Towards efficient computational fluid dynamics simulations with physics-informed machine learning

Purpose and goal

This project strives towards significantly speeding up existing CFD solvers by integrating them with state-of-the-art machine learning (ML) methods. Within the project we will implement a graph-based neural acceleration model that aims to accelerate an iterative CFD solver, thus enhancing computational efficiency. The goal is to assess how much faster our hybrid physics-ML approach can make predictions, with a focus on turbulence modeling, while maintaining the same level of accuracy as fully physics-based alternatives.

Expected results and effects

CFD simulations are fundamental building blocks within technologies across a broad set of sectors and industries. Therefore, approaches that improve the speed or accuracy of such simulations will have massive impact. In this project we will evaluate simulations using our hybrid physics-ML approach and compare them with pure CFD methods as well as purely data-driven methods. The outcomes of Step 1 will inform us of which application areas and use-cases are most suited to the technology and will guide the choice of project partners for Step 2.

Planned approach and implementation

Modeling and model training will be carried out iteratively to ensure that data collection generates relevant results for the modeling at all modeling stages. Model evaluations will occur during M4-M7, where validation data is used to guide hyper-parameters choices during the modeling stage. Model comparisons will be conducted on a separate test set to avoid overfitting on validation data. Project management will run throughout the whole project, with early-stage focus on use-case validation with key stakeholders, and late-stage focus on reporting and preparations for Step 2.