Siemens' physics AI is 1,000x faster for design, but it can't certify a single safety-critical part

Siemens has a number that grabs you by the lapels: Simcenter PhysicsAI can predict design performance up to 1,000 times faster than a traditional solver. It’s a geometric deep-learning system that doesn’t compute physics from scratch but learns from historical simulation data to spit out an estimate in seconds. For an engineer staring down thousands of design variations, that kind of speed is genuinely useful. But what the software cannot do is the one thing that actually matters for a brake pedal or an airbag inflator: sign off on a final, safety-critical design. Sam Mahalingam, who leads the business building this technology at Siemens Digital Industries Software, doesn’t dance around the limitation. “Is this good for safety-critical applications?” he asked on the sidelines of Realize LIVE Asia-Pacific in Bengaluru. “No, it is not.”

That blunt admission cuts against two years of an industry-wide AI hype cycle that has relentlessly promised near-total autonomy. The real value here isn’t the 1,000x speed—it’s knowing precisely where that speed stops being trustworthy. Mahalingam frames the surrogate model not as a replacement for validation but as a filter you place in front of it. Engineers can explore a vast design space, zero in on two or three promising candidates, and only then switch to a full physics-based simulation for the detailed work. A Continental airbag case Siemens has showcased sits squarely inside that boundary: it was for initial exploration, not a manufacturing recommendation. The accuracy is tight—typically within 1% to 3% of a physics-based solver, according to the company’s case studies—but tight isn’t certified.

There’s a second limit the eye-popping speed numbers tend to obscure. Several of Siemens’ headline results, including work with Magna and Continental, rely on AI trained on synthetic data generated by Siemens’ own solvers. The surrogate is only ever as good as the simulation that taught it. Mahalingam acknowledged the circularity directly: where a customer had no data, they first ran broad design explorations using Simsolid and HEEDS, then fed that simulation output back into training the physics AI model. It’s a constraint that could turn into a trap, but Siemens has built a guardrail to stop the model from hallucinating on unfamiliar ground. If you ask it to predict a shape radically different from its training data, the system is designed to refuse. “We have put in guardrails where it comes back and says, hey, I cannot predict this,” Mahalingam said. “So the engineer cannot shoot themselves in their own legs.”

There’s a quiet strategic bet in this honesty. Every simulation vendor is now bolting AI onto its portfolio, and the credibility risk is that buyers simply stop believing any performance claims. By openly marking the edge of the technology—safe for exploration, useless beyond its training envelope, and never a substitute for final sign-off—Siemens is gambling that engineers, a famously skeptical bunch, will trust a tool more when it tells them what it cannot do. It’s a message that lands differently from a company rooted in simulation rather than pure-play AI. While chip-design and enterprise-AI vendors have spent the hype cycle promising full autonomy, a firm whose DNA is validating physical things against reality is drawing a line in the sand. For an industry that has often treated AI’s limitations as a temporary bug rather than a permanent feature, that boundary-setting might be the most useful product of all.