Project Description

(ML-assistance for faster computational fluid dynamics) The goal of this project is to develop a new machine learning based method to simulate fast-moving vehicles with complex geometries in compressible flows, called Domain-transferable Graph Operator (DGO). Our approach is able to query a numerical solver (also developed by us) when it is not confident about its prediction, and is able to continually learn out-of-distribution dynamics from CFD solver. DGO is also designed to be mesh-independent, that is able to work with different mesh structures, geometries, and resolution. This allows the model to always have reliable solutions and learn also while it is deployed, making it more reliable and partially avoiding the limitations of existing ML-based methods that are seen as black-boxes. We expect that our method will be more computationally efficient than numerical methods since DGO's metalearning generalizes numerical methods over small mesh-segments to other mesh segments and also to new flows.

Start Date

May, 2025

Postdoc Qualifications

Ph.D. in one of the following relevant fields:
* Computational Fluid Dynamics (CFD)
* Computer Science (Machine Learning/Artificial Intelligence)
* Applied Mathematics
* Mechanical/Aerospace Engineering

Machine learning expertise:
* Expertise in Python and deep learning frameworks (e.g., PyTorch)
* Experience with graph neural networks and metalearning
* Familiarity with domain adaptation and transfer learning

Computational fluid dynamics expertise:

* Extensive experience with numerical methods for CFD
* Proficiency in using and developing CFD solvers
* Knowledge of compressible flows and aerodynamics

Software development:
* Proficiency in version control systems (e.g., Git)

Publications:
* Strong track record of publications in relevant journals and conferences
* Experience presenting results at conferences and workshops

Problem-solving skills:
* Ability to develop innovative solutions for complex problems
* Experience in interdisciplinary research

Co-Advisors

Carlo Scarlo, Associate Professor of Mechanical Engineering (scalo@purdue.edu)
Bruno Ribeiro, Associate Professor of Computer Science (ribeirob@purdue.edu)

Bibliography

Takahiko Toki, Victor C. B. Sousa, Yongkai Chen and, Carlo Scalo, Sub-filter-scale shear stress analysis in hypersonic turbulent Couette flow, Journal of Fluid Mechanics, 2024, Vol 984, A53

SC Mouli, M Alam, B Ribeiro, MetaPhysiCa: Improving OOD Robustness in Physics-informed Machine Learning, The Twelfth International Conference on Learning Representations (spotlight paper), 2024