A 10% improvement in aerodynamic performance can lead to roughly a 5% increase in the range of an Electric Vehicle (EV) [1]. Range is one of the most important performance differentiators of an EV, so automotive manufacturers are doing everything they can to increase it. This is making aerodynamics more of a priority than ever before.
Optimising the geometry of a vehicle to minimise drag and maximise downforce whilst meeting the styling requirements of the brand is no easy task. It requires constant collaboration between the design and aerodynamic departments and typically takes multiple years of development involving potentially hundreds of design iterations and thousands of numerical simulations.
However, the recent surge of Artificial Intelligence (AI) techniques, such as Machine Learning, is revolutionizing this process. To investigate the potential of Machine Learning in Computational Fluid Dynamics (CFD), we conducted a proportion study using NAVPACK AI software on an AeroSUV model in ELEMENTS. Continue reading to discover what we found.
The Conventional Approach To Proportion Studies
One of the first tasks of improving the aerodynamics of a vehicle is to first understand how its shape influences the airflow, which is achieved with a proportion study. This essentially takes a CAD model of the initial design and morphs the main features through a batch of parameters, calculating the sensitivity of drag and lift.
‘Typically, you are analysing big-ticket proportion changes to the vehicle’s shape such as the front and rear overhang, roof height, windshield angle and boat tailing,’ highlights Angus Lock, General Manager at ENGYS. ‘By doing this early in the development program you gain feedback on the areas of the car that are particularly sensitive to aerodynamics. This helps guide the design studio on how much they can move surfaces around and the likely effect on aerodynamic forces.’
Instead of morphing the features through every incremental change for every parameter, the Latin Hypercube sampling method is used to run simulations at random points across the parameter space. This significantly reduces the number of simulations, whilst still providing a representative picture of the sensitivities. Once all these simulations are complete, this data is then used to generate a response surface which helps aerodynamicists identify the optimal geometry.
In the case of the AeroSUV model, 12 different parameters were analysed through 117 samples. Each steady state simulation was run on 300 cores and took around 3 hours, totalling 2 weeks of simulation time.
To accelerate this process, companies typically run half car models or truncate the run time. The normal run time of a simulation results in fully converged flow fields, but the forces tend to converge a little earlier. Consequently, the run time can be shortened, and the aerodynamic forces still obtained.
‘Despite these tactics, proportion studies can still be time consuming,’ explains Lock. ‘If the aero or design teams want to investigate a new shape, then new geometry has to be created, or morphed. This then has to be run in a traditional manner, so there is no immediate feedback. But by using AI, you can get results in seconds.’
The Machine Learning Approach to Proportion Studies
Over the last few years AI has gradually infiltrated into CFD workflows across industry. Automotive companies, Formula 1 and America’s Cup teams are already benefiting from its capabilities. AI is essentially a clever form of interpolation, where Machine Learning algorithms analyse training data to understand trends and predict the next data point.
In CFD, existing data sets are used to train AI models which correlate an input with an output, such as a geometry change with drag coefficient. This model is used within a CFD simulation and because it does not need to resolve the Navier Stokes equations at each point, it can generate results much faster. Consequently, aerodynamicists can simulate more iterations, in shorter run times and view results immediately.
For the AeroSUV model, NAVPACK software trained a model using Graph Neural Networks (GNN). This is a Machine Learning technique which processes objects as graphs made up of points (nodes) connected by lines (edges). It can generate predictions at the node, edge and graph levels as well as the relationships between them.
‘NAVPACK uses GNNs to map the behaviour of the function you are analysing directly to the shape of the object you are training,’ highlights Lock. ‘So you can take historic data, train an AI model, morph the model in real time and then click predict to get CFD results instantly. This has the potential to fundamentally change how aerodynamic and design departments work together.’
Training An Accurate AI Model
One of the common concerns with AI is trusting its accuracy. However this is simply determined by the quality of the model, which in turn depends on the training data used. Typically, around 200 simulations are required to train an AI model effectively, although this varies according to the resolution of the geometry being analysed. This may sound like a lot of simulations, but historical or existing data can be used.
‘The amount of samples you need to train an AI model depends on several factors,’ reveals Matthias Bauer, CEO of NAVASTO. ‘The first is the quality of the data. If the data is noisy then you need more samples, but if it is high quality then you need less samples.’
‘The way the data is produced also has an effect,’ continues Bauer. ‘If you run a design of experiments that samples in a space filling manner across the entire design space then less samples are required. But if you use existing data that has been generated for other studies, this can be less representative. This is because these studies typically focus on a specific aspect of a vehicle, which leads to clusters of data points and areas of the design space that have not been sampled.’
NAVASTO have spent the last 10 years developing their software to reduce the amount of training data required. They’ve achieved this through clever statistical techniques, allowing AI models to be trained with as few as 30 samples, with the AeroSUV model requiring only 110 samples.
‘It wasn’t that long ago when you needed thousands of CFD simulations to train an AI model,’ says Lock. ‘This made integrating Machine Learning in aerodynamic development expensive for companies. One of the benefits of NAVASTO is that they have lowered this threshold to a few hundred simulations, making AI much more achievable.’
Hardware Requirements Of Machine Learning
Companies can also be apprehensive about the hardware costs of integrating Machine Learning in CFD, but these are often less than you think. Like CFD, the complexity and run time of a simulation differs greatly if it is run on a laptop compared to High Performance Computing (HPC). However, most AI models can train on a single GPU, as it is the Video Random Access Memory (VRAM) that is most important.
‘The amount of VRAM decides how large the AI model can be, how detailed the geometry is and how many variables can be included within the model,’ explains Bauer. ‘The most efficient approach is to store this onto a single GPU as otherwise you lose performance if you distribute it across several GPUs.’
‘We aim to store the geometry, the Machine Learning model and the variables on a single 40GB or 80GB GPU. But of course the more, the better,’ continues Bauer. ‘This is a similar level of investment to purchasing a high-end workstation.’
Case Study Results
Once the AeroSUV model was morphed in Blender, CFD simulations were run in ELEMENTS software which then provided the training data to NAVPACK’s GNN AI model.
In just two days, the model showed strong correlation with the numerical training data, achieving a correlation coefficient of R² = 0.951 for drag and R² = 0.958 for lift. The predictions of this model compared to unseen data also showed good accuracy with a correlation coefficient of R² = 0.912 for drag and R² = 0.918 for lift.
‘We determined several things from this study,’ highlights Lock. ‘Firstly, we were impressed with the accuracy of the AI model, particularly as we only used 110 simulations to train it, which is relatively low. This did mean we were limited to fairly small morphing changes, so if we were to do this again we would grow the AI model to investigate larger modifications and simulate more parameters.’
‘I would also adapt the morphing cage around the model,’ continues Lock. ‘This approach is great for traditional parametric style optimisation because it’s easy to go back into the results and tease out the drag sensitivity of the windshield angle for example. But to complete more freestyle changes, you require more freedom. Overall, if you want to optimise particular parts of a vehicle through a bunch of different parameters then using AI is a way of getting to the optimum very quickly.’
Conclusion
There is no question that adopting AI techniques such as Machine Learning in CFD workflows is a faster way to reach results. While the scalability and open-source nature of CFD solutions, like ELEMENTS, allows engineers to run hundreds of cases to train AI models with greater concurrency than traditional CPU or instance-based CFD licence schemes. Combine this with the performance of the latest CFD software and the lead times for generating AI models significantly reduces.
This has helped automotive companies shorten CFD processing times from weeks to seconds, especially in the early design phases. While America’s Cup teams who typically complete 1,000 simulations, now only have to run 200 cases and the AI model predicts the remaining 800.
‘AI and Machine Learning will not replace simulations or experiments, it is simply another tool in an aerodynamicist’s toolbox,’ concludes Bauer. ‘So actually, it is not solely a question of accuracy, because what is the ground truth? Is it wind tunnel data, real world testing or CFD? As with any tool, the trends are more important, than the absolute values.’
‘We need to leverage the advantages of AI, which is speed and effectiveness,’ continues Bauer. ‘AI should not be used to replace first principles, because that’s not possible. AI has the power to streamline aerodynamic development whilst also connecting departments within companies. Instead of throwing CFD simulation data away, it can be used to build AI models which are then accessible to everyone in the organisation, whether they are an aerodynamicist or not.’
References
[1] 2019. Extending electric vehicle range through a more aerodynamic design [Online]. Automotive Testing Technology International.
You must be logged in to post a comment.