Dassault Systèmes SIMULIA, current

An AI assistant for simulation

Engineers make real design decisions based on simulation results, so an AI assistant for simulation has to be more than fast. It has to get the setup right, show how it got there, and check with the engineer before anything important. This is the approach I’ve been proposing and prototyping for fluid simulation.

RoleUX lead for fluids, one voice in a company-wide initiative
CompanyDassault Systèmes SIMULIA
My partA proposed interaction model, principles, coded prototypes
StatusIn progress, internal, direction still open
Simulated airflow over a car with a vortex wake behind it, and the assistant asking whether the road should move with the car before it builds anything

A note on scope

AI at Dassault Systèmes is a large effort, with many teams and people far more senior than me working on it. This page is about my own contribution: a point of view on how the assistant should work for fluid simulation, which I’ve been prototyping and presenting inside the company and to other domain leaders.

The work isn’t public, so the pictures and the walkthrough here are my own illustrations made with AI, not the actual product.

The question

Most AI assistants are built around a chat box. Simulation is different. Setting up a CFD case is highly technical work that involves dozens of decisions about geometry, physics parameters and mesh, and even one wrong value can produce results that look fine but aren’t.

So the design question is how an engineer stays in control of a result they’ll be accountable for, while the agent does more of the work.

My hypothesis

My proposal starts from a hypothesis about how engineers want to work with an agent, based on some early research I did on simulation augmented by AI. They want to build the case together with it, not hand it over to an AI to build on its own. Users want to stay in control of every decision that affects the answer. And they want to check things as they go, because that’s how trust gets built.

Control isn’t fixed, though. Once an engineer trusts a setup, they’re happy to hand off the repetitive part, like running a parametric design study across dozens of variants. Early on, they want to do it themselves. So the assistant has to earn autonomy, not assume it.

Converge before you compute

I think of the assistant’s job as a triple diamond, a bit like the double diamond in design. Engineers come in with a hundred different kinds of requests: a vague question, a half-built setup, a result that looks off. Whatever the starting point, the assistant has to pull the conversation toward three convergences, in order. In between, each engineer can open things up again in their own direction, and the assistant’s job is to bring it back to the next decision.

123

1. The objective

What are we actually trying to find out? Drag at 70 mph, whether a pump cavitates, how hot a board gets. Until we find this out nothing else matters, because the objective decides everything after it.

2. The features to get there

The AI assistant works out the physics and features needed to reach that objective: flow regime, boundary conditions, materials. It proposes, and the engineer agrees or makes changes.

3. Ready to run

The last pass covers fidelity and anything still missing: mesh refinement, solver settings, gaps in the geometry. Only after this does anything get solved.

Each diamond closes with the engineer saying yes. That’s where control and trust live. An early version of my proposal had separate planning and execution modes, but over time, watching real workflows showed our product ontology has some natural checkpoints, so I rebuilt it around them.

Principles I’m proposing

  • Organize help around what the engineer is trying to do, not around the software’s menus. People think “check if this pump cavitates,” not “open the boundary condition dialog.”
  • Show where every value came from: the user, the CAD model, a default, or an assumption the agent made. Engineers can only trust what they can trace.
  • Tie every result to the exact model and scenario that produced it, so nothing silently goes out of date.
  • Keep one consistent character across every task, so it feels like the same assistant whether it’s meshing a pump or setting up airflow over a car.

Try it

Play the engineer in a concept-car drag study. The tag at the top shows which diamond you’re in. The assistant proposes, asks and explains, and nothing is built or solved until you say so. The tree on the left fills in as it works.

Concept car, external aerodynamics
{{ phase }}Step {{ n }} of 7
Model tree
Concept car
{{ caseNote }}
Analysis Case.1{{ c1status }}
Physics: steady
Air Domain.1
Boundary Conditions (5)
Road: {{ floorShort }}
Mesh: medium
Close small gaps
Result: {{ d1 }} N
Analysis Case.2Solved
Geometry: no mirrors
Result: {{ d2 }} N
made by the assistant
What’s the drag on this concept car at 70 mph?
So the goal is the drag force on this car at 70 mph. Before I set anything up, one thing defines that number: should the road move with the car, like real driving, or stay fixed, like your wind tunnel?
{{ floorAnswer }}
Here’s the setup I’d build. Nothing is created yet.
Proposed setup
Speed: 31.29 m/s (70 mph)From you
Speed: 35.76 m/s (80 mph)From you
Road: {{ floorText }}From you
Car body and wheelsFrom the CAD model
Air at 25 °CDefault
Steady, k-ε RealizableFrom your team’s aero workflow
Tunnel: 3 car lengths ahead, 8 behindFrom your team’s aero workflow
Mesh: medium, about 6 million cellsEstimated
Build this setup.
Case built. The build check found one problem. The left mirror doesn’t quite meet the door. A 0.3 mm gap leaves the car body open, so the air volume can’t close. I can close gaps smaller than 0.5 mm. That changes only the geometry used for the simulation. Your CAD stays as it is.
Show me where.
It’s highlighted in the viewport: the base of the left mirror, where it meets the door panel.
Close the gap.
Gap closed. Build check passed. Before we spend compute, here’s the whole case for your sign-off.
Solve check: 17 features, each linked in the tree
PhysicsSteady, k-ε Realizable
Boundary conditions5
MeshMedium, about 6 million cells
MeasurementsDrag force, Cd
Stopping criteriaCd within 0.5%
Estimated run: about 14 minutes on GPU.
Run the solver.
Drag at {{ mph }} mph: {{ d1 }} N Cd {{ cd1 }}. Most of it comes from the rear wake and the wheels. This result belongs to Analysis Case.1, with exactly the setup you approved.
What if we remove the mirrors?
I made Analysis Case.2 from Case.1. The only change is the geometry: no mirrors. Case.1 and its result stay exactly as they were.
Without mirrors: {{ d2 }} N Cd {{ cd2 }}, about {{ pct }}% less drag than Case.1.
Who decided what?
You made every call that changes the answer.
You decidedSpeed: {{ mph }} mphRoad: {{ floorShort }}Build the setupClose the mirror gapRun the solverTry it without mirrors
The assistant filled inPhysics model, from your team’s workflowTunnel size, from your team’s workflowAir properties, a defaultMesh size, an estimateCase.2, a copy of Case.1
All of it is kept with the result, so whoever reviews it can see where each value came from.
Why this step {{ note }}
An illustrative example with made-up values. It is not the product or its interface.

Where it stands

I built interactive prototypes, storyboards and videos across different kinds of flow problems, from pumps to external aerodynamics to cooling. Having something people could watch or click through got developers, product managers and other designers reacting to the same thing, and the response has been encouraging.

My proposal is one of several directions being discussed, and nothing has been finalized yet.