A Single Industrial Simulation Would Take 10 Hours. AI is Cutting That Down to Seconds

For decades, engineering teams have had to balance simulation accuracy with the time and computing resources needed to run it. A product can be tested virtually for structural strength, aerodynamics, vibration, and other properties. Every additional design iteration can add hours of computation and engineering work.

Ravikiran Pothukuchi, Director of Enterprise Applications of industrial AI company Dassault Systèmes India, draws a line between general-purpose generative AI and AI built for engineering. 

"For industrial AI or scientific AI—what we call 'virtual companions' in engineering—you can't put a disclaimer on a design," Pothukuchi tells AIM. "We combined generative and agentic AI, grounded in physics and built on top of statistical foundations, to create AI that has to be right, not guess or hallucinate."

From Simulation to Design Exploration

Simulation has traditionally been separated from product design. Engineers create a design and pass it to specialist teams for structural, fluid dynamics, or noise and vibration analysis, with changes from one discipline often triggering another round of work.

Dassault Systèmes has sought to reduce this back-and-forth through MODSIM, its approach to unified modelling and simulation on the 3DEXPERIENCE platform. AI-powered MODSIM uses machine learning and surrogate models to accelerate simulation and let engineers explore more design variations.

Pothukuchi said engineers working within fixed timelines are traditionally limited to testing 10–20 design alternatives. With AI-assisted exploration, that number can run into the thousands within the same window, lowering the odds of missing the strongest possible design.

High-fidelity simulations still require significant computing resources. Pothukuchi says a single run could take about 10 hours on a high-performance computing cluster. Surrogate models trained on existing simulation data can estimate results in seconds, reserving full physics-based runs for a small set of final candidates.

"Instead of running 1,000 high-fidelity simulations at 10 hours each, you run the surrogate model for most of those 1,000 iterations, fast and approximate, narrowing in on the best candidates, and only run the full high-fidelity simulation two or three times at the end," he adds. Dassault Systèmes says this can cut total simulation time from roughly 1,000 hours to about 20.

The savings show up on both sides of a company's ledger. Lower compute time and fewer HPC cycles reduce cost directly. Faster development also gets products to market sooner, which Pothukuchi frames as a revenue advantage tied to being first to market.

Three Agents, Three Jobs

Dassault Systèmes has built three AI agents that work across different parts of a customer organisation. 

Aura acts as a business analyst, drawing on industry and customer data to answer questions about cost, sourcing and margins.

Leo, named after Leonardo da Vinci, is positioned as an engineering assistant, carrying technical knowledge across materials, structural stiffness, and other design decisions engineers face during development. 

A third agent, Marie, named after scientist Marie Curie, is focused on fundamental science and works at the level of physics, chemistry and molecular modelling.

Aura's role becomes clearest in supply-chain decisions. Pothukuchi cites tariffs as an example.

"If tariffs hit global supply chains, the business agent can provide intelligence suggesting, for example, sourcing a component from a different country plant of the same supplier, since tariffs are country-specific, rather than just switching suppliers entirely," he explains. He points to a case involving a 20% cost increase moving through a supply chain, where this kind of analysis helped identify a lower-tariff sourcing option and reduce the hit to margins.

Beyond Cars and Planes

Pothukuchi pushes back on the idea that these tools serve only heavy industry. Consumer products like bottles, laptops and phones go through similar testing, including drop tests to check whether small, loosely connected internal parts survive a fall.

Simulation software suite SIMULIA and computer-aided design software CATIA are also used together to calculate a product's carbon footprint at the design stage, a measure Dassault Systèmes calls an "eco-bill." 

Pothukuchi says this lets manufacturers compare material choices, such as paper against plastic packaging, and weigh durability against environmental impact before a product is built. 

The same approach extends to manufacturing, where designs have to be turned into production and sourcing decisions. 

"Producing a car might involve around 10,000 components. You won't manufacture all of them in-house, you decide what to make versus what to buy," Pothukuchi illustrates. Dassault Systèmes says its AI capabilities support bill-of-materials creation and manufacturing process planning.

Virtual validation matters here because physical mistakes on the shop floor are hard to reverse. "Unlike software, where you can hit 'undo,' physical manufacturing errors are expensive and can't be undone," he notes.

From Prompts to Physical Products

Dassault Systèmes is looking further ahead to prompt-driven design. "Prompt-driven design, the longer-term vision is a designer giving a detailed prompt and generating a fully designed model, for example a car," Pothukuchi says.

That vision depends on keeping engineers in the loop rather than removing them. Repetitive work, including parts of early-stage modelling, is increasingly handled by AI, shifting engineers toward reviewing and validating results instead of producing every iteration by hand. 

SIMULIA sits at the centre of that validation step. It tests whether a car can survive a crash, whether a building can withstand an earthquake, and how a design performs under fluid and vibration loads before production begins. 

Three or four decades ago, that kind of testing meant building and destroying hundreds of physical prototypes. Companies today run as few as one or two physical tests, with the rest validated virtually through simulation.