Parametric reduced order modelling and deep learning to accomplish pattern recognition and regression tasks

The use of high-fidelity physical/numerical models is still challenging when dealing with inverse problems aiming at para- meter estimation. To enhance the solution of these problems in computational mechanics, we propose to train statistics-based regressors via input-output data ad-hoc generated from the physics-based model. The goal is to obtain an easy to use data analysis tool capable of performing pattern recognition and regression tasks, in compliance with the governing equations of the model. In particular, we combine: (i) parametric model order reduction techniques, to reduce the computational burden connected to the input-output pairs generation; (ii) deep learning architectures, for the sake of pattern recognition and regression tasks.