Artificial Intelligence in Car Manufacturing

The automotive industry faces challenges; design teams in automotive companies are under pressure. Can artificial intelligence (AI) help product design teams in the automotive industry? How can supply chains become more efficient with AI and data science?

AI technology is all around us. Many have heard of - or are already using - advanced driver assistance systems. Final users will perceive the driving experience of AI-assisted cars. The automotive industry can benefit from IoT sensors acting in the production line to enhance quality control at unprecedented levels for automotive manufacturers.

The examples mentioned come from user experience or production lines.

However, we will focus on manufacturing applications:

  • What is the impact of Artificial Intelligence upstream of manufacturing?
  • What is the effect of Artificial Intelligence on product design and program management?
  • Can the supply chain between Tier 1 suppliers and OEMs become leaner?
  • At the same time, should it be more productive on the high-tech aspects of design, such as CAE simulation?

Automotive Industry - Role of Tier 1 in the Supply Chain

Let's take a step back before speaking about AI technology. We will look at the organization of the supply chain in the automotive industry. We will identify where AI technology can help in any automotive sector.

Who Are Tier 1 Suppliers in the Automotive Industry?

Tier 1 suppliers play a fundamental role in traditional, hybrid or electric vehicles. We must pay attention to the part of Tier suppliers. Their relationship with the OEM is different from commodity suppliers.

Automotive manufacturers (OEMs) assume that Tier 1 will provide added value to them. Tier 1 suppliers must be competent in design and manufacturing. This competence is both a power and a responsibility when dealing with the OEM customer.

IMAGE SOURCE https://4.bp.blogspot.com/-g4lifUhJv8s/Uat5pVe7j7I/AAAAAAAAKng/zannRtE00bs/s400/Peugeot+Citroen+1.jpg

From RFQ to Production

The RFQ (request for quotation) phase is when the competition with other suppliers officially starts. Tier 1 companies are invited to submit technical and commercial proposals to become equipment suppliers for one or more vehicle models branded by an OEM.

Continuous collaboration between the selected suppliers and the OEM continues after closing the contract. What is the difference between aftermarket supply and original equipment supply? Aftermarket sells off-shelf components; instead, an OEM supplier must adjust to every new need until and often after SOP (start of production).

OEMs need suppliers to meet functional requirements, such as durability, energy efficiency, etc. How to ensure meeting them? Until the 90s, suppliers relied on physical prototypes and testing cycles. Later, Computer-Aided Engineering (CAE) coupled with Computer-Aided Design (CAD) helped to shorten the above cycles' timing and costs. This revolution was from analogue to digital (software 1.0) engineering. Finally, after the analogue to digital software 1.0 revolution, a 2.0 revolution is taking place.

Artificial Intelligence: Software 2.0 Comes to Help Tier 1 and OEMs

Software 2.0 is "software eating other software" that gets its superpowers (i.e. real-time predictive analytics capabilities) by feeding on data. Data come from "traditional" software 1.0, such as the above-mentioned CAD and CAE.

The Software 2.0 revolution is creating design tools that learn from our experience and help us designing products better and faster.

Automotive Industry - Technical Specialization in the Tier 1 Supply Chain

We can focus on the details of a traditional car radiator. We can see very different manufacturing processes, engineering skills and supply chain members at play.

In the figure, water pumps, coolant tanks and radiator engines are three components involving three suppliers.

IMAGE SOURCE: https://i.stack.imgur.com/nNHPT.jpg

Different laboratories and equipment and different engineers will manage them. All this underlines a high degree of technical specialization.

How Is Artificial Intelligence Changing the Automotive Industry?

We will show examples of how car manufacturers (OEMs) and their Tier 1 suppliers use AI technology in the automotive industry. It will be clearly demonstrated how Artificial Intelligence can improve business support functions for traditional methods. We will talk about self driving cars and autonomous driving technology, the manufacturing floor and finally, more in detail, we will speak about its impact on product design.

Autonomous Vehicles

We will touch on autonomous vehicles and move on to the manufacturing floor and the design department. Autonomous vehicles rely on various technologies to face driving challenges. Such challenges can be navigating roads, avoiding obstacles, and take crucial decisions. Some of the key AI technologies used by autonomous vehicles include the below. We will pass them in a quick review and move on to manufacturing and design topics in the next section.

Autonomous vehicles use computer vision algorithms that interpret the data from sensors. This allows them to understand the environment around them. Artificial Intelligence can identify objects in the background. Examples of obstacles for autonomous cars are pedestrians, other vehicles, and traffic signs (figure below).

IMAGE SOURCE https://www.cvlibs.net/datasets/karlsruhe_objects/objects.jpg

Applications of AI in Car Manufacturing

IMAGE SOURCE: https://62e528761d0685343e1c-f3d1b99a743ffa4142d9d7f1978d9686.ssl.cf2.rackcdn.com/files/67279/area14mp/image-20141215-24297-6jsbmp.jpg

Artificial intelligence (AI) has many potential applications in car manufacturing. Some examples, to only name three, include quality control, predictive maintenance and supply chain optimization.

The Role of Generative Design

Imagine that car manufacturers want to design a new sports car optimized for aerodynamics with AI systems. A machine learning model, trained on a dataset of existing sports car shapes, could generate thousands of designs in a short time.

IMAGE SOURCE: Anthony Massobrio

Those designs incorporate various features, such as spoilers, air dams, and diffusers.

The first question is, what will be the aerodynamic performance associated with those changes?

There is a need for predictive analytics at play in sync with design changes. Here, AI-based prediction can come to help AI-based geometry changes. Working together, shape change and engineering prediction can produce AI-optimized aerodynamic shapes.

The manufacturer could then downselect the best design options in these simulations. Final refinement through human design and final testing/simulation with traditional methods.

DEBOSH (Deep Bayesian Optimization) can morph a baseline design (left) into an aerodynamically optimal shape (right)
DEBOSH (Deep Bayesian Optimization) can morph a baseline design (left) into an aerodynamically optimal shape (right)

Practical Use Cases of AI Applied to Car Design

We started with general considerations and an overview of AI in autonomous driving and manufacturing. Now, we will dive deep into three applications of AI applied to vehicle design, with the specific aid of data coming from CAD and CAE.

Automotive Industry Case 1 - Passenger Car Aerodynamics

Several important topics need to be considered in the aerodynamic design of a passenger car. Many of them can be tackled with aerodynamic simulations (CFD) or AI predictions.

We will review three topics. They impact battery/fuel consumption, vehicle handling, and driver/passenger comfort. The topics are drag, downforce and (aero)acoustics. Software 1.0 such as CAE / CFD can simulate all within hours. The Software 2.0 paradigm makes it possible to get answers with AI / Deep Learning within seconds.

  1.  Drag

Reducing the drag of a car can improve its fuel efficiency and peak speed. There are many ways to reduce the drag coefficient. Classic approaches include streamlining the car's shape and using active aerodynamic devices. Examples of aero devices are spoilers and air dams. We will talk more about spoilers (see figure).

Aerodynamic devices such as spoilers can contribute to better performance in terms of downforce and aerodynamic resistance
Aerodynamic devices such as spoilers can contribute to better performance in terms of downforce and aerodynamic resistance
  1. Downforce

Generating downforce (negative lift) can help improve the vehicle's handling and stability at high speeds or in specific situations such as curves. This can be achieved using active aerodynamic devices such as spoilers and airfoils.

  1. Noise

A source of annoyance for drivers and passengers is internal and external noise. External noise is generated by air flowing around the car and specific devices such as side mirrors (see figure). The aerodynamic design of a car should consider the noise generated by the airflow around the vehicle to reduce it.

Generative design of a side mirror in a car to optimize its aerodynamic / aeroacoustic performance
Generative design of a side mirror in a car to optimize its aerodynamic / aeroacoustic performance

Passenger Car Aerodynamics: Reduce Resource Requirements

All the mentioned topics can be reduced to KPIs that need optimization. Drag should be minimized for consumption. Downforce should be maximized for stability. Aeroacoustic noise should be minimized for comfort.

In theory, it is easy to conceive a CFD simulation supporting an optimization campaign for the KPIs.

The two questions are: how many hardware resources and highly specialized skills are required? Many. How many designers can access CFD? Just a few.

Bringing Advanced Skills to more Ressources and Without Toils: AI

The learning process of AI systems can bring CFD, under controlled situations, to the vast majority of engineers. The outcome, shown in the figure as "prediction", is virtually identical to the CFD simulation, shown as "Fields" in the figure.

As a target, let's envisage engineers who are not CFD specialists. Also, we envisage avoiding requirements for HPC (local cluster or cloud) resources, such as in the figure.

AI applications can run on local computers or even laptops.

This is a piece of very welcome news for the automotive industry. "AI in automotive" was usually associated with collaborative robots or self driving vehicles.

We will see how AI can impact the automotive value chain much earlier, not only in the new autonomous vehicles market.

Snowball Effect of AI on Design in the Automotive Industry

What is the relationship between snowballs, AI and the automotive industry?

An AI investment could have a revolutionary effect when the AI solution is deployed in early concept phases.

We propose the analogy with a snowball rolling down a snow-covered mountainside; as the ball rolls, it will pick up more snow, gaining more mass, and so on. In aerospace engineering, it is also used to describe the multiplication effect in an original weight saving.

In the automotive design department, implementing AI in the early concept phase (e.g. when C-Levels discuss the launch of a new car) can have a deep impact on the rest of the car's projects. The positive and beneficial avalanche will be a much wider range of design possibilities from which optimal solutions will be selected at earlier stages than with traditional methods.

In fact, AI allows the simulation of car performance well before the production and assembly lines, and even before detailed CAD work. Thus, AI-based simulation integrated with shape modifications is an ideal tool to support high-level meetings. During meetings, technical assessments on design changes are obtainable in real-time on any platform, such as a laptop. This is quite a dramatic change from the usual process where management and designers had to raise a request for a simulation to the CAE department, and receive the feedback and discussion about the simulation results after hours or even days or weeks.

New Demographics of Engineers in the Automotive Industry

There are probably, in most companies, 10 to 100 times more design engineers than specialized CFD and CAE engineers. Mass deployment of CFD and CAE is conceivable with AI-based simulation. These are the three main enablers making AI-based simulation democratic and therefore mass-deployable.

  1. AI can be deployed in an easy App for Design Engineers in the automobile industry who do not need to be experts in AI.
  2. Thanks to an artificial network structure, AI acts in seconds instead of hours or days
  3. AI processes industrial geometries (CAD) without requiring time-consuming and specialist software such as "meshers" or solvers.

Therefore, Software 2.0 will not require any "Effort 2.0". The final deployment of Deep Learning is not even easy, it is just straightforward.

Technical Details - What Does It Mean That AI Systems Learn?

In this figure, we see a CNN at work. But what is a CNN (convolutional neural network), and how can software learn?

A CNN is a neural network designed for image recognition and processing. It comprises multiple layers of interconnected neurons. Layered artificial neurons process and analyze visual data using a hierarchical approach. The lowest layers recognize basic features such as edges and shapes. The higher layers combine and interpret these features to identify more complex patterns and objects. This process can be seen in the figure, with more details learnt while proceeding from left to right. CNNs are particularly effective for tasks such as image classification. Associating an engineering result to a CAD input is a sort of image classification, although a very advanced one.

Geodesic Convolutional Neural Network learning to associate an input F (right) to a CAD input X (left).
Geodesic Convolutional Neural Network learning to associate an input F (right) to a CAD input X (left)

Usual popularizations of Deep Learning use 2D images. The nature of output data (FEA or CFD results) and input data (3D CAD geometry) is more complex than that!

Technical Details - Efficiency in the Learning Process

We will show how the efficiency of the learning process can be increased thanks to specific Transfer Learning techniques. The envisaged adoption of AI for simulation becomes much less consuming than expected. As example; Neural Concept Shape decreased by 97,5% the CFD simulations needed for car aerodynamics while meeting all functional requirements.

Automotive Industry Case 2 - F1 Car Aerodynamics

A fascinating and challenging benchmark for any technological application is Formula 1. To summarize, from the aerodynamic point of view, there are three significant technical challenges and a political-organizational challenge in F1. We will see how AI in the Automotive sector and specifically Aerodynamics can solve technical and organizational challenges.

Technical Challenges in F1

The technical challenges in aerodynamics are different from passenger cars. In an F1 car, one should focus on creating as much downforce as possible while sustaining straight-line speed. This allows the driver to have a better grip and to handle through turns. Downforce is so powerful that it could enable driving on a ceiling!

Reducing drag force is also essential in F1 aerodynamics because it can help increase the car's top speed. However, it is less relevant than in passenger cars.

Another aspect more relevant than in passenger cars is aeroelasticity. The F1 car has a lot of external structures exposed to aerodynamic loads. The car's body and wings can flex and deform under such loads, which can affect the car's performance, as seen in a few spectacular accidents.

Organizational Challenges in F1

A final point that could be crucial in AI development is the Regulations. The F1 world has strict regulations on the design of the car's aerodynamics, including rules on the shape and size of the wings, the front and rear diffusers, and other aerodynamic devices. These regulations can limit the design options available to the aerodynamics team. But, there are also regulations on the number of wind tunnel testing runs and CFD simulation hours that can be utilized to develop a car. This is different from car manufacturers, where external controls do not regulate the usage of resources.

Practical Case of Aerodynamic Design in F1 - Traditional Approach

Let us now investigate the case of an F1 designer who needs to examine five different car designs as in the figure (superposed).

Five different F1 designs
Five different F1 designs

Traditionally, the designer would iterate with her CFD department in order to continuously have insights on what is virtually happening in the flow field (figure below) and computed aerodynamic parameters such as the downforce.

Aerodynamic F1 simulation
Aerodynamic F1 simulation

With traditional simulation approaches (this notion is essential, in order to understand the benefits of AI!), the passage from the CAD to the CFD image is not immediate. In the CFD department of the F1 team, full-time specialists will take care of several phases.

A more advanced scenario envisages the users sitting in front of a portal (web-based service). Users give as input their CAD file and a few parameters, waiting again some time to get an automated aerodynamic report. In any case, an essential factor is the time of response.

Practical Case of Aerodynamic Design in F1 - AI Approach

Can AI cut the waiting time down from 3 hours to 0.3 seconds? This would allow designers to manage several orders of designs to sweep a more extensive and detailed design space with AI. One can focus on final converged solutions with CFD and/or wind tunnel testing.

The designer would want surface and volumetric data. The most important surface data is represented by surface pressure (figure below).

Examples are pressure and velocity fields in the vehicle cross-sections.

Here the designer can check aerodynamics in a reduced timeframe (0.3 seconds on a laptop Vs 3 hours on a compute cluster). The aero detail is the evolution of vortices in various vehicle sections. There is nothing significantly different in the results. What is essential is that the CFD solver time becomes orders of magnitude smaller with AI predictions. Also, the meshing part is virtually inexistent.

Aerodynamics of F1 car in a cross section showing streamlines. Mirrored comparison of CFD (known as Groundtruth) and AI (known as Prediction).
Aerodynamics of F1 car in a cross section showing streamlines. Mirrored comparison of CFD (Ground truth) and AI (Prediction)

This type of capability is soon going to revolutionize F1 as well as other applications of CFD and FEA to automotive and yacht racing.

Automotive Industry Case 3 - HVAC Simulation

Computational fluid dynamics (CFD) simulation can be used to optimize the design of a car's heating, ventilation, and air conditioning (HVAC) system by predicting how the air will flow through the system and how it will affect the temperature and comfort of the occupants.

However, CFD is an expert topic reserved for a few specialists and only for some HVAC designers; as it is the case throughout the automotive industry. How can AI technology help HVC designers? Can AI in the automotive industry bridge the gaps between customers (OEM), designers and simulation specialists?

Neural Concept Shape shows how AI enables real-time simulation for Tier 1 components and assemblies.

Big data coming from years or decades of experience in the auto industry, coupled with recent advances in AI and machine learning, are transforming the automotive industry for next-generation cars and can come to help HVAC designers in two ways:

  1. Fast thermal and flow predictions associated with designs with AI-based algorithms. AI algorithms can be used to build data-driven predictive models that can estimate the performance of an HVAC system based on its design and ambient temperature conditions. These models can help designers optimize the size and configuration of the system.
  2. Creation of new designs (optimization) with an AI system. AI optimization algorithms can be used to search for the optimal design of an HVAC system based on a set of constraints and objectives and, again, on the previous AI predictive capabilities. These algorithms can consider many design variables in almost real-time and find the best combination that meets the desired performance targets.

 Retrieving Data and Using Data

We will now examine a practical use case of HVAC simulation (figure below).

Improved business support functions in Engineering are shown by the implemented AI system from Neural Concept. The AI tool (NCS) is combined with datasets taken from the company's historical experiences and adequately stored in the PLM system. It is essential when talking about "Big Data" (one of the emerging keywords when talking about technological trends), to practically track down where those Big Data are located.

In the PLM, all types of data following the product history, ranging from CAD and simulation data to sensor data, can be stored. They are just waiting to be leveraged by Neural Concept Shape in the learning process of its Deep Learning AI technology!

In the HVAC design scenario, the inputs to the AI tool provided by product designers are raw CAD geometries and operating parameters (e.g. inlet air mass flow rate and temperature). The expected outputs are air pressure, velocity and temperature, available in 3 seconds after importing a CAD.

The predictive tool is much more than "easy-to-use software for designers". Its usage is immediate in a design environment because it encapsulates a workflow commissioned by the final users and tailored to their personalized needs.

HVAC Case Summary

We could demystify AI by calling it a "Software 2.0" approach, i.e., software fed with traditional software inputs such as CAD and CFD. The data shown in the above Case History N°1 is much smaller than expected. Instead of the scary scenarios of 1'000+ needed samples, here, with only 100 samples, AI was able to reach the desired accuracy.

The key ingredients in data efficiency are two:

  1. Using a non-parametric, generalized free reading of CAD data. This is possible with Neural Concept Shape thanks to its non-parametric nature.
  2. Leveraging transfer learning. With transfer learning, one can "restart" AI Deep Learning not from the default scratch but from a pre-trained predictive model.

Conclusions

AI and deep learning can potentially revolutionize the automotive industry design process, starting from the earliest product concept stage. We have shown how AI can assist the automotive industry in the following ways, starting from the concept phase of vehicle design:

  • Democratization: Deep Learning can extract knowledge from previous CFD/CAE simulations and make it available to design engineers.
  • Broader & Faster Exploration of Design Space: Deep Learning can perform +1'000 times faster than CFD/CAE, thus allowing the exploration of +1'000 designs within the same amount of time.
About the author
Anthony Massobrio
Anthony is a CFD expert since 1990, working initially as a senior researcher, then moved to Engineering, acting also as technical director in a challenging Automotive Tier 1 supplier environment. Since 2001, Anthony has worked in Software & Engineering Consultancy as a Sales Engineer and manager. In 2020, Anthony fell in love with AI and has worked since then in the field of “AI for CAE” at Neural Concept and as an independent contributor.
About the author
Anthony Massobrio
Anthony is a CFD expert since 1990, working initially as a senior researcher, then moved to Engineering, acting also as technical director in a challenging Automotive Tier 1 supplier environment. Since 2001, Anthony has worked in Software & Engineering Consultancy as a Sales Engineer and manager. In 2020, Anthony fell in love with AI and has worked since then in the field of “AI for CAE” at Neural Concept and as an independent contributor.
About the author
Anthony Massobrio
Anthony is a CFD expert since 1990, working initially as a senior researcher, then moved to Engineering, acting also as technical director in a challenging Automotive Tier 1 supplier environment. Since 2001, Anthony has worked in Software & Engineering Consultancy as a Sales Engineer and manager. In 2020, Anthony fell in love with AI and has worked since then in the field of “AI for CAE” at Neural Concept and as an independent contributor.
About the author
Anthony Massobrio
Anthony is a CFD expert since 1990, working initially as a senior researcher, then moved to Engineering, acting also as technical director in a challenging Automotive Tier 1 supplier environment. Since 2001, Anthony has worked in Software & Engineering Consultancy as a Sales Engineer and manager. In 2020, Anthony fell in love with AI and has worked since then in the field of “AI for CAE” at Neural Concept and as an independent contributor.
About the author
Anthony Massobrio
Anthony is a CFD expert since 1990, working initially as a senior researcher, then moved to Engineering, acting also as technical director in a challenging Automotive Tier 1 supplier environment. Since 2001, Anthony has worked in Software & Engineering Consultancy as a Sales Engineer and manager. In 2020, Anthony fell in love with AI and has worked since then in the field of “AI for CAE” at Neural Concept and as an independent contributor.
About the author
Anthony Massobrio
Anthony is a CFD expert since 1990, working initially as a senior researcher, then moved to Engineering, acting also as technical director in a challenging Automotive Tier 1 supplier environment. Since 2001, Anthony has worked in Software & Engineering Consultancy as a Sales Engineer and manager. In 2020, Anthony fell in love with AI and has worked since then in the field of “AI for CAE” at Neural Concept and as an independent contributor.
About the author
Anthony Massobrio
Anthony is a CFD expert since 1990, working initially as a senior researcher, then moved to Engineering, acting also as technical director in a challenging Automotive Tier 1 supplier environment. Since 2001, Anthony has worked in Software & Engineering Consultancy as a Sales Engineer and manager. In 2020, Anthony fell in love with AI and has worked since then in the field of “AI for CAE” at Neural Concept and as an independent contributor.
About the author
Luca Zampieri
Luca joined the Application team in 2018, aiming to build the next generation Deep-Learning tool dedicated to CAD and CAE.
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
About the author
Thomas von Tschammer
Thomas joined the team in 2018 as Director of Operations, aiming to empower engineers with next a next generation Deep-Learning tool dedicated to CAD and CAE.
About the author
About the author
About the author
About the author
About the author
About the author
About the author