Physics-informed Machine Learning (Φ-ML) Meets Engineering Webinar

Thursday 25 May 2023, 13:00 → 14:00 UTC

Title: Promoting data-efficiency of deep learning for dynamical systems

Abstract: Deep learning has emerged as a promising alternative approach to the first principle-based methods in accelerated simulation and inverse modeling. However, deep learning normally requires a large amount of data and cannot generalize with limited data, which hinders its effective implementation in the context of scientific research where data is often expensive to acquire. In this talk, we will show that by incorporating widely applicable Euclidean symmetry, a neural network-based reduced-order model approach, named Euclidean symmetric neural network (ESNN), is able to learn from a single demonstration and generalize to a variety of unseen boundary conditions, initial conditions, and different material properties, on a demonstration dynamical system: 3D Slinky. The ESNN is trained under the neural ordinary differential equation (Neural ODE) framework to learn the 2D latent dynamics from the motion trajectory of a reduced-order representation of the 3D Slinky system. We show that by construction the ESNN preserves energy invariance and force equivariance on Euclidean transformations of the input, including translation, rotation, and reflection. The model order reduction leads to simulation acceleration by one to two orders of magnitude compared to traditional numerical methods. We also demonstrate through ablation study the necessity and influence of each component in the learning pipeline.

Speaker: Dr. Qiaofeng Li

ZOOM link: tbc

More Information / Subscribe / Past Seminars
https://andreapizzoferrato.notion.site/ML-meets-Engineering-fa48aefc1a7f40e4b98cb6f861f766cd