Read on to discover more about the work done and the remaining challenges in this interview with Dr. Eugene de Villiers.
In this written interview the managing director and co-founder of ENGYS explains the objective of the UPSCALE’s Work Package 1, leaded by his company: Machine Learning Enhanced simulation tools.
The main objective of the WP 1 is to enhance the performance of existing CFD and FEM Crash tools and processes using machine learning and model order reduction. These tools serve as the primary source of training data for more comprehensive real-time meta-models that may be used in the context of product design or directly for multi-point regulatory compliance testing. Work Package 1 is crucial to the success of UPSCALE. During the first year of the project the partners involved have been ramping up and building infrastructure to generate training data for the machine learning routines. Read on for details!
ENGYS offers CFD software products and services for engineering analysis and design optimisation based on open-source technologies.
The main picture shows the other partners involved on this work package.
Artificial Intelligence (AI) is not only changing what a vehicle can do, it is also changing how vehicles are built! In this sense the UPSCALE project could revolutionize the automotive sector, boosting the production of electrical vehicles by saving up to 20% of the development time. As leaders of the Work Package 1: Machine Learning Enhanced simulation tools, which are the main expected results?
The main goal of WP1 is to integrate machine learning into CFD and FEM (Crash) methods in an effort to improve the turnaround times of these tools. The expectation is that many of the critical and most computationally costly components of CAE tools can be replaced by equally accurate ML-derived surrogates. In a general sense then, the main anticipated outcome of WP1 is to realise the successful integration of machine learning methods with physics-derived algorithms such that the process as a whole becomes more efficient. In specific terms, we aim to create faster and more efficient solvers, mesher generators, collision detection algorithms, battery models, decomposition methods and turbulence models.
How the research on convolutional neural networks (ConVet) will improve the aero-thermal Computational Fluid Dynamics and Crash simulations?
For the most part, the project will use existing CNN technology. The innovation comes from the configuration of the training networks and application specific integration strategies. For example, in the context of external aerodynamics, our efforts will include:
- enhancing standard aerodynamic solvers with ML to speed up pressure projection so that they can be used to generate training data for the real-time models more quickly
- improving the accuracy of RANS turbulence models to provide the same accuracy as high-fidelity methods like LES/DES for a fraction of the cost
- configuring the training system that will allow neural networks to predict the flow field around electric passenger vehicles in real time
The vast majority of CFD problems are grid-based. How can the project improve and streamline the portioning of computational grids?
Grid partitioning algorithms are typically subject to a variety of inputs, many of which are suboptimal in terms of the eventual algorithm’s performance. We aim to use data driven methods to construct a surrogate model which can provide more appropriate graph decompositions. Improving the decomposition will in turn lead to more robust execution and faster solution times.
The UPSCALE project aims at optimizing the Finite Volume Acceleration solver based on the application of machine learning derived algorithms. How does this technology work?
At the core of machine learning is the idea of using past experience to predict the outcome of a given scenario. For the solver acceleration part of WP1 we aim to use this approach to replace at least a portion of the pressure solution stage of the algorithm. Since the pressure calculation consumes 50-60% of the solution time, significant gains can be had by using ML models in this context. We expect further advances to be possible by using a similar approach for other components of the algorithm.