Dominic Jainy brings a sophisticated perspective to the industrial software sector, where the marriage of artificial intelligence and traditional physics is reshaping how we build the physical world. With a background deeply rooted in machine learning and high-stakes IT infrastructure, he understands the delicate balance between the raw speed of neural networks and the absolute precision required in heavy engineering. In this discussion, we explore the rise of surrogate models in simulation, the critical boundaries of AI in safety-rated environments, and why the industry is beginning to value radical honesty about technology’s limits as a primary competitive advantage. The conversation traverses the technical landscape of geometric deep learning, the use of synthetic data in training complex models, and the enduring necessity of human oversight in a world where design variations can now be explored 1,000 times faster than ever before.
How does the introduction of physics-based AI fundamentally alter the traditional engineering workflow when exploring thousands of design variations?
The shift we are seeing is nothing short of a paradigm change in how we conceptualize the “trial and error” phase of engineering. Traditional solvers are incredibly precise, but they are also notoriously slow, often taking hours or days to chew through a handful of complex simulations. By utilizing geometric deep-learning software like Simcenter PhysicsAI, we can now produce predictions up to 1,000 times faster because we aren’t computing the physics from scratch every single time. Instead, a surrogate model learns from historical data to provide an estimate in seconds, allowing an engineer to feel the immediate impact of a design change. It transforms the workflow from a series of long, isolated wait times into a fluid, conversational exploration of what is possible.
Why is it so critical for vendors and engineers to establish firm boundaries for AI, particularly when it comes to safety-critical components?
When you are dealing with life-or-death components like a Continental airbag or a jet engine turbine, “close enough” is never actually enough. Even though Siemens cites a very impressive 1% to 3% variation compared to physics-based solvers, that tiny margin represents a massive risk in a safety-critical application. The value of this technology is not in replacing the final validation but in acting as a high-speed filter that sits right in front of it. We must resist the industry urge to claim AI can do everything and instead use these surrogates to zero in on the best two or three designs. Only after those finalists are cleared by a traditional, rigorous physics-based check should they ever move toward a manufacturing line.
What are the practical implications and risks of training these physics AI models on synthetic data rather than real-world measurements?
Training on synthetic data, such as simulation output from tools like Simsolid or HEEDS, creates a specialized ecosystem that is incredibly efficient but inherently bounded. In cases like the work done with Magna, the AI is essentially learning from the “digital twin” of a process, which means the surrogate is only ever as good as the simulation beneath it. This circularity means the AI isn’t discovering new laws of physics; it is learning to replicate the behavior of existing solvers at a fraction of the temporal cost. It is a powerful way to generate data when none exists, but we have to be mindful that the AI’s ceiling is strictly defined by the quality and accuracy of the original solver that taught it.
How do the built-in guardrails in modern simulation software prevent the “hallucinations” that often plague other forms of generative AI?
The most dangerous thing an engineering tool can do is provide a confident answer to a question it doesn’t understand. To prevent this, specific guardrails are integrated to stop the model from predicting outcomes on geometric ground it has never encountered during training. If an engineer attempts to test a shape that is radically different from the training set, the system is designed to trigger a hard stop and admit it cannot predict the result. This transparency is vital because it ensures the engineer cannot, as the saying goes, “shoot themselves in their own legs” by trusting a baseless prediction. It turns the AI into a partner that knows its own limits, which is far more useful than a tool that guesses.
By being so candid about what physics AI cannot do, is the industry actually increasing the long-term trust of the engineering community?
Absolutely, because engineers are inherently skeptical of “black box” solutions and marketing hype that promises total autonomy. When a company like Siemens stands up and explicitly says their AI is not for final safety sign-offs, they are separating themselves from the noise of the general AI hype cycle. They are betting that an expert will trust a tool more when it clearly defines its own “envelope” of operation—where it is safe for exploration and where it must hand the pen back to a human. This honesty creates a more durable relationship with the user because it positions AI as a high-speed assistant rather than a flawed replacement for human judgment.
What is your forecast for the role of the human engineer as these 1,000x speedups become the industry standard?
I believe we are entering an era where the human engineer transitions from being a “calculator” to a “curator” of complex systems. As we automate the grunt work of iterating through thousands of variations, the human’s role in defining the right questions and interpreting the final safety risks becomes more important, not less. We will see a future where the AI provides the breadth of search—the 1,000x acceleration—but the human remains the final authority on depth and integrity. My forecast is that the most successful engineering firms will be those who use AI to widen their creative search while keeping their physics-based solvers as the ultimate, uncompromising gatekeepers of the physical world.