Virtual vibro-acoustic prototyping using physics-informed neural networks
Researchers
Research Areas
Mechanical engineering, Physics, Vibrations, Acoustics
Project Brief
In many industrial applications, noise and vibration performances of manufactured products need to be controlled to meet comfort and regulatory requirements. Virtual prototypes or digital twins consist of numerical models for predicting the vibro-acoustic behaviour of mechanical structures, in order to assess their radiated noise at the earliest design stages of a project. Conventional numerical methods based on discretization approaches, such as the finite element method or the boundary element method, present some limitations, in particular to address the mid and high frequency ranges and to consider uncertain parameters.
Although there have been advances in vibro-acoustic simulations over the last few decades (notably through the growing evolution of computing capabilities), there are no well-established and effective methods for predicting the behaviour of complex structures at mid to high frequencies.
Quite recently, supervised neural networks have shown promising results with numerical simulation data in various scientific fields. Until now, the data acquired was not always sufficient to accurately model physical phenomena. Moreover, in some cases, this technique does not respect the underlying physical laws. A few years ago, the Physics-Informed Neural Network (PINN) paradigm was proposed. Partial differential equations representing physical laws are directly included in the neural network using automatic differentiation. Numerous studies have focused on the modelling of steady or unsteady phenomena in fluid mechanics. However, by now, little attention was paid to represent the vibration and the noise radiated from elastic structures. This thesis aims to fill this gap.
The PhD works will consist in investigating the ability of PINNs to model the vibroacoustic behaviour of complex structures. Models of increasing complexity will be investigated. The modelling of bending motions of thin beams and plates will be considered at first to learn how to implement PINNs, to study the influence of the “loss” definition on the prediction of frequency response as well as how to take the boundary conditions.
We will also seek to introduce uncertain mechanical or geometrical parameters into the modelling in order to predict the dispersion of the vibratory response of the considered systems. Then, investigations will carry out on assembling of different thin structures on one hand, and on the coupling with the acoustic domain in another hand. Uncertainties at the coupling junctions could be introduced to describe imperfectly characterized assemblies.
We will investigate different neural network architectures as well as different “losses” to obtain reliable results within reasonable computation times. We could envisage to develop data-driven PINNs, including experimental data in the training process, as well as physics-based constraints.
At the end of the study, a practical application will be studied to highlight the interests of the proposed modelling and to demonstrate the ability of PINNs to deal with acoustics and vibrations engineering problems.