At the CTI Symposium 2025 in Novi, Luca Zampieri, Engineering Director US at Neural Concept, gave a presentation on “Engineering Intelligence Driving the Future of Powertrain Development”. We spoke with him about how engineers and the products they design can benefit from using AI engineering software.
Luca, what can Neural Concept do for the automotive industry? What is your product?
At Neural Concept, we develop AI software that accelerates the engineering process. Our idea is to connect and speed up the internal workflow, all the way from design to manufacturing. Our platform allows you to stay connected to today’s workflows – moving back and forth with your current tool stack – while also leveraging new technology and big data: state-of-the-art algorithms and computing infrastructure, all in a single environment. That’s what we offer the automotive industry: faster, better design and engineering processes.
So far, AI has remained a rather abstract topic for many engineers. To begin with, what are the milestones in AI history that led to today’s capabilities in powertrain development?
It goes back to the first AI revolution, which was about statistics and optimization – Gaussian processes, traditional machine learning. But today we focus more on what came after the advent of deep learning: convolutional networks – which scan and compress visual data into patterns a model can learn from – and transformers, which let a model weigh the relevance of different parts of an input against each other. A key milestone was the creation of surrogates predicting physics as well as displaying generative capabilities for 3D shapes. These technologies matter greatly for powertrain development. And now we also use large language models, also powered by transformers, for an agentic control and consumption of the platform.
How does your platform learn the patterns for engineering requirements – through integration with common engineering tools?
Exactly. Our platform includes such tools and interfaces. But the main idea is to access the whole workflow, from design and conception through to manufacturing, removing the bottlenecks along the way. One such bottleneck is generative modeling, or surrogacy – the idea of replacing a physical simulation with a learned model that approximates it much faster. For some applications, we provide pre-trained models that come with their own data. For most others, we provide an agentic framework to create your own surrogates as well as automations to collect and label data.
What does that mean – this is a huge amount of knowledge you need. How do you train this as a user, as an engineer?
For a physical surrogate, you don’t need much knowledge as a user, because everything is already set up. What you need is data and your engineering knowledge. We work with small amounts of data because it’s essential to focus only on what matters to you. By default, engineers tend to replicate the full simulation output simply because that’s what they are used to seeing. But most of the time, what matters is only part of that output – and capturing just that part is often much easier, and requires far less data, than reproducing the whole simulation.
Isn’t there a risk that engineers lose expertise and creativity by being assisted by an AI engineering assistant?
I don’t think it’s one versus the other. The question is how the product designer can leverage simulation to better judge when a design is a good one. The engineer still needs all of the expertise. Andrej Karpathy, founding member of OpenAI and former Tesla Autopilot AI director, once said that you can outsource your thinking to AI models, but not your understanding. The same applies here:
these tools can help the expert gain even more expertise, because quick feedback lets you better understand the physical phenomena behind your design. Today, engineers spend much of their time on repetitive, low-skilled tasks, leaving less time to focus on the depth of the problem. Our AI tools allow them to focus fully on the problem they’re actually trying to solve.
Let’s assume I’m a designer working with your tool. Do I use natural language to communicate with the agent?
That’s one option. A lot of knowledge lives in natural language, so you can interact a great deal through text. But you can also interact naturally – clicking, moving, manipulating the interface – so it’s not only text. Just as a picture is worth a thousand words, I’d say an interactive interface is worth a thousand pictures. If I needed to explain something to you, I’d probably sketch it. We encourage sketching because it conveys design intent far better. The goal is for you to convey your design intent to the machine, and natural language is just one way of doing that.
What benefits does your AI software offer in terms of time savings, accuracy, performance, or quality?
The accuracy of the AI matters less than you’d expect. What matters is how fast you can design something, and how much better a design you can find. We often say AI is a thousand times faster than simulation – but that figure alone is meaningless. What matters is how much faster you reach the final result, and that depends heavily on the workflow and the product. It’s usually a trade-off between speed and performance. We’ve seen products designed in three days instead of three months. We’ve seen performance double. And we’ve sometimes seen performance improve by just one percent – which can still be extremely impactful. In powertrain development specifically, we typically target performance gains of around 10 to 15 percent, and time savings of about a quarter of the original development time.
You also spoke about combining parametric and non-parametric development methods. What would be an example in the development of an electric powertrain?
It’s both parametric and non-parametric – think of parametric as a design defined by a fixed set of adjustable variables, like the dimensions of a known shape, and non-parametric as a free-form shape with no such constraints. It’s also a case of combining new technology with old. Manufacturing teams use CAD, while newer techniques let you explore a much richer, more organic design space. But hand manufacturing teams something too organic, and they’ll push back. So the idea is to use traditional parametric methods with CAD where that makes sense, and non-parametric freedom where it’s needed – and to keep both connected to the tools used today, adopting new ones only where they add value. If we think about an e-motor, for example, the aim is to increase performance while reducing noise and vibration. In this case, you might want the magnet shape to remain parametric, but the air pockets to be non-parametric and organic.
You also mentioned so-called world models. What’s their relevance to automotive powertrain development?
A world model is one trained on enough data that the usual distinction between “in-distribution” and “out-of-distribution” data – cases the model has and hasn’t effectively seen before – no longer applies; the whole domain you want to explore falls within what it can predict. That doesn’t mean it’s a fully generic physical model. Building a robust world model requires considerably more data, and its accuracy is typically lower than that of a specialized model built close to the specific design you’re examining. For heat exchangers, for instance, world models generalize well across many configurations. For e-motors, though, it makes more sense to combine a world model with a specialized model close to the design being explored to get better accuracy.
Generative AI, as many of us use it in daily life, tends to produce answers at all costs, even when it combines facts incorrectly. How do you avoid such reasoning errors in engineering?
It’s essential to validate whatever the AI model produces. In engineering, optimization happens under constraints – and if you’re good at optimizing, you’ll very likely produce a result that violates some rule you hadn’t specified, whether a manufacturing limit or a physical consideration from another department. That’s why validation is so important. When using AI for design and optimization, it doesn’t really matter if the model hallucinates, because in the end, you’ll have a set of optimal results to cross-reference and verify.
To put it bluntly, could one say that the role of developers is thus shifting more toward that of a highly qualified supervisor?
I think engineers already work this way; they’ll simply do more of it. Designers from other departments already bring designs that engineers need to check and adjust. Here, that process is streamlined, and validation itself can be automated – not as fast as AI, but AI helps automate it, so you don’t need to spend time on it manually. You can automatically trigger higher-fidelity tests instead.
How does working with your tools change the prospects of engineers?
It changes them considerably. Skills like tool use and repetitive work become less important. But creativity and a deep understanding of the product you’re working on – already the hallmark of a good designer or engineer – will matter even more tomorrow. The good designers and engineers will become even better; the weaker ones risk becoming obsolete.
Interview: Gernot Goppelt
