a time-consuming fine-tuning process. Moreover, the risk of over-fitting is unavoidable because of inherent noise in the data collection process. In this talk, we propose to use the Bayesian Recurrent Neural Network (BRNN) to characterize composite history-dependent plasticity constitutive law leading to an uncertainty-aware paradigm. Meanwhile, we compare the performance of state-of-the-art Bayesian inference approaches for composite material law modeling, and identify the most suitable approach considering the predicting accuracy, implementing complexity, and executing time. Results show that BRNN has better prediction than typical RNN and additional uncertainty prediction that could be used to design and optimize composite material. In addition,