Accelerating FEM with ML: an introduction to the Integrated Finite Element Neural Network
HTML-код
- Опубликовано: 21 авг 2024
- Speaker: Panos Pantidis (New York University Abu Dhabi, United Arab Emirates)
Title: Accelerating FEM with machine learning: an introduction to the Integrated Finite Element Neural Network (I-FENN).
Time: Wednesday, 2023.06.21, 10:00 a.m. (CET)
Place: fully virtual (contact Jakub Lengiewicz to register)
Abstract: Complementary to conventional numerical methods, physics-informed neural networks (PINNs) have recently emerged as alternative approximators for the solution of partial differential equations (PDEs). The main benefit of PINNs versus the conventional methods is their tremendous computational efficiency in terms of predictive speed, once the PINN model has been trained. Leveraging on their swift predictive capability, we integrate PINNs within the finite element solver and utilize them to approximate the solution of mechanics-governing PDEs. This step allows us to bypass the direct numerical discretization of the PDE, approximating the solution field at a fraction of time compared to traditional FEM. The developed framework is termed I-FENN (Integrated Finite Element Neural Network), and in this talk I will present its application in the case of nonlocal gradient continuum damage. A series of benchmark numerical examples will be presented, showcasing the computational efficiency and generalization capability of I-FENN, along with an extensive error convergence and hyperparameter analysis to solidify its background.
Additional material:
Pantidis and Mobasher (2023), Integrated Finite Element Neural Network (I-FENN) for non-local continuum damage mechanics, Computer Methods in Applied Mechanics and Engineering, 404: 115766 (link).
Pantidis and Mobasher (2023), Error convergence and engineering-guided hyperparameter search of PINNs: towards optimized I-FENN performance, Computer Methods in Applied Mechanics and Engineering, accepted (link for arXiv version).
Panos Pantidis obtained his BSc from the Aristotle University of Thessaloniki, Greece, in 2015. In 2019 he received his PhD from the Civil Engineering Department of the University of Massachusetts, Amherst. He then joined Thornton Tomasetti, working at their New York City Forensics practice. Dr. Pantidis is now a Postdoctoral Associate at New York University Abu Dhabi. His research focuses on computational mechanics and machine learning for multi-physics problems.
The aim of the Machine Learning Seminar series is to host presentations on fundamental and methodological advances in data science and machine learning, as well as to discuss application areas presented by domain specialists. The uniqueness of the seminar series lies in its attempt to extract common denominators between domain areas and to challenge existing methodologies. Therefore, the focus is on theory and applications to a wide range of domains, including Computational Physics and Engineering, Computational Biology and Life Sciences, and Computational Behavioural and Social Sciences. The seminar aims to bring together young and experienced researchers from various disciplines to exchange ideas on Machine Learning techniques. It is currently affiliated with the University of Luxembourg and is run under the auspices of the DTU DRIVEN PRIDE project, funded by the FNR, and the widening participation DRIVEN project, funded by H2020. The seminar also welcomes talks by researchers from a wider collaborative network, including but not limited to early-stage researchers in RAINBOW ITN, as well as current and incoming individual Marie Skłodowska-Curie fellows.
The usual format is as follows: a short presentation (20-30 minutes) followed by a longer discussion (30-40 minutes). The usual time is Wednesdays at 10:00 a.m. (CET). If you are interested in joining, please contact Jakub Lengiewicz. See www.jlengineer.....
nice
Very nice work... congratulations! The fact that you need the scale the strains for the PINN training is related to feature scaling. Feature scaling is very common on machine learning, especially when you have data with very different ranges. Quote "Feature standardization makes the values of each feature in the data have zero-mean (when subtracting the mean in the numerator) and unit-variance. This method is widely used for normalization in many machine learning algorithms (e.g., support vector machines, logistic regression, and artificial neural networks).[3][4]" (en.wikipedia.org/wiki/Feature_scaling).