Siemens Digital Industries Software held the Simcenter Technology Conference (STC) at the Westin Seoul Parnas on May 21. It was the first integrated technology event held since the completion of the Altair acquisition a year ago. The company unveiled a unified portfolio spanning simulation, testing, HPC, and AI - but following the keynotes and brief conversations with those involved, it became clear that the weight of the event rested somewhere else entirely. Siemens was not announcing an upgrade to its simulation tools. It was laying out its entire industrial AI platform strategy for manufacturing.

By Sang Min Han_han@autoelectronics.co.kr
한글로보기






A New Competitive Standard - From Design Quality to Verification Speed

"We asked that 'Altair Technology Conference' be printed in small text at the bottom of the banner. We were worried customers who loved Altair might wonder where ATC had gone."
Byung Joon Oh, Korea Country Manager at Siemens Digital Industries Software, opened by addressing the event's name. Last year's Altair Technology Conference (ATC) had become this year's Simcenter Technology Conference (STC). The message was clear: Altair customers' investments would be protected and technical support would continue. One year after the merger, the name had changed - but the assets would not be cut off.
Sam Mahalingam, Senior Vice President and Head of Simulation, took the stage for the first keynote and described the reality facing manufacturers. The numbers speak for themselves. An average of 23 weeks from concept to prototype. Sixty percent of engineering time consumed by data management. Engineers navigating an average of more than 15 disconnected systems. Add to that roughly 10 hours of pre-processing for 300 components, up to 20 hours for numerical analysis, and additional hours for post-processing. Exploring the full design space takes about two weeks. Because systems are not connected, data gets regenerated - and regenerated data requires verification, which takes still more time. The problem is not tool performance. It is that data cannot flow.
Sam noted that the pace of technology adoption itself had changed. It took a decade for personal computers to go mainstream, another decade for the internet. Generative AI reached the mainstream in just two years.
"How can AI learn the language of engineering? How do we teach AI to understand manufacturing?" The question Sam posed carried the full weight of the keynote.
Unlike most generative AI systems trained on internet text, the world of engineering runs on fundamentally different data - 3D geometry, physical laws, material properties, load conditions. These cannot be learned from text. What is needed is not a general -purpose generative AI, but an industrial AI architecture capable of understanding engineering data, connecting it, and managing the full verification workflow. This is where Siemens' acquisition of Altair begins, and where the new Simcenter portfolio takes shape.
Oh framed the stakes through a comparison with China.
"Why is China so fast? Korea is ahead in efficiency within each domain - design, analysis, production - but slow when it comes to connecting those silos. BYD has cut more than 18 months from design to mass production, and it keeps getting shorter."
Disconnected silos mean data gets regenerated, and regenerated data requires verification that takes still more time. When Sam raised AI, he was proposing technology as the means to reduce that verification time. When Oh raised silo connectivity, he was arguing that the underlying architecture itself must change so that data does not need to be regenerated in the first place.



Sam Mahalingam, Senior Vice President, presented the realities facing manufacturers: an average of 23 weeks from concept to prototype, 60 percent of engineering time spent on data management, and engineers navigating more than 15 disconnected systems.



Physics AI - AI Trained on Engineering Ground Truth

Siemens' answer begins with Physics AI. Built on Altair's Geometric Deep Learning technology, the surrogate model is trained on decades of simulation data accumulated by customers, delivering up to 1,000 times the speed of conventional physics-based solvers.
Sam placed three approaches side by side in the keynote: traditional simulation analysis, Simcenter SimSolid - which operates directly on geometry without meshing - at 30 times the speed, and Simcenter Physics AI at 1,000 times. But the numbers are not the point. What changes is the number of design alternatives that can be evaluated within the same amount of time.
"Our Physics AI model is not a foundation model trained on internet data. It is a model that customers train directly on their own data."
This is not a rejection of foundation models. Siemens is actively building its own industrial foundation model. The training base is not generic internet data - it is customers' CAD files, simulation results, test data, and physics-based analysis outputs. Physics AI predicts and validates design candidates rapidly on top of that foundation. The concept is a foundation model trained on engineering ground truth.
Customer IP stays in-house while gaining the inference speed of AI. The system runs on-premises, in hybrid environments, or on public cloud - but in every case, data remains within the customer's single-tenancy environment. Model training requires no data scientists or coding; it is completed via point-and-click in a web browser. Simulation assets that individual engineers have managed on personal computers for decades are converted into AI assets shared across the entire enterprise.
Real customer results support this. Rolls-Royce reduced gas turbine engine structural analysis from months to days using SimSolid. Hyundai Motor Group cut its AI-based subsystem parameter optimization process from one week to 15 minutes - a 94 percent reduction. In a separate case, Hyundai Motor Company achieved 30 percent energy savings compared to conventional rule-based control, and a 16 percent improvement in cooling efficiency at the same power consumption. LG Vehicle Solutions integrated mechanical shock, thermal stress, and dynamic analysis - previously handled across multiple software packages - into a single model and solver, reducing development time by approximately 20 percent. SK Innovation improved battery electrode coating uniformity by 11 percent and reduced the defect rate by 50 percent. The numbers differ across cases, but the direction is the same: more designs verified in the same time, better results at lower cost.
Kyuwon Lee, Senior Executive Director, unpacked this Physics AI architecture from the perspective of engineers in the field. He presented three types of Simcenter AI: ShapeAI, which shortens analysis preparation time; Physics AI, which uses geometric deep learning to deliver fast analysis predictions; and romAI, which learns from high-fidelity analysis and test data to generate real-time models. Where conventional deep learning excels at predicting trends from structured data, Physics AI works directly with unstructured data - 3D CAD, mesh files, and test data.
"Combining Physics AI with HEEDS allows you to automatically generate hundreds of design alternatives with a single click and review them simultaneously within a single workflow."
On top of this sits a three-tier AI assistant concept: an intelligent assistant that checks analysis models and retrieves information via natural language; a skilled partner that analyzes results and recommends optimal solutions; and an autonomous executor that independently carries out repetitive tasks. The agentic AI is currently in development, with release expected by year-end or next year.
"Physics delivers trustworthy results. AI adds intelligence and speed. Data provides the context that connects those results to business outcomes. Only when all three operate together does the real value of engineering AI emerge."




A three-way comparison: traditional simulation analysis, SimSolid (30x faster), and Physics AI (1,000x faster). The numbers are not the point ? what changes is the number of design alternatives that can be evaluated within the same amount of time.


The Siemens AI Fabric architecture. Three layers: AI Foundation at the base, AI-native industrial applications in the middle, and the agentic enterprise platform at the top. Compatible with any cloud and any agent framework.



The Conditions for Trust - The Industrial AI Platform Siemens Is Building

If the first axis is reducing verification time, the second is connectivity. What Sam introduced under the name Trusted Outcomes is not simply a bundle of AI features.
The architecture Siemens calls the Siemens AI Fabric consists of three layers. At the bottom sits the Industrial Foundation Model; in the middle, AI-native industrial applications (Simcenter, Teamcenter, Opcenter, Designcenter, and others); at the top, an agentic enterprise platform. Each product comes with built-in copilots and agents, and MCP (Model Context Protocol) endpoints and skill files are exposed externally. Customers can build their own agents, or bring Siemens' agents into their own orchestration platforms. The architecture connects with any agent framework, any cloud, any data platform.
At the core of this architecture is the Knowledge Graph - the semantic layer. It connects disparate data silos, builds industrial ontologies, and provides the foundation for AI to understand the meaning of data as it operates.
"To trust agents and their outputs, a semantic layer is absolutely necessary," Sam said.
The trustworthiness of an agent does not come from the accuracy of data alone. It comes from whether AI understands the meaning of relationships between data. Siemens plans to build and open-source this industrial ontology. It will connect not only its own software - PLM, BMS, ERP - but also third-party data sources. Altair's RapidMiner becomes a core infrastructure component of this semantic layer.
The basis for that trust lies in the architecture itself. The LLM used in the copilot is an external model, but the RAG database and graph RAG connectors reside inside the enterprise. Every action executed by an agent is recorded with traceable data provenance. Security is not a policy problem. It is a design problem.



"Internalization May Actually Be the Slower Path"
- The Reality Facing Korean Manufacturing

Reducing verification speed with Physics AI, connecting data with the semantic layer, automating workflows with agents - that is the full picture Siemens has drawn. But no company can do all of this alone. Oh acknowledged that reality directly.
His opening remarks were less a technology introduction than a strategic argument about manufacturing. When technology was static, top-down decision-making worked. Today, AI is an environment that every employee accesses. The next transition begins before the current technology reaches maturity. Gartner's hype cycle no longer applies. New technologies emerge too quickly, and the judgment calls about which ones to bet on grow harder by the day. Working-level practitioners need a gateway to explore and experiment with new technologies, while leadership must restructure communication so those experiments stay aligned with the company's direction.
"The blind spot in internalization efforts is that they can be slow. What happens if a new technology emerges before the one you chose has taken root - Then you have to redirect internalization again, and the speed at which you produce innovation outcomes can actually slow down."
The philosophy of internalization is not wrong. But the pace of technological change has outrun the time that internalization requires.
Oh proposed a partnership architecture. If a company can handle 50 percent of what it needs on its own, the question becomes which partners will fill the remaining 50 percent - and whether those partners are open, innovative, and capable of two-way communication. Siemens is already building an open innovation ecosystem with NVIDIA, AWS, SAP, and SKT. With NVIDIA, the collaboration centers on digital twin rendering and physical AI training environments. With AWS, they are co-developing cloud-based service models. With SAP, both companies have been jointly developing interfaces that match PLM and ERP systems at the data model level for five to six years.
"The openness, innovativeness, and capacity for two-way communication of the partners you choose matters more than ever."
There is a place where this partnership architecture is already working. That place is HD Hyundai.




HD Hyundai Executive Director Young-woong Yoo presented the Design Spiral at his session. Formalized in 1959, this methodology has been the foundation of ship design for over 60 years. Its cycle of iteration - converging from requirements toward an optimal solution - cannot be resolved by upgrading individual tools.


"Having Nothing Gave Us the Freedom to Start Fresh"
- HD Hyundai's Proof

The presentation by Young-woong Yoo, Executive Director of Technology Planning at HD Korea Shipbuilding & Offshore Engineering, was the most grounded account of the day. He began by explaining why digital transformation in the shipbuilding industry had taken so long.
The CAD systems in use were old, and getting design data to the shop floor was a persistent bottleneck. That was the industry-wide problem. Ships are humanity's oldest form of transport, yet the way they are built has changed remarkably little. A single cruise ship contains 900,000 parts and requires 10 million hours of manufacturing time. The Newport News shipyard in the United States - specializing in aircraft carriers - produces one vessel every seven years. HD Hyundai builds 150 ships a year. The volume and complexity of design data that must be generated and delivered to the floor every day exceeds that of almost any other industry.
"Ships are the messiest product in existence, in terms of part count and complexity," Yoo said.
At the heart of ship design is the Design Spiral - a methodology that has guided the industry for more than 60 years, originally formalized by J. Harvey Evans at MIT in 1959. Starting from mission requirements, the process iterates through preliminary, contract, functional, detailed, and production design phases, converging toward an optimal solution. At every stage, hull resistance, stability, propulsion, structural strength, and electrical routing collide endlessly. Redesigning for resistance breaks stability; fixing stability reduces cabin space; enlarging the engine means redesigning the engine room. This cycle of iteration cannot be resolved by upgrading individual tools. What is needed is a fully integrated solution that brings every design process into a single centralized environment.
HD Hyundai began its digital transformation review in earnest in 2022, conducted proofs of concept with multiple vendors over two years, and selected Siemens as its partner. Full enterprise platform migration begins in 2028.
"The precision of design and the depth of analysis have come together on one platform. This is not a tool adoption - it is a decisive opportunity to establish a single digital backbone connecting design, analysis, and production."
The Siemens-Altair integration read, for HD Hyundai, as a foundation for connecting CAD and analysis tools into a single digital thread. Altair had long been a trusted platform for its ability to rapidly generate meshes from geometry and deliver actionable insights.
"I think the fact that we had nothing gave us the freedom to approach this with a more open mind."
The absence of legacy systems meant less resistance to adopting a new platform. In this sense, HD Hyundai's position resembles that of BYD - which entered electric vehicle development unburdened by an internal combustion engine heritage.
HD Hyundai's digital transformation roadmap runs in three phases: a visible shipyard by 2023, a connected, predictive, and optimized shipyard by 2026, and an intelligent, autonomously operated shipyard by 2030. The company is currently in the second phase. It has deployed the Palantir system to build a foundation for connectivity, prediction, and optimization - and is already seeing results. A demonstration showing the entire shipyard rendered as a digital twin, developed in collaboration with NVIDIA, was also unveiled. The partnership architecture Oh described was already operating on the shipyard floor.




The Simcenter Press Machine booth on the exhibition floor. A loop in which a digital twin controls a physical machine and physical data feeds back into an AI model was running live during the event. Virtual sensors replaced expensive physical sensors - design and operations connected through a single Simcenter environment.




Physics AI-driven generative engineering
The Next Advance in Simulation

If HD Hyundai demonstrated what open innovation looks like in practice, Sam turned to what comes next.
Physics-based simulation is not going away. AI broadens the scope of design exploration, while final designs are still validated by physics-based solvers. In safety-critical product design, statistical outputs are not enough - physical ground truth is required. When safety is at stake, people do not trust statistical results. The role of simulation solvers is not diminishing; it is becoming more important as the definitive validation checkpoint for the expanded design space that AI makes possible.
Physics AI-driven generative engineering is the next step. AI receives performance targets, automatically generates designs, and - when a design falls short in validation - revises it and outputs a manufacturable CAD file. It is a complete loop: from requirements to design, through physics validation, to manufacturable geometry. Where simulation has until now been a tool for verifying designs, here it becomes a tool for generating them. That reality is close at hand.
In electromagnetics, structures, and aerodynamics, the gaps that Altair has filled have raised the completeness of the integrated portfolio. Siemens' direction is to provide seamless workflows regardless of whether the software is its own or a third party's.



The Future of Simulation Is Not Faster Analysis

What Siemens presented at STC was not an upgrade to simulation software. Physics AI, semantic layers, agent orchestration, industrial ontologies, digital threads - these point toward a future in which engineering data flows freely, AI reasons and acts on top of it, and design is connected to production within a single continuous context.
"The technical content matters, but what needs to happen is for the entire lifecycle - materials, components, equipment, processes, factories - to become digitized and data-sharing. A system in which the entire supply chain between large enterprises and their suppliers is connected by a digital thread is the direction Korean manufacturing must move." Oh said.
The Altair acquisition is not simply a portfolio expansion. Altair's geometric deep learning and HPC technologies, along with RapidMiner's semantic layer capabilities, form the core infrastructure for Physics AI and the Knowledge Graph. Siemens did not buy Altair. It acquired the pieces needed to complete an architecture it had already been designing.
The future of simulation is not faster analysis. It is a system in which AI trained on engineering ground truth explores the design space, agents execute workflows, and a digital thread connects the results all the way through to production. Siemens is now writing the OS for that system. STC was not an event with a new name. It was something else entirely.


 

AEM(오토모티브일렉트로닉스매거진)