Siemens Says Physics AI Can Run 1,000x Faster, But Not for Final Safety Approval

Siemens says its PhysicsAI technology can dramatically accelerate engineering design exploration, but the company is drawing a clear line around where artificial intelligence should not be trusted on its own: safety-critical validation.

Physics-based AI can evaluate thousands of design variations in roughly the time traditional simulation tools might process only a handful, with Siemens claiming performance improvements of up to 1,000 times faster.

But speed does not mean engineers can skip traditional physics validation.

Sam Mahalingam, who leads the business behind the technology at Siemens Digital Industries Software, made that distinction clear during Realize LIVE Asia-Pacific in Bengaluru.

Asked whether the technology is suitable on its own for safety-critical applications, Mahalingam gave a direct answer.

“No, it is not.”

The statement highlights one of the most important limitations surrounding the growing use of AI in engineering. Siemens is positioning PhysicsAI not as a replacement for traditional simulation, but as a tool for rapidly narrowing down design possibilities before engineers perform more rigorous validation.

What Siemens PhysicsAI actually does

The technology, called Simcenter PhysicsAI, uses geometric deep learning to predict how a design is likely to behave.

Unlike traditional physics solvers that calculate a problem from the ground up for every new design, PhysicsAI learns patterns from previous simulation results.

Once trained, the AI can estimate what will happen to a new design within seconds.

This makes it particularly useful during the early stages of engineering, when teams may want to evaluate hundreds or thousands of possible designs.

Instead of performing computationally expensive simulations on every variation, engineers can use the AI model to identify the most promising candidates first.

Those candidates can then undergo more detailed physics-based simulations.

AI predictions still need validation

Accuracy remains one of the biggest questions surrounding AI-driven engineering.

According to Mahalingam, traditional simulation software earned engineers' trust over decades by being repeatedly compared against physical testing.

PhysicsAI is now being benchmarked against those established physics-based simulations.

When enough training data is available, Siemens says predictions can come within roughly 1% to 3% of traditional solver results in certain applications.

That level of accuracy may be useful for identifying promising designs.

It is not necessarily sufficient for certifying components where failure could threaten human life.

Mahalingam described PhysicsAI as an exploration layer rather than the final authority.

Engineers can use the AI model to rapidly test large numbers of design variations, narrow the results down to perhaps two or three strong candidates, and then run those designs through conventional simulation tools.

Only after completing the more detailed physics-based validation would a design proceed toward manufacturing.

Continental airbag project shows the boundary

Siemens has highlighted applications of PhysicsAI involving companies including Continental.

One example involved airbag design exploration.

But Mahalingam stressed that even in cases like this, AI is used during the early exploration phase.

PhysicsAI is not intended to become the only validation system before a safety-related design is approved for production.

The distinction matters because applications involving automotive structures, aerospace components and industrial machinery can carry serious consequences if simulation predictions are wrong.

Physics AI still depends on good training data

Another limitation sits behind the headline speed improvement.

PhysicsAI models need data.

In several Siemens examples, including work involving Magna and Continental, that training information included synthetic data produced by traditional simulation software.

This creates an important dependency.

If an AI model learns from simulation output, its reliability ultimately depends on the quality of the simulations used to train it.

Mahalingam explained that customers without enough existing engineering data can first generate synthetic training data using Siemens tools.

For example, engineers can explore multiple designs through Simcenter HEEDS and generate simulation results using Simsolid.

Those results can then become the training dataset for a PhysicsAI model.

Once trained, the AI can predict results for similar designs much more quickly.

The approach creates a workflow where conventional simulation produces the knowledge and AI later accelerates how that knowledge is applied.

Guardrails prevent predictions outside training range

Siemens has also developed safeguards intended to stop PhysicsAI from confidently predicting results for designs that are too different from anything in its training data.

A model trained on variations of one type of component may not reliably understand a radically different geometry.

Instead of generating an unreliable prediction, the system is designed to warn engineers that the new design falls outside the model's training range.

Mahalingam said the system can effectively tell users that a shape is too different from what the AI has previously learned.

This is especially important in engineering AI because incorrect confidence can be more dangerous than simply refusing to provide an answer.

AI as an engineering accelerator, not the final judge

The Siemens approach presents a more restrained view of artificial intelligence than the autonomy-focused messaging increasingly common across the technology industry.

PhysicsAI is not being positioned as a system that eliminates engineers, physics solvers or validation procedures.

Instead, its strongest role may be much narrower and more practical.

AI performs the rapid first pass.

Traditional simulation performs the detailed verification.

Human engineers still decide whether the design is ready.

For engineering teams, the real advantage may not simply be the headline 1,000x speed improvement.

It is the ability to explore a much larger design space without removing the established validation processes required when the final answer has to be correct.

In safety-critical engineering, Siemens' message is clear: AI can help engineers find the answer faster, but it should not be the system that gives the final approval.