Insights · Industry Insights · Automotive

Industry Insights · Automotive · Article 01

AI in Automotive R&D and Engineering: What the Evidence Shows

AI is changing established engineering work and opening new research possibilities across vehicle manufacturers, suppliers and research institutions. The public evidence reveals different approaches, different levels of maturity and no single path for the industry.

Automotive research and development spans work as different as refining combustion engines and hybrid controls, developing batteries and electric motors, assessing aerodynamics and crash performance, designing lighter structures and writing vehicle software. Engineers also investigate functions that were once beyond the capabilities of a car. Each advance must eventually be made reliable, manufacturable and affordable. Much of this research takes place beyond the vehicle manufacturer, among component suppliers, specialist engineering companies and research institutions.

Artificial intelligence is beginning to change this landscape in different ways. In some activities it improves the efficiency of established engineering work. Elsewhere it influences which experiments specialists undertake or how they examine competing designs. In materials research and autonomous systems, it is expanding the possibilities that can be investigated or realised. The public evidence is uneven: some results have been measured in specific engineering activities, others are research demonstrations, and still others remain company ambitions. Those differences are essential to understanding what is changing.

These developments are emerging across many parts of the industry, without a single approach that all manufacturers appear to be following. Their significance depends not only on what AI can do, but on which engineering problem it changes, who is developing the capability, and what it would take to use it in a production vehicle.

AI-assisted calibration, simulation and vehicle testing

Engine calibration illustrates how AI can improve an established, measurement-intensive activity. An engine must perform across different loads, temperatures, speeds and operating conditions. Engineers spend considerable time deciding what to measure, conducting tests and adjusting control settings. Mazda's 2025 technical report describes AI-assisted methods that reduced development labour for the control-calibration activities studied by more than half. The approach was partly applied during development of the CX-90's 3.3-litre gasoline turbo engine. Mazda also reports work to apply similar methods to electric-motor development. Its collaboration with Secondmind shows how existing powertrain expertise can be combined with specialist machine-learning tools.

Methods such as active learning use earlier experimental results to identify which measurement is likely to be informative next. They can help engineers investigate uncertainty without following a fixed, exhaustive test plan. A comparable approach appears in Nissan's work with Monolith. The partners used historical data and predictive modelling to prioritise chassis bolt-joint tests for the new Leaf, reporting a 17% reduction in physical testing for that activity. In both cases, AI changes the choice and sequence of engineering work, rather than merely performing an unchanged calculation faster.

Emissions requirements create another reason to improve how calibration work is selected and assessed. Ashok Leyland and MathWorks have studied tree-based machine learning for calibrating emissions-monitoring conditions under India's BS-VI Stage 2 rules. Europe's Euro 7 regulation introduces requirements covering areas such as brake particles and battery durability. Such policies influence engineering priorities; the direct AI evidence here concerns particular calibration methods, not a demonstrated change in the adaptability of entire vehicle programmes.

Vehicle aerodynamics illustrates a different use of AI. Computational-fluid-dynamics simulations can take considerable time, particularly when designers need to examine variations of a body shape. In a 2025 study involving Volkswagen engineers, learned models estimated drag at least 600 times faster than the reference simulations, using 343 geometries derived from 32 baselines across five car projects. But they predicted the direction of drag changes correctly in only about 77% of comparisons and performed less reliably on subtle variations within a design family. The models may help screen more early concepts, but if their predictions determine which shapes receive closer attention, errors can also affect the direction of the engineering search. Finer comparisons still require higher-fidelity analysis.

Structural design provides a related example. Chery describes a platform that combines learned predictions with parametric modelling, simulation and optimisation; the company reports removing 13.8 kilograms from side- and pole-impact subsystems on one project. In the United States, General Motors has described reducing certain virtual engineering analyses from approximately 15 hours to about a minute. Both are company-reported results, and existing computer-aided engineering methods contribute alongside AI. The common effect is that more alternatives may be considered before engineers commit to detailed simulations or physical prototypes.

These approaches use different techniques. Some choose experiments adaptively; others learn approximations of computationally expensive physical models; others combine optimisation with established engineering software. Their contribution is often to reduce the cost of finding promising engineering alternatives. That can change what is worth investigating, although testing, tooling, system integration and regulatory evidence retain their own demands. Faster analysis of one task does not, by itself, establish a shorter vehicle-development programme.

How AI may change automotive engineering work

Easier exploration does not remove the need for engineering expertise, but it can change where that expertise is applied. Specialists still define the problem, decide which variables and constraints matter, recognise when predictions extend beyond the available evidence and determine what must be tested before a design can be accepted. These responsibilities are apparent in emerging workflows, although there is little public evidence measuring how much time engineers have actually been able to redirect from routine work.

At component supplier Denso, researchers have automated parts of materials experimentation and are investigating AI methods that can propose further experimental conditions based on results already obtained. They describe an ambition to spend less time on repetitive procedures and more on deciding which materials are worth pursuing and whether a finding has scientific or practical value. This remains an intended shift in the researchers' work, not a measured change in time allocation. It illustrates how better experimental tools may make the choice of research questions more consequential.

Vehicle software presents another kind of engineering work. Modern cars contain interconnected control systems, embedded code and software that must remain maintainable as vehicle functions evolve. GM's 2026 research describes using language models for contained tasks such as code refactoring, translating legacy scripts and preparing test artefacts. Its researchers found that well-specified, reviewable tasks produced more dependable results than open-ended requests. A ZF technical paper sets out a related approach to chassis-component inspections: a computer-vision model classifies observations, while specialists review uncertain cases and record corrections. The methods differ, but both place the model within an engineering process in which its output can be examined and improved.

New possibilities in battery research, vehicle design and autonomous driving

Other applications are changing what can practically be investigated. A 2020 Nature study combined early battery-life prediction and Bayesian optimisation to find promising fast-charging protocols among 224 candidates in 16 days. Testing every candidate exhaustively was estimated to require more than 500 days. This was not the invention of a new automotive battery. It was a demonstration that researchers could use partial results to guide experiments through a search that would otherwise have consumed far more laboratory time.

The scale of a materials search can change as well. In a study involving Microsoft and Pacific Northwest National Laboratory, researchers combined machine learning, physics-based models and cloud computing to screen more than 32 million candidate structures for solid-state battery electrolytes. They narrowed the search to 18 promising compositions and characterised a selected electrolyte in the laboratory. Finding candidates computationally is still far removed from qualifying a battery material for vehicle use; durability, safety and production methods matter. A published scientific critique of another AI-generated materials database also questioned how many predicted structures were genuinely novel, credible and useful. Google DeepMind has defended its discovery claims. The disagreement illustrates why the number of proposed compounds is not itself a measure of successful discovery.

Physical design introduces another possibility. Bugatti's Tourbillon includes an AI-developed, 3D-printed hollow rear-airfoil arm alongside other organically designed structural parts. Such components show how computational methods and additive manufacturing can together produce geometries difficult to make through conventional processes. The result cannot be attributed to AI alone, and the economics of a limited-production hypercar differ from those of a mass-market vehicle. It nonetheless shows how advances in modelling, materials and manufacturing can reinforce one another.

Autonomous driving is different again, because AI enables a vehicle function rather than simply improving the design process. Waymo reports 271.3 million rider-only miles through June 2026 in the cities where it operates. The company-reported figure demonstrates sustained driverless operation within defined service areas, not universal autonomy. That capability rests on specialised perception and control software, sensors, simulation, vehicle integration and extensive testing. It establishes progress in a particular technical domain, without providing a basis for ranking companies across automotive R&D as a whole.

These examples have reached different stages. AI-guided experiments can be evaluated against measured laboratory outcomes; proposed materials must still progress through development and qualification; novel structures depend on manufacturing economics; and driverless vehicles have reached commercial operation in some environments. Their significance for vehicle manufacturers depends on where each capability stands between research, practical engineering and deployment.

Automotive R&D across suppliers and research institutions

The origins of these developments are as varied as the applications. Mazda and Nissan have worked with specialist software providers. Denso and ZF undertake component and materials research of their own. Battery and tyre suppliers possess testing facilities, physical data and design expertise that manufacturers do not necessarily maintain internally. The distinction between in-house R&D and outsourced development is therefore too simple to describe the range of relationships involved.

Tyres offer a particularly concrete example. Bridgestone brought a driver-in-the-loop simulator into operation in 2026, combining high-fidelity models, historical data, AI methods and feedback from professional drivers. It can examine a broader selection of tyre designs before final physical validation and supports closer development work with automakers. Sumitomo Rubber and Fujitsu reported a proof of concept reducing one tyre-deformation analysis from about 45 minutes to five, with use as a development tool targeted for 2027. These are different levels of maturity, but both suggest how suppliers' research capabilities may influence the options available to vehicle designers. Whether those wider choices are already producing better vehicle-level outcomes remains to be demonstrated.

Applied-research institutions also contribute. Germany's Fraunhofer SCAI, working with engineering software firm SCALE and with Audi, Volkswagen and Porsche as associated partners, is investigating how machine learning can connect engineers' descriptions of design changes with virtual crash-test data. Other Fraunhofer programmes examine the evidence needed for automated-driving safety and new materials and production methods. These are ongoing research projects rather than measured vehicle-programme improvements. They show how methods and shared research infrastructure can develop beyond one manufacturer before being adapted to particular products.

These relationships include licensing specialist software, co-developing engineering methods and participating in shared applied research. They also distribute responsibility for data, problem definition, testing and integration differently. The resulting capability depends on what each participant contributes, not simply on which organisation owns an AI model.

The public record consequently does not support a single ranking of manufacturers by AI innovation in R&D. Some organisations have demonstrated strengths in particular fields, while others investigate narrower problems important to their products. Public announcements reveal only part of their work. What the evidence supports is a landscape of distributed experimentation and collaboration, without implying that every company is advancing equally or that less visible activity means a company is falling behind.

When robotics and manufacturing influence vehicle design

Automakers are using established industrial automation while testing more adaptable forms of physical AI. Xiaomi's automotive factory uses conventional automated production equipment, while its 2026 investor presentation describes humanoid-robot trials on specific tasks. Google DeepMind reported in November 2026 that Apptronik robots achieved 96% task success in automotive parts-kitting trials at Mercedes-Benz plants. Those results do not establish sustained economic performance across general assembly work. They do, however, show why present technical limitations are not necessarily an adequate forecast for a multi-year product or factory investment.

The more direct connection to vehicle R&D is that production capabilities can influence component design. Flexible wiring harnesses remain difficult to manipulate reliably with robots, but the response need not be limited to making robotic hands more dexterous. In Audi's Next2OEM project, partners developed harness production, automation-compatible connectors and installation processes together. It is a design-for-automation example, not evidence that AI itself delivered the improvement. As robotics develops, the possibilities for product and process design may change together, even when the timing and economic value of particular systems remain uncertain.

Different priorities, and the need to keep choices informed

The evidence now available supports different conclusions at different levels. Selected calibration, simulation and testing activities show measured or company-reported improvements against specific baselines. Changes in how specialists allocate their time remain less well documented. Research into materials and computational design has broadened the range of candidates that can be explored, while driverless operation has moved into commercial use within defined areas. The differences in maturity matter when assessing what any of these developments can contribute to a vehicle programme.

Nor is there a single course that every manufacturer needs to follow. A company with substantial hybrid-engineering commitments may have compelling reasons to focus on calibration and energy management; another may concentrate on batteries, software architecture or autonomous functions. These differences reflect engineering priorities and commercial choices. Public evidence provides no sound basis for assuming that the companies pursuing the greatest number of visible AI initiatives are making the best overall decisions.

The cases also suggest a consequence beyond faster individual tasks. Vehicle platforms, supplier arrangements, control strategies and validation programmes involve commitments made against the technical and economic possibilities understood at the time. Mazda's reported reduction in calibration labour, the battery experiments and the aerodynamic screening study do not show that earlier decisions were mistaken. They do show that the cost of investigating some alternatives can fall substantially. As engineers use new methods to select which designs and hypotheses deserve further work, AI can influence the direction of exploration as well as its speed.

Those choices remain constrained by physical testing, production costs, safety, regulatory evidence and each manufacturer's priorities. A more affordable experiment may reinforce an existing programme, reveal a viable alternative or show that an attractive idea is still impractical. Meanwhile, capabilities developed by suppliers and research institutions, changes in production methods and shifts in competitors' offerings can alter the options available without originating in a manufacturer's own AI programme. The longer-term significance is that the assumptions behind multi-year engineering commitments may change while the commitments themselves remain. The number of announced AI projects says little about which organisations will recognise and respond to those changes.