Optimising automotive performance with AI
The automotive industry is accelerating into an uncertain future where increased design complexity coupled with mounting regulations and customer demands are impacting not only tomorrow’s vehicles but the engineers developing them, today.
Gary Brotman, chief executive officer at machine learning company Secondmind, said: “Today’s automotive engineers face a herculean task in efficiently and effectively modelling and simulating the systems that comprise today's cars. The complexity of vehicle software alone is 400 % greater than the pace of development productivity - complexity driven by an increasing number of design parameters and constraints, including tighter emissions and safety regulations, fuel economy, aerodynamics, drivability, cost and consumer desires, to name a few.”
Brotman added: “There's also the challenge of integrating multiple systems for hybrid and electric vehicles and achieving optimal performance in development and throughout the vehicle lifecycle. And for traditional car makers, the growing pressure to reduce emissions means engineers must develop innovative, long-term solutions while optimising legacy ones, which is costly, time-consuming and makes it harder to maintain competitive advantage, let alone survive.”
In short, automotive engineers are expected to do more with less. Increasing vehicle complexity sits at the heart of the many design and development challenges that engineers face. This is because today’s vehicles have moved away from mechanically dominated systems and into multi-physics, fully connected entities, driven by data and powered by electricity.
Zed Tang, technical account director at Ansys, explained: “The demand for designing more efficient and powerful electrified powertrains at a lower cost has incrementally increased. When we consider where a great source of cost savings and efficiency in the overall system can be found, the Electric Drive Unit (EDU) is it.”
An EDU comprises the power electronics consisting of the control software, the gearbox, and the electric machine – all of which must seamlessly work together to move the vehicle. “So, the designer must consider system performance instead of a single component and the interactions between the different units to meet these challenges. The only way to understand the system performance is through system-level simulations,” Tang added.
The development of a battery EV powertrain, for example, is another complex systems problem. Ansys Motor-CAD simulations were recently used to help engineers determine whether an interior permanent magnet (IPM), an induction magnet (IM) or a wound field synchronous magnet (WFSM) is the best motor design for an EV, examining the resulting trade-offs to identify the best solution.
The e-motor and battery pack are both further examples of where multi-physics designs are required. The design and development of each system require an understanding of thermo-electric behaviour and cooling, in addition to traditional structural mechanics.
Royston Jones, global head of automotive at Altair, explained: “Considering the tight limits on costs and the industry’s competitive time-to-market, developing innovative electric powertrains is a challenge for any manufacturer. Altair’s simulation optimisation technology can minimise weight and increase performance by rapidly exploring thousands of design permutations.”
Altair e-Motor Director lets users easily define a broad design space for a single baseline concept, where they can then copy, paste, and change design concepts to build a design of experiments (DoE) database with design information.
The solution creates competitive motor designs while considering all project requirements, according to Jones, who said: “Altair e-Motor Director mixes the advantages of simulation- and data-driven design to weigh conflicting constraints from multiple physics so you can accelerate the development of cutting-edge e-powertrains.”
“Altair e-Motor Director creates, manages, and stores study descriptions for multi-domain analyses, and handles design of experiments (DoE) data associated with your design studies. It also enables optimisation to identify a motor family or a specific motor based on a single DoE or multiple DoE studies.”
But advancing automotive electrification encompasses much more than a shift from an internal combustion engine (ICE) to battery power. Tang explained: “Infrastructure, maintenance, and a host of other variables must be considered. Currently, the automotive industry is facing a big challenge from other industries and other locations looking for expertise.”
Tang added: “There’s a workforce shortage everywhere, but even more so on the electrification side. If automakers can’t hire enough engineers, they need to think of ways to empower their engineers to be more efficient and effective.”
Optimisation is a key technology to help engineers who are working with increasingly tight deadlines and complex designs.
Altair’s solutions in automotive vehicle development, such as multi-disciplinary optimisation (MDO), are designed to optimise “skateboard” platform designs for a wide range of use cases (including Battery EV, Hybrid EV, Plug-in Hybrid EV), as well as different vehicle types to improve performance and reduce development costs.
Jones added: “Engineers will need to adopt optimisation tools – if they don’t their competitor OEMs will. Optimisation fits the artificial intelligence (AI) narrative. Traditionally, engineering has been data poor, so optimisation technology has been key.”
Jones added: “In the future, we will be data rich and that data will come from everywhere (for example: the field, warranty, physical testing, manufacturing etc) including massive amounts of synthetic (simulation) data. Engineers will become more conversant with data analytics techniques to provide increased insight into the product since the product complexity and refinement level will continue to increase.”
There are now a range of novel machine learning tools coming online to address these design challenges. Cambridge-based start-up Secondmind, for example, uses a specific branch of machine learning to help optimise the development of tomorrow’s vehicles (see the ‘Intelligent Engineering’ box).
Altair's digital twin solutions also use simulation, machine learning, and artificial intelligence to create virtual representations of physical assets. Jones explained: “Optimisation has been in Altair’s DNA for well over two decades and recently, with the ability to generate large volumes of synthetic (simulation) data we can efficiently utilize machine learning (ML).”
Jones added: “In Altair e-Motor Director, we utilize both optimisation and ML technologies to ensure that from a vast array of motor family permutations the correct e-Motor topology is selected with the ML ensuring a manufacturable shape has been selected.”
Likewise, Ansys has expanded its multiphysics capabilities “beyond automotive structure to meet challenges involving electromagnetics, controls, functional safety, reliability, materials intelligence, and more,” Tang explained. “Beyond Ansys multiphysics simulation capabilities, we have streamlined workflows, design automation and optimisation, high-performance computing (HPC), and cloud-enabled solutions to assist engineering teams.”
Whatever the underlying technology that’s sitting under the hood, it’s clear that a range of tools are now vital to streamline the design and development of tomorrow’s vehicles - and that machine learning and optimisation will continue to play a significant role, going forward.
Secondmind is focused on using model-based design and optimisation solutions to overcome the many challenges today’s automotive engineers face. Scientific Computing World caught up with the company’s chief executive officer, Gary Brotman, to find out more about the Secondmind’s machine learning approach and work in the automotive industry.
Q: How is Secondmind addressing the challenges that today’s automotive engineers face and what makes it different from other simulation solutions?
A: The challenge of virtually achieving design and performance reliability through simulations that translate to validity and effectiveness in real-world scenarios, an already difficult exercise, is made even harder by the growing glut of data bogging down the design process. As a result, existing vehicle design and development technologies used to build complex models for simulating and optimising new components, systems and materials are struggling to keep pace at a time when they are needed the most.
The Secondmind Optimization Engine powers our cloud-based solutions for vehicle system design and control system calibration. And because it’s cloud-native it has the ability to continuously optimise the performance of complex systems throughout the vehicle lifecycle.
The Optimization Engine is designed from the ground up to solve the most difficult engineering problems in automotive and address the shortcomings of other AI-based solutions, by enabling intelligent, automated experiments, modelling of physical and virtual data, and providing engineers with better choices in design simulation and testing. The result is higher precision prototype designs, faster and more accurate performance optimisation, and less rework throughout the design and development process.
Most importantly, Secondmind slashes data dependencies by up to 80 percent to facilitate design and performance optimisation of high-dimensional problems more efficiently and accurately.
A good example of powertrain optimisation is the calibration of the e-motor and inverter pair, one of the most critical subsystems in electric vehicles. Calibration involves optimising myriad parameters and constraints to achieve optimal performance, such as effective use of energy from the battery, in as little time as possible.
The engineer is particularly interested in the rotor temperature because the motor’s magnetic field decreases as the magnet heats up, so a wide range of measurements is required. However, during the testing process the rotor temperature rises and the engineer must wait for it to cool and return to a stable state before they can take the next measurement, resulting in many lost hours.
Secondmind incorporates domain knowledge into its machine learning models in order to understand more precisely the physics of the e-motor. These models quickly learn and adapt to the e-motor’s unique personality, reducing the amount of data to collect for calibration. This further accelerates the calibration process and reduces the overall time required to generate accurate calibration maps that tell the e-motor and inverter how to function under certain conditions.
Q: Can you tell me a little more about the technology behind the Optimization Engine?
A: The technology fuel that powers the Optimisation Engine is Secondmind Active Learning, which intelligently automates the process of data acquisition, modelling, analysis and experimental design to achieve optimisation objectives faster. Traditional design of experiments (DoE) approaches are manual and linear, with engineers spending a lot of time upfront planning and acquiring more data than they need before running the experiment. This brute force approach more often than not results in a suboptimal outcome and the process needs to be repeated many times.
As an integral element of the Optimization Engine, Secondmind Active Learning significantly reduces the upfront preparation time and effort by giving the engineer a simple template to define the parameters, constraints and objective for a given optimisation session, and with a small sample of data, Active Learning algorithms intelligently design and run smaller experiments based on data from the targeted regions of interest deemed important enough to collect.
Knowledge of the problem increases with each iteration and, if left to run in a fully automated fashion, the optimisation objective is typically reached in half the time of existing DoE tools with results being as good or better. If engineers prefer a deeper level of engagement, they can participate by leveraging their domain knowledge in the process - knowledge not captured in the data, but potentially vital to achieving successful results.
Q: Can you tell me more about your System Design and Calibration products - why have you focused on these two areas for the automotive sector?
A: The broad design and calibration phases of vehicle development are the most complex, time-consuming, and costly and this is where we believed we could make the biggest initial impact. Our System Design and Calibration solutions offer capabilities that are unique to each development phase, application and engineering end user. At its core however, the Optimization Engine is designed to be system and application agnostic, offering flexibility in optimising system, subsystem and component-level designs, and in addition to the calibration of any number of vehicle control systems without the need for bespoke software development.
Secondmind for Calibration helps calibration engineers design high-value control strategies and produce high-precision calibration maps for complex systems like e-motors, internal combustion engines, and hybrid systems. A key capability of the Calibration solution is the intelligent automation of experiments to more quickly and easily generate calibration maps. Car makers like Mazda are using Secondmind for Calibration to halve calibration time using just 20% of the data they would otherwise need, and both modelling accuracy and less time on test benches has resulted in projections of significantly less prototype fabrication costs in future new vehicle programs.
Secondmind for System Design reduces design and simulation time, and error correction costs helping design engineers discover more design options and make better system configuration choices. By quickly identifying optimal design spaces, engineers can explore and innovate with better choices than they would have otherwise had due to data and time constraints.
Multiple engineers contributing to the design of complex vehicle systems are also empowered to experiment and make design trade-offs without having to worry about team or component-specific dependencies, resulting in efficient parallel planning that improves collaboration and ensures development schedules remain on track.