Sam Mahalingam on the Speed of Physics AI and the Need for Physics-Based Validation

Physics AI can explore vast design spaces at remarkable speed, but its results still need to be validated by physics-based solvers and engineers before they can be trusted. Siemens is bringing design and simulation together with open ontologies and an engineering stack in which AI is embedded directly into the workflow. In industrial AI, competitive advantage will come not from speed alone, but from the ability to ground decisions in real-time data, physical evidence and what Siemens calls “Engineering Truth.”

By Sang Min Han _ han@autoelectronics.co.kr
한글로보기

“Yesterday’s Data Is Dangerous”
Daimler Truck Cuts Ties With the Past to Connect the Future




Some in the industry argue that design and simulation must first be unified into a single workflow before AI can deliver meaningful value. Does Physics AI depend on such an integrated workflow?
Mahalingam    Design and simulation need to be unified within a single workflow for AI to add value. I believe an integrated working environment is necessary for Physics AI to deliver real value. However, that does not necessarily mean integration has to come first.
The first step is to bring designers and simulation engineers into the Physics AI workflow. I believe Physics AI models should initially be built by simulation analysts using their existing simulation data and then published as a library of models for different types of analysis.
Once that library is available, you can shift left. Designers can use those Physics AI models directly within their design software to assess whether the designs they are developing are feasible, without first going through a detailed design and simulation process. That is how the two areas will come together.
Further into the future, simulation and design environments may be brought directly into the design tool itself. Even then, however, simulation experts will still be needed. In my view, the two roles will never completely merge. Once the design space has been narrowed and a few designs have been selected for more detailed analysis, CAE experts will still have to take those designs through the traditional workflow to properly validate the selected design space.
This is how I expect the teams to mature. Physics AI models will first be created by analysts and then made available for designers to use. Over time, the two teams will develop a natural handoff and improve the models through an iterative process.


How are traditional physics-based simulation and Physics AI coexisting in engineering practice today? What determines which approach an engineer uses for a particular task?
Mahalingam    We provide our customers with an ecosystem of tools. Today, we have an application called Simcenter Engineering Intelligence. It provides a single pane of glass for all engineers, whether they are designers or simulation analysts.
By a single pane of glass, I mean that it manages all CAE work-in-progress data, extracts the associated metadata and brings tools from across the Simcenter portfolio into the same experience. It also manages a library of Physics AI models.
When a new design comes in, the application looks at what the engineer wants to do. If the objective is to explore different variations of the original design, for example, it will recommend using Physics AI as the inference engine rather than a physics-based solver.
Once the design space has been narrowed and only a few simulations require further validation, it will recommend using a physics-based solver. The application is intelligent enough to understand the stage of the simulation workflow and recommend the appropriate tool.
It can also make recommendations based on other service-level objectives. If compute resources are limited but a design still needs to be assessed quickly, it may recommend using a Physics AI approach even for detailed simulation. If sufficient compute power is available and the goal is to fully validate the design, it will almost always recommend a physics-based solver.
The two approaches therefore coexist within the same environment. The system makes intelligent recommendations, but the engineer ultimately decides whether to follow the recommendation or choose a different approach.



How Is Trust Established?

Physics AI is a surrogate model trained on a customer’s simulation data. If its prediction is wrong in a safety-critical automotive application, is there enough traceability to satisfy a regulatory audit?
Mahalingam    We are not saying that the inference produced by Physics AI for a particular type of analysis should be treated as the final answer. Physics AI helps engineers explore complex designs and very large design spaces extremely quickly. Once that process has narrowed the options down to one, two or perhaps three designs, a physics-based solver is used to properly validate them.
The same principle applies to safety-critical automotive design. We are not suggesting that a company should take the result from a surrogate model and move directly into manufacturing.
Two decades ago, physical testing was used to validate CAE. Today, CAE—the physics-based solver—becomes the validation tool for designs explored using Physics AI. With that validation structure in place, we do not expect the use of Physics AI to create regulatory noncompliance.


In Hyundai Motor Group’s case, Physics AI reduced subsystem optimization time from one week to 15 minutes. Could results generated by Physics AI eventually be accepted as evidence in crash-safety or ISO 26262 compliance processes? Or will final certification continue to depend on conventional physics-based simulation and physical testing?
Mahalingam    It took almost two decades for physics-based simulation to become an accepted certification tool in certain use cases and, in some of those cases, to replace physical testing. I believe Physics AI will need more time as well.
Until these models have examined all possible designs and design variants, and have been trained on enough permutations and combinations of data, they will not yet be treated as a certification mechanism.
In my view, physics-based simulation will remain the basis for certification for some time. But as these models continue to evolve with more and more data, I can see Physics AI maturing to the point where it could serve as a certification mechanism in perhaps five or six years.


Siemens says industrial AI is evolving beyond assistance toward autonomous AI agents capable of performing engineering work. Do you expect these agents to remain within a human-in-the-loop structure, or could some engineering tasks eventually be performed fully autonomously without human approval? Where should that boundary be drawn?
Mahalingam    In my personal opinion, achieving full autonomy at the enterprise level will be extremely difficult. Companies will continue to hire more industrial experts because the human brain is still the fastest pattern-recognition machine in the world.
Experts will still be needed to validate autonomous designs. That is how we see the human remaining in the loop, and I expect that structure to remain in place for a long time. I do not expect engineering to become a completely autonomous environment.
What will change is that repetitive and routine daily tasks will become fully autonomous. Experts will become involved when their judgment is required for a decision, and they will still provide the final approval. In my view, it will not be full autonomy.


Siemens plans to make the knowledge graphs and industrial ontologies at the core of Siemens AI Fabric available to customers. In Korean engineering environments that use a mix of third-party CAD and CAE tools, will these ontologies connect openly with third-party data, or will they remain self-contained within the Siemens portfolio?
Mahalingam    Siemens has always believed in an open architecture. That is why “adaptive” is such an important part of our overall strategy. Adaptive means not only that our products integrate well with one another, but also that they integrate with competing products in the market, open-source products and homegrown products.
Our approach to building ontologies is based on our experience in PLM and MES. Drawing on the domain knowledge we have accumulated over the past three decades, we plan to publish ontologies that we believe can be commonly used across industry.
Companies will be able to extend these ontologies across the heterogeneous toolchains used in their own environments. Even if a company does not use our PLM product, it can still map the ontologies and connect them to another PLM, MES or ERP product.
Our ontologies are based on open standards, including RDF. That allows us to publish them into any enterprise system and provide the meaning, relationships and context that AI requires.



Where Will Competitive Advantage Come From?

Physical AI ultimately learns about the real world through testing and virtual simulation. From Siemens’ perspective as a company with both physical testing systems and simulation software, will the real competitive advantage for future OEMs come from having more AI models or from owning physical test and simulation data?
Mahalingam    In my view, the future competitive advantage for automotive manufacturers will come from the speedups that AI provides. The question, then, is whether they need to own more and more physical test and simulation data to gain that advantage. I do not believe that is necessarily the case.
Even without historical simulation data, it is possible to synthesize a significant amount of data that is highly correlated with physical test results and physics-based simulations.
That is one reason we take such pride in Simcenter SimSolid. It is a very fast solver that works directly on CAD geometry without requiring a separate meshing process. Yet compared with conventional FEA, the difference is only around 1 to 3 percent and, in some cases, just 1 percent.
I therefore do not believe that a company must own both simulation data and physical test data to gain an advantage. What matters is whether it has the ability to generate and use data that correlates closely with physical testing and physics-based simulation. With that capability, it can gain a competitive advantage through AI.


If AI models themselves gradually become commoditized, what will truly differentiate Siemens and Simcenter in the simulation market five years from now: the AI technology itself, or the validation framework that makes its physical predictions trustworthy?
Mahalingam    That is a very good question. I recently posted about this exact subject on LinkedIn, and I also discussed it in a presentation to our simulation team.
Many AI startups are approaching engineering organizations and offering to build AI at the periphery. They are essentially offering to build surrogate models: “Give us your data, and we will build the surrogate model for you.”
This is where we differentiate ourselves. We are not simply providing AI at the periphery or as a wrapper. We are injecting AI directly into our simulation tools. This allows us to retain that differentiation and prevents the value from moving upstream to the startups.
This is why trusted outcomes are so important to us. A startup can build a surrogate model, but that model can only tell you what happens based on historical data and designs that are already known. If you introduce a completely novel design, that surrogate model is no longer useful.
By introducing technologies such as Physics AI, we are keeping that value within the simulation tools and within physics-based simulation. That is fundamental to our business model. We will provide our own AI agents and embed AI directly into the existing software.
Because we provide the full stack, from physics-based solvers to AI, customers do not have to use one company for peripheral AI tools and then come back to the Simcenter physics-based solvers to validate the results. This is how we preserve our value and differentiate Siemens from AI startups.


So, ultimately, this is all about “Engineering Truth.” Is that the correct way to understand it?
Mahalingam    Yes, that is exactly right. Consider the example of a wildfire. The parameters and surrounding conditions change very quickly and dynamically. If you use data from one minute ago to develop a response plan, conditions such as wind speed may already have changed by the time that plan is ready.
You cannot operate on a snapshot of old or stale data. You need real-time information. In the engineering world, that information must be grounded in Engineering Truth. Providing that foundation is the role of engineering companies like Siemens.


 

AEM(오토모티브일렉트로닉스매거진)

---