Gasgoo Munich- "Why does AI hallucinate? Because it bears no accountability—it can spout nonsense without consequence."
At the recent 14th Automobile & Environment Innovation Forum, those words from Zhao Fuquan laid bare a stark reality facing AI in automotive engineering. Zhao is a lifetime honorary president of FISITA, a professor at Tsinghua University, and dean of its Automotive Industry and Technology Strategy Research Institute.
AI getting things wrong is nothing new. But when it steps out of a chat window and into the vehicle development process, the cost of an error is no longer just an inaccurate answer.
Consider body lightweighting: AI prediction models can help engineers rapidly assess the performance of various structural designs. Yet even a flawless prediction doesn't guarantee the design will pass crash validation—let alone clear it for production.
AI can crunch the numbers, but it cannot shoulder the burden of engineering judgment.
As AI weaves into more industrial software, automakers are gaining new tools to accelerate development. Yet actually wielding those tools effectively is far more complex than simply deploying a model.
As Development Cycles Shrink, Where Do Better Designs Come From?
The hardest part of automotive R&D isn't finding a workable design—it's finding the better one among a sea of options.
Take body lightweighting again. Thinning a steel panel might cut weight, but it could also compromise crash safety. Add reinforcing structures, and you might cancel out the weight savings you just achieved.
Engineers must constantly balance weight, safety, cost, and manufacturability. Improving a single metric doesn't mean the overall design is actually better.
Liu Yegang, director of the lightweight technology committee at Chery, told Gasgoo that EV lightweighting has evolved from a single-minded quest to shed pounds into a multi-objective, multi-constraint optimization problem.
R&D teams are no longer hunting for the optimum in a single metric. They need designs that satisfy multiple engineering demands at once.
The catch? The number of possible design combinations is staggering.
Image Source: AI
Computer-aided engineering (CAE) simulation is already ubiquitous in auto development. Engineers can analyze structural strength and crash performance before building physical prototypes, then tweak designs based on the results.
Optimization algorithms aren't new, either. By defining design variables, objective functions, and constraints, software can sweep through different parameter combinations for engineers.
Even so, having these tools doesn't mean development efficiency is no longer bottlenecked.
In traditional workflows, engineers iteratively modify structures, run simulations, and adjust again. Under complex crash scenarios, high-fidelity computing can devour immense time and processing power.
As design variables multiply and performance requirements grow more complex, the cost of finding a superior design climbs right alongside them.
Making matters trickier, the industry is giving engineers less and less time to work with.
Fierce competition in China has steadily compressed new-model development cycles. Automakers want faster launches, yet safety validation, reliability testing, and manufacturing prep don't magically get easier just because a deadline moves up.
At a recent Gasgoo AI forum, Wang Xianbin—a partner and VP of research at Gasgoo—spoke bluntly about the supply-chain strain from accelerated R&D: "If you've cut every corner and still can't hit the milestone, the only thing left to cut is validation."
Behind that remark lies a widening time paradox in vehicle development.
Designs demand iterative refinement, and essential validation can't simply be slashed. If efficiency gains come mainly from skimping on validation, risks just get deferred to late-stage development—or even post-delivery.
The real opportunity, then, is cutting dead-end attempts during the design phase—focusing finite computing and validation resources on the most promising concepts.
That is exactly the opening AI is finding in structural design.
How Much Time Can AI Actually Save Engineers?
If high-fidelity simulations take too long, can AI predictions at least filter out the unnecessary compute cycles?
That’s the logic pulling AI into structural optimization.
Chery and Jiangsu University jointly developed the iDO lightweight intelligent design platform. It marries AI prediction models with parametric modeling, CAE simulation, and proprietary optimization algorithms to accelerate complex structural design.
Liu noted that the platform is now on its 2.0 iteration and has been used in early-stage development for models and projects including the Sterra ES, Fulwin A9/T11, C21A, and E02.
Parametric modeling, simulation, and optimization algorithms are already industry staples. AI’s main contribution is rapid prediction based on existing data, letting engineers evaluate certain designs at a fraction of the computing cost.
Still, prediction is not validation.
A model can estimate the performance of various parameter combinations, but prediction alone cannot confirm that a design meets safety, manufacturing, and reliability standards.
AI, then, isn't designing cars on its own. It operates as a node within the existing engineering toolkit, helping to hunt for better solutions.
Data from Liu illustrates what this’s tool suite delivers in practice.
On the E02 project, the team trimmed 13.8 kg from side- and pole-impact subsystems over four optimization iterations.
- At the platform level, iDO-driven optimization has cut body weight by up to 21 kg—provided all performance metrics hit their targets.
Elsewhere, the optimization work has opened up 100–150 mm of additional crash space and boosted bumper beam energy absorption by 32%.
These figures map to different components and metrics—they can't simply be tallied, nor should they be read as simultaneous gains on a single vehicle. Together, however, they make one thing clear: lightweighting isn't just stripping out material. It's finding a smarter balance between mass and structural performance while hitting every engineering target.
Beyond the optimization outcomes, efficiency is the other key measure of these tools' worth.
Liu reported that iDO slashed the component modeling cycle by 75% and lifted optimization efficiency by over 40% compared to commercial optimization software.
AI prediction models don't deserve all the credit. Parametric modeling cut out repetitive manual steps, proprietary algorithms sped up parameter searches, and better tool orchestration played its part, too.
In other words, AI's impact hinges not just on the prediction model, but on how seamlessly it plugs into existing R&D workflows.
Overseas automakers are chasing similar efficiency gains.
Image Source: AI
Spanish automaker SEAT, working with engineering software firm Synera, once automated an engineering workflow and slashed the processing time for a body-in-white optimization from 42 hours to just one.
The partners later layered AI agents onto those automated workflows. According to Synera, a 2026 demo for a manufacturing feasibility assessment saw processing time shrink from roughly four weeks to 30 minutes.
Those two datasets reflect different tasks—they aren't directly comparable, and they certainly don't equate to a slashed whole-vehicle development cycle.
Even so, the case proves that R&D efficiency gains don't always start with a more powerful AI model. Overhauling the engineering workflow itself can deliver dramatic results.
Chery is betting on AI predictions paired with optimization algorithms to explore more structural designs, faster. SEAT is leaning on workflow automation and AI agents to eliminate redundant tool operations and the friction of task handoffs.
The paths diverge, but they converge on one truth: AI only generates real value when it is embedded directly into engineering workflows.
As these AI tools become accessible to more automakers, a new question emerges: Why can similar software still yield such different R&D outcomes?
Same Algorithms, Different Designs — Why?
Commercial industrial software is racing to integrate AI.
In March 2026, Synopsys released Ansys 2026 R1. Its GeomAI module assists in exploring geometric concepts, while SimAI trains prediction models on simulation data to support design studies and performance optimization.
That gives automakers more options. Some AI capabilities no longer require in-house development; they can be tapped through off-the-shelf industrial software.
But acquiring the tools is just the start.
Even within body lightweighting, different models use different material systems, structural layouts, crash requirements, and manufacturing processes. Even with identical software, R&D teams must redefine which design variables can flex and which performance metrics are non-negotiable.
Set the design goals poorly, or overlook a critical manufacturing constraint, and the algorithm might find a mathematically superior solution that is utterly unbuildable in reality.
Before AI can truly enter automotive R&D, the first hurdle is translating messy engineering requirements into well-defined optimization problems.
That translation is impossible without an automaker's own institutional knowledge and engineering data.
Liu argues that as AI enters the R&D space, industrial software competition will hinge on more than algorithms. The winners will be those that best fuse engineering data, domain expertise, and real-world application scenarios.
Knowing which structural tweaks improve crash performance, or which material pairings suit a given process—these are lessons learned over time. If that knowledge stays scattered across projects and teams, it can't fully power downstream AI optimization.
Zhao's previously proposed "1+N+X" model framework addresses this split between general capability and enterprise application. Here, "1" is the foundational large model, "N" covers domain-specific models, and "X" represents company-specific models. He argues that manufacturers needn't all build foundational models; they should focus on training proprietary models tied to their own operations.
Yet owning the data and the model doesn't mean the real-world problems are solved.
AI prediction models rely on historical data. When a new project introduces different materials, structural forms, or load cases, the old model's accuracy can degrade.
Liu stresses that AI predictions must stay bounded by model accuracy and applicability. In critical scenarios like complex crashes, high-fidelity CAE simulation, engineering judgment, and physical testing remain irreplaceable.
R&D teams, then, need to know not just what a model can do, but exactly where it might break down.
By cross-referencing predictions against simulation and test results, engineers can map a model's limits and adjust their approach accordingly. Validated experience and workflows can then be reused in future projects.
That reuse has limits, though. Throw new materials, structures, or load cases at the model, and it must be re-evaluated from scratch.
Which means that even with the same AI tools, different automakers must still lean on their own engineering experience to decide which results to trust—and which to re-validate.
AI is speeding up the search for designs, but it hasn't made engineering judgment optional.
That judgment springs from engineers' hard-won expertise—and from the data, design processes, and validation systems the company has built over years.
As algorithms become commoditized, what may prove hardest to copy is an automaker's ability to turn AI's raw compute into a reliable car design.
