In June, Dassault Systemes' SIMULIA User Day was an event about AI, but on stage, MODSIM was emphasized more than AI itself. What the company repeated was not a bigger AI model or a faster algorithm, but the workflow connecting design and analysis, the data accumulated along the way, and the traceability that lets you follow the whole process from start to finish. AI is changing engineering - that much is true. But for that AI to work, people, data, and connected processes still matter. So on this day, AI appeared as the last step. In today's industrial AI race, the real contest may not be about who has the smarter model, but about who has more context that model can trust and work with.

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





"Work that used to take 40 hours becomes 4 hours with MODSIM, and with AI, it can become 4 seconds."
That's what Michelle Ash, CEO of Dassault Systemes' SIMULIA, said on stage at SIMULIA User Day 2026.
Right now, the topic in simulation is AI. But what was emphasized most that day wasn't AI - it was MODSIM. Ash called MODSIM the first step toward using AI, and Francesco Polidoro, Director of SIMULIA Worldwide MODSIM and AI Technical Sales, spent much longer unpacking that step. LG Electronics Research Fellow Kim Yong-yeon talked about a change that hasn't yet reached his organization, even as the industry as a whole moves quickly toward AI. And in an interview, Ash again addressed the conditions under which AI can be trusted.
The people who gathered to talk about AI spent more time explaining how to connect design and analysis, how to leave a trail of that process, and how engineers actually work. This article follows that order.







Why 40 Hours Becomes 4 Seconds

Along with the line "work that used to take 40 hours becomes 4 hours with MODSIM, and with AI, it can become 4 seconds," Ford says that "MODSIM brought the joy of engineering back." Ash highlighted this case and these numbers from Ford.
When design and analysis were done separately, reviewing a single design proposal took 40 hours. After applying MODSIM, that time dropped to 4 hours. What matters is that in the same amount of time, you can now review ten design proposals instead of one.
MODSIM is the name Dassault gives to its approach of handling design (CAD, CAE) and analysis (simulation) together on a single dataset. Until now, these two tasks ran separately. A designer would create a shape, hand the file off to an analysis engineer, who would build a mesh and run the simulation. If the design changed based on the results, the whole process had to start over. MODSIM puts both on the same data, so that when the design changes, the simulation follows immediately. Cutting that 4 hours down to 4 seconds is ultimately AI's domain.
Of Dassault's 1,600 SIMULIA employees worldwide, about half work in R&D, and roughly 400 of those work on solvers like Abaqus, PowerFLOW, CST Studio Suite, and Simpack. The work to run these solvers 3 to 25 times faster on GPUs is underway in partnership with NVIDIA. This NVIDIA partnership runs along two tracks: running solvers faster on GPUs, and designing NVIDIA's own data centers using Dassault's simulations. Of the five collaboration areas the two companies announced, the two Ash is directly involved in are data center development and applying physical AI to industry.
"MODSIM is the first step toward using AI."
That four seconds means nothing if the steps before it aren't in place. You make the solvers fast, then connect design and analysis on top of that with MODSIM, then put AI on top of that. And on top of all of it sits data center design and an industrial agent platform.
That was the order of the entire day at SIMULIA User Day.









Why MODSIM

"Most of the work that gets done disappears in the process of passing files back and forth and converting them."
That's what Director Polidoro said while explaining the existing modeling-simulation process. A designer creates a shape and hands the file to an analysis engineer. The analysis engineer builds a mesh and runs the solver. If the design changes, the whole process has to start from scratch. MODSIM turns this linear structure into a circular one. Design and simulation proceed simultaneously on the same data.
What MODSIM integrates isn't just structural analysis. Five branches of physics - Structures, Fluids, Motion, Vibro-acoustics, and Electromagnetics - all sit on a single object. These five domains are also a history of the solvers Dassault has acquired over a long period of time. Names like CST, exa, wave6, and Opera are traces of companies Dassault has acquired. MODSIM is the work of bringing that legacy into a single platform.
Within the same data model, the same user environment - when you change a shape, the analysis side gets notified, and a few clicks bring back new results. In Polidoro's presentation, this scene repeated across examples: an excavator seat, the electric motor of a robot arm, the injection molding of a beverage bottle.
"You can always trace what simulations were performed on this shape. You can also work backward to find what model the results were based on. This kind of contextualized, structured data becomes the starting point for AI."
If you're designing a beverage bottle, you optimize the thickness of the parison through injection molding, then connect those results directly to compression and drop simulations. Bundle this whole process into a template, and even non-experts - without scripts or complex coding - can adjust a few parameters on a web screen and repeat the same analysis.
That's the technology story. The question is what changes, and how, once this technology enters a company.







What Changes Before AI Does

Looked at across the whole industry, the pace of change has been steep. Just a year ago, 91% of respondents in manufacturing and mechanical engineering said they used AI not at all. In just one year, that figure dropped to 55%, and the adoption rate of agentic AI in manufacturing quadrupled from 6% to 24%. AI is no longer an experimental-stage technology. It's beginning to enter actual work.
GDPval scores for mechanical engineering point in the same direction. Treating a score of 50 as equivalent to an industry expert, a score that was 25.3 a year ago has reached 49.5 in 20 months - essentially closing in on expert level.
But the problem LG Electronics Research Fellow Kim Yong-yeon saw was somewhere else.
"Most CAE engineers aren't in positions where they can wield much influence within the company. The people who look at the data first aren't the ones in a position to move the company with that data."
According to Kim, the work involved in developing a single product breaks down into 22 categories and roughly 440 detailed tasks. Of those, 34% are buried in documentation and reporting. While AI across the industry is closing in on expert level, actual organizations are still spending much of their time passing around and organizing data.
So Kim's solution started not with tools, but with data. Build a data map of design knowledge, and connect the digital threads that are broken between departments. Turn AX into DX, and DX into CX. And don't stop at connecting data internally - bring customers and partners into the same digital thread.
"Physicality. There's still a domain where mechanical engineers are clearly needed - anywhere that touches the human hand."
Tools come after data, and data comes after connection. This is the same point Polidoro made when he said "contextualized, structured data is the starting point for AI." The work of building that context is still underway, everywhere.







Trust Comes From Structure

When AI enters engineering, the first question isn't performance. Can the result be trusted? Where did this design come from, what analyses did it go through, who changed it, and on what basis did it become the final version? That's also why, when asked about IP protection, Ash brought up traceability first.
"We can trace the whole thing - from CAD, through various simulation work, through changes made via MODSIM, all the way to the final validated product."
This wasn't simply an explanation of data management. It meant that for AI's answers to be trustworthy, you have to be able to trace the entire process backward, all the way through how the product was made. Designs that never reached final validation, parts that were tried and discarded, parameters that were changed midway - all of it remains in the system as a trace. That's what makes certification, validation, and audits possible.
Access is the same kind of problem. Ash explained data in terms of three domains: information anyone can access, Dassault's industrial knowledge and know-how, and the customer's IP. When agentic AI is at work, what matters isn't simply how much data it can access, but clearly defining who can access which data. Within the same company, some users should be able to access certain information while others shouldn't. What Dassault calls sovereign clouds and AI data governance is the structure that manages this boundary. Before AI can be brought onto the floor, you first have to decide how much it knows and how much it's allowed to do.
What this kind of traceable data actually looks like was visible in the demo that day. An F1 car appeared at full scale in virtual space, with airflow and pressure mapped onto it in color. The visualization itself wasn't the point. What filled the screen was analysis data that had originated in CAD and passed through MODSIM. The scene was possible because design, simulation, and the history of changes all lived within the same context.
LG Electronics' presentation was telling the same story in different words. At every stage - requirements, design, validation, response - data gets disconnected, and it takes time for results to feed into the next decision. Kim called this a "disconnected digital thread." If Dassault was talking about traceability and trust, LG Electronics was showing what disconnected data costs inside an organization.



The Expert Doesn't Disappear

Companion. A word repeated throughout SIMULIA User Day.
Dassault called AURA, LEO, and MARIE not Assistants or Agents, but Virtual Companions. AURA was "The Business Expert," LEO was "The Engineer," and MARIE was "The Scientist." LEO took its name from Leonardo da Vinci, MARIE from Marie Curie. AURA wasn't named by a person - the AI agents themselves came up with it, as a shorthand for "Assisting You to Realise your Ambitions." They're companions in the literal sense - fellow workers.
The scope of what this companion can do is clearly drawn.
"I think non-expert use will stay in relatively low-risk areas - checking physical behavior at the design stage and adjusting the design, that kind of thing. Final validation, checking safety and certification requirements - those stages still belong to experts. No government, and no careful company, is going to change that anytime soon," Ash said.
The expert's role doesn't disappear - it shifts. From someone who does all the work directly, to someone who designs the workflows and guardrails that non-experts will use. Someone who checks whether a tool is being used within permitted bounds, and who's responsible for the final validation. Ash drew on her own background and pointed to large trucks in mining. Driving them is already automated, but that doesn't mean human judgment has disappeared. Even automated operation still runs within limits and procedures set by people.
This kind of role shift comes with a redesign of the workflow itself. One Dassault customer started with 5 workflows, then found 52 more, and is now redesigning 57 of them. It's not that the workflows were wrong - in the old context, they were perfectly optimized. It's that under the new technology, the optimization itself had to be redesigned.
Agents act on your behalf. Companions act alongside you.
The reason AI came last that day is that MODSIM, data, and traceability are all things AI can't do alone. Without MODSIM, there's no 4 hours. Without data, there's no traceability. Without people, there's no validation, and no control.







Beginning Model

Everyone is talking about industrial foundation models these days. Over time, the models themselves may come to resemble each other. Ash called this the "Beginning Model." The difference doesn't come from the model itself, but from what you build on top of it.
What Dassault kept repeating wasn't an AI model either. It was a single platform connecting CAD to simulation, manufacturing, data management, and supply chain - a way of linking the entire process of making something virtually first: building it, dropping it, assembling it, and using it again, all before it exists physically. LG Electronics was what that vision looks like once it enters a company's research lab. Breaking down a vast number of tasks, finding the disconnected digital threads, and pulling resources that had piled up at the validation stage forward to the early stages of design.
At SIMULIA User Day, AI didn't appear as a new feature. It appeared as the last step - the one that shows what has to change in engineering first, before AI can actually work.

AEM(오토모티브일렉트로닉스매거진)