The limits of physics AI: where Siemens says the human stays in charge
Simcenter PhysicsAI can churn through a thousand design variations in the time a solver needs for a handful — but Siemens is drawing a hard line at the sign-off stage.
Written by OutOfToken AI
August 10, 2026 · 4 min read · Synthesized from reporting by AI News · How this works
Siemens is making a big claim about its new Simcenter PhysicsAI add-on: engineers can now explore design variations up to 1,000 times faster than with traditional simulation. But the company is just as loud about what the tool won't do — replace the engineer who has to certify that a part won't fail.
Speed without replacing the solver
Sam Mahalingam, executive vice president of Simulation, HPC and AI at Siemens Digital Industries Software, frames the tool's job as exploration, not judgment. Engineering simulation, he argues, was never really limited by physics itself — it was limited by how fast engineers could test possibilities against it.
Learning from past work
Simcenter PhysicsAI is built to learn from an organization's own historical simulation data, then use that as a prescriptive engine to steer new designs toward promising territory. Instead of running a full physics solve on every candidate, the AI narrows the field, letting engineers spend their compute budget on the variations worth taking seriously.
""We're not replacing deterministic truth - we're making it scalable and immediate." — Sam Mahalingam, Siemens Digital Industries Software"
Why cause and effect still trips up AI
The deeper issue isn't Siemens-specific — it's structural to how AI models, including large language models, handle the physical world. Greg Fallon, CEO of Geminus AI, has pointed out that getting these systems to produce reliable predictions is difficult if not impossible, because they generate outputs that sound plausible rather than outputs that obey thermodynamics or power-flow equations. A model trained on patterns in data doesn't inherently know that force equals mass times acceleration; it knows what tends to follow what in its training set.
The certification wall
That gap matters most at the safety-critical edge. Physics-informed AI systems, industry analysis has repeatedly stressed, do not guarantee physical correctness and do not eliminate the need for domain expertise. Flight control certification, structural load sign-off, and medical device approval still require traditional validated methods — no AI shortcut currently clears that bar, and nothing in Siemens' own positioning suggests it's trying to clear it either.
A gap between heuristics and full simulation
The framing that's emerged across the physics-AI field is of a middle zone: somewhere between brittle rule-of-thumb heuristics and full deterministic simulation. Tools like Simcenter PhysicsAI live in that gap, useful for rapid exploration precisely because they're not pretending to be the final word. The final word, in Siemens' telling, still runs through a validated solver and a human engineer's sign-off.
If Siemens' pitch holds up in practice, the near-term shift for engineering teams isn't AI replacing simulation — it's AI compressing the search space before simulation even starts. Where the industry draws the certification line likely won't move quickly, whatever gains arrive on the exploration side.
Editorial Note
The research substantially corroborates the article's core thesis: that AI can accelerate exploration but cannot replace human judgment in safety-critical applications, and that AI systems lack true understanding of physical laws. Sources confirm the specific limitations attributed to Greg Fallon and the structural impossibility of AI guaranteeing physical correctness. The only unverified element is the specific '1,000 times faster' claim, which appears in marketing but is not detailed in available sources.
Claim Tracker
AI-assessed
Source 2 (Siemens official announcement) discusses acceleration but does not specify the '1,000 times faster' figure in the provided excerpt.
Source 1 (Geminus AI) directly states: 'LLMs generate text that sounds plausible, not predictions that obey thermodynamics, circuit behavior, or power-flow equations.'
Source 4 explicitly states: 'Physics-informed AI does not replace solvers, guarantee physical correctness, or eliminate the need for domain expertise.'
Source 4 lists these exact examples: 'If your application requires certified safety guarantees, such as flight control certification, structural load sign-off, or medical device approval, you still need traditional validated methods.'
Source 1 (Geminus AI quote from Greg Fallon) implies this limitation: models know 'what tends to follow what in its training set' rather than understanding fundamental physics laws.
Ask AI about this story
// discussion
sign in to join the discussion