Model fusion with physics-guided machine learning: Projection-based reduced-order modeling

The unprecedented amount of data generated from experiments, field observations, and large-scale numerical simulations at a wide range spatiotemporal scales has enabled the rapid advancement data-driven especially deep learning models in fluid mechanics. Although these methods are proven successful for many applications, there is grand challenge improving their generalizability. This particularly essential when employed within outer-loop applications like optimization. In this work, we put forth physics-guided machine (PGML) framework that leverages interpretable physics-based model with model. Leveraging concatenated neural network design multi-modal sources, PGML capable enhancing generalizability effectively protects against or inform about inaccurate predictions resulting extrapolation. We apply as novel fusion approach combining Galerkin projection long- to short-term memory (LSTM) parametric order reduction flows. demonstrate improved purely through injection physics features into intermediate LSTM layers. Our quantitative analysis shows overall uncertainty can be reduced approach, test coming distribution different than training data. Moreover, our used an inverse diagnostic tool providing confidence score associated observations. proposed also allows multi-fidelity computing by making use low-fidelity online deployment quantified models.