Data-physics driven reduced order homogenization

A hybrid data-physics driven reduced-order homogenization (dpROH) approach aimed at improving the accuracy of the physics-based reduced order homogenization (pROH), but retain its unique characteristics, such as interpretability and extrapolation, has been developed. The salient feature of the dpROH is that the data generated by a high-fidelity model based on the first order computational homogenization (i.e., without model reduction) can improve markedly the accuracy of the physic-based model reduction. The dpROH consist of the offline and online stages. In the offline stage, an enhanced model reduction strategy based on the Bayesian inference (BI) that employs the gated recurrent unit (GRU) neural network surrogate is conceived. In the online stage, the dpROH (rather than the GRU surrogate employed in the BI process) is utilized for the component level predictions. Numerical examples comparing various variants of the dpROH with the pROH, and the reference solution demonstrate its improved accuracy.

Automated summary

A hybrid data-physics driven reduced-order homogenization (dpROH) approach has been developed to improve the accuracy of physics-based reduced order homogenization (pROH) while preserving its interpretability and extrapolation. The dpROH utilizes data generated by a high-fidelity model to enhance the accuracy of the physics-based model reduction. The dpROH consists of an offline stage employing a Bayesian inference (BI) strategy with a gated recurrent unit (GRU) neural network surrogate, and an online stage utilizing dpROH for component level predictions. Numerical examples demonstrate improved accuracy compared to pROH and reference solution.