Artificial Intelligence (AI) has exploded into almost every industry over the last few years and is expected to grow from $86.9 billion in revenue in 2022 to over $407 billion by 2027 [1]. This extraordinary growth may feel like it has sneaked up on us, but the machine learning modelling techniques that lie at the heart of AI have actually been around since 1956. The only difference is that now, advanced computing capability has enabled AI to be more accessible, operational and scalable.
What Is AI?
With the increasing prevalence of AI, its meaning has become somewhat confused. AI is essentially where machines, such as computers, simulate human intelligence. These machines use algorithms to consume and analyse training data to establish trends which can then be utilised to predict future conditions. Consequently, AI is capable of executing tasks and emulating human cognitive activity. Machine learning is a form of AI that enables computers to learn without being explicitly programmed and these algorithms can be grouped into three primary categories; supervised learning, unsupervised learning and reinforcement learning. Each category encompasses several branches of machine learning which are based on a distinct mathematical model, these include Neural Networks, Deep Learning, Decision Tree, Bayesian and Regression. There are then multiple variations of these mathematical models, with more discovered every year.
The Role Of AI In CFD
Whether it is a car, boat or plane, the changing landscape of powertrain technologies and regulation along with new companies entering the market has made the transport sector more competitive than ever. This is forcing manufacturers to respond quicker and significantly reduce the time to market of their products.
‘We are shortening development times by 25%,’ says Thomas Ulbrich, Member of the Board of Management of the Volkswagen brand in a recent interview. ‘In the future, vehicle projects will be created in 40 months instead of the previous 54 months.’
Alongside these shorter development times, the transition to hybrid and electric powertrains has placed even more importance on aerodynamics. A recent study showed that at speeds of 130km/h (78mph) around 80% of a vehicle’s power is used to overcome aerodynamic losses. Reducing this air resistance and improving aerodynamic performance by 10% can increase the range of an electric vehicle by 5-8% [2]. Therefore, as manufacturers hunt for every opportunity to improve the range of vehicles, the number of CFD simulations has rocketed.
This is where AI can help. Machine learning algorithms apply a new approach to CFD and use training data to develop models that allow engineers to simulate more iterations, in shorter times and view results instantly.
‘The conventional approach to updating a design involves completing numerous simulations and physical tests to assess its performance,’ explains Mattias Bauer, CEO of NAVASTO who supply AI CFD software solutions.
‘The design will then be optimised and re-tested until the performance goals are achieved. This can be a time-consuming process because you have to wait for the simulation results before modifying the design. Whereas, with AI models, the turn around time for simulation results is reduced to seconds which means that feedback is available in almost real time. This is game changing because it cuts the design cycle from days to seconds.
How Does AI Work In CFD Simulations?
AI is incorporated into CFD simulations through data driven models. This is essentially where an existing data set is used to train an AI model which correlates an input, such as an object’s shape and boundary conditions, with an output, such as drag coefficient or surface pressure.
This existing data can either be specifically generated for the project or based on legacy data. Once this AI model has been fully trained, it can then be used within a CFD simulation where it outputs the same results as a conventional simulation, but in real time. ‘AI is not magic, it is ultimately just mathematics, so an AI model will not be able to generate accurate results if the data used to train it is poor,’ highlights Bauer. ‘That’s why it is paramount that the training data is well converged and distributed within the design space.’
The CFD Problems AI Can Solve
There are two main approaches to using AI in CFD. The first is problem specific, where an engineer is working on a specific problem, such as optimising the performance of an active spoiler, for example. Here, the initial design space is defined and simulations are run on a number of sample cases to generate the necessary training data. Once complete, the output is an interactive AI model where parameters can then be changed and the corresponding results are instantly displayed.
‘This approach results in a model that was specifically trained to solve a particular problem and typically you only need around 20 to 30 initial sampling simulations,’ says Bauer. ‘So instead of waiting 12 hours for the results of each iteration, you can view the results in real time within your post processing software.’
The second approach is to create a general AI model that can predict the aerodynamic behaviour of completely new geometries using existing data. ‘By adding more and more data to the model over time, it’s possible to grow the model to incrementally increase its predictive scope,’ says Bauer.
Optimising The Active Spoiler of An Audi Sedan
The engineers at Audi wanted to establish the optimum position of the active spoiler. To achieve this, data from a total of 32 Design of Experiments (DoE) and Surrogate Based Optimisation (SBO) samples were used to train a problem specific AI model.
‘Instead of using trial and error, we worked with Audi to build an optimisation workflow that allowed them to identify the optimum position for the active spoiler,’ explains Bauer. ‘But we also generated an AI model which meant they could start from the aerodynamic optimum, but extend the search for a solution by including stakeholders from other departments,’ continues Bauer.
‘Within a meeting, aerodynamicists pitched ideas for alternative solutions and tested the impact on performance instantly. The real time feedback of the model allowed them to identify a solution that only marginally deteriorated aerodynamic performance, but was much more suitable from a structural point of view and acceptable to the style department.’
Reducing The Number Of CFD Simulations In The America’s Cup
Yacht racing teams competing in the America’s cup typically run thousands of CFD simulations to optimise the aerodynamic performance of the sails in a variety of wind speeds and directions. Existing CFD data from previous simulations was used to train an AI model which could then predict the performance of new sail geometries.
‘Our AI solution reduced the amount of simulations signficantly which not only saved the team money, but it also meant that they received the results weeks earlier than usual,’ highlights Bauer.
Unlimited Iterations In Formula 1
Without downforce, a Formula 1 car can only corner at approximately 1.5G. However, add a downforce package, and the same F1 car can achieve around four times the lateral acceleration at the corner apex. This is why aerodynamics is so vital in Formula 1 and why teams complete as much aerodynamic testing as possible, both on and off track.
To cap the amount of money spent on aerodynamic testing, Formula 1’s governing body, the FIA, has restricted the amount of wind tunnel and simulation time teams can have since 2009. The latest regulations are the most stringent yet and scale the allowable testing time on championship position; giving teams at the top end of the grid less testing time then those at the back of the field.
‘In Formula 1, the limited test times allowed by the rules means that the maximum amount of data needs to be extracted from each simulation,’ highlights Bauer. ‘Using our solution, Formula 1 teams trained AI models on relevant available datasets and could then test an unlimited number of additional configurations without using up any of their allowable CFD time.’
Can you trust results from AI?
As with all simulations, if the input data is of low quality then regardless of the modelling techniques used, the output data will also be low quality, and it is the same with AI. That’s why it is vital to provide the models with high quality training data.
‘Within our software we use a variety of machine learning and deep learning methods but these methods only give you a prediction quality,’ explains Bauer. ‘This prediction quality depends on the quality of the data provided to the models, and this is the most important consideration when using AI.’
‘People also think that AI can replace the first principal simulations entirely, but in my opinion, this is definitely not the case,’ continues Bauer. ‘Those first principal simulations are essential to generating accurate training data for the models, so AI might be able to decouple problem solving from running simulations, but it won’t replace them. AI is a tool that can help inform your decisions, it is not going to make the decisions for you.’
References
[1] R.W., K.H., 2023. 24 Top AI Statistics And Trends In 2023[Online]. Forbes Advisor.
[2] P.S., 2019. Extending electric vehicle range through a more aerodynamic design[Online]. Automotive Testing Technology International
