We are pleased to announce the seminar of

Antoine GOICHON, PhD student of team VAST-FM, on the "Hybridization of physical model and machine learning technique for the study of the structural condition of aeronautical structures". 

On 11/05/2023 at 13h15 via Teams.

Abstract :

With the increasing performance of deep learning models, their wide application domain, and their quality as universal approximators, fully data-driven approaches using these algorithms are gaining popularity, especially for solving physics-related problems. They generally require less expert knowledge and can learn very complex relationships that link data together. However, these models often have limited interpretability and explainability. Conversely, physical models based on a foundation of scientific and technical knowledge are inherently interpretable. The identification of their parameters generally requires less data. The hybridization of physical models and machine learning techniques attempts to bring together the advantages of both approaches while limiting their drawbacks. The goal is to combine the performance and flexibility of deep learning algorithms with the interpretability and lower data requirements of physical models.
This work aims at predicting the mechanical behavior of aeronautical structures using hybrid approaches mixing learning techniques and physical models. We will first discuss a hybrid architecture that has attracted a lot of interest in recent years : the Physic-Informed Neural Network (PINN). This architecture is based on the integration of physical equations within the objective function that trains the deep learning model. PINNs seek to learn from the experimental data, but also from the equations integrated in this way. We will then study another hybrid architecture aiming at integrating "classical" numerical solvers within a deep learning architecture : Neural Ordinary Differential Equation (Neural ODE). This approach allows to improve some aspects of the deep learning model while keeping some interesting properties of the solvers that have benefited from several decades of development, especially in terms of computational error estimation. These two hybrid approaches will be studied in the context of the development of a structural virtual sensor capable of reconstructing the in-flight behavior of aeronautical structures, solely from data obtained from aircraft instrumentation. This virtual sensor will be trained, tested and validated on two types of data : Synthetic data created thanks to mastered physical models and real data coming from flight tests of real or reduced size aircraft.