Rethinking Simulation: How CompLabs Is Building Foundation Models for Mechanical Design

Mechanical engineers need to simulate how their product will perform under various physical conditions. However, these simulations can take days to run and cost thousands of dollars. CompLabs is developing an AI model that can run them in seconds and at negligible cost.

Why Simulation Matters

Designing a product is fundamentally an optimization problem. And this optimization relies on iteratively testing and improving designs based on feedback. The faster you can test, the faster you can learn, the faster you can optimize.

It was for this reason that in the 1960s, numerical solvers were popularized to approximate real-world physical conditions. These algorithms enabled engineers to test products in silico, rather than through costly and time-consuming physical prototypes.

“But the complexity of designs and simulations kept increasing, causing the solvers to have exponentially longer runtimes,” says Chinmay Shrivastava, co-founder and CTO at CompLabs.

Further, these simulations require specialized expertise and often weeks of meshing (describing the geometry to the computer), costing large organizations hundreds of millions per year. The high cost of each simulation results in slow iteration and forces engineers to explore a restricted design space—leading to sub-optimal products.

What If AI Could Understand Designs and Physics?

The impact of AI understanding language has been significant, but what if AI could also understand designs and physics

Surrogate models are being used by advanced engineering teams to speed up simulations. They are narrow ML models trained on a specific simulation, requiring significant data and ML expertise. However, they can’t accurately model nonlinear physics, capture high fidelity, or maintain accuracy with changes to the geometry, material, or physics.

“Many teams experiment with surrogate models, but often invest significant effort for limited return. We’re providing them a faster, more generalizable way to accelerate simulations,” says Noah Evers, co-founder and CEO at CompLabs.

Rather than training highly-specific single-purpose models, CompLabs is developing a general AI model that understands 3D geometries and how they’re affected by physical conditions.

Their model is pre-trained on a diverse corpus of 3D geometries and physics data. This enables it to replicate the performance of a company’s solver after finetuning on prior simulation runs. Users can then run high-fidelity simulations in seconds on new geometries, materials, and physical conditions. Because the model can intuitively understand geometries, there’s no need for meshing.

For example, one large engineering company is planning to use the model to evaluate material and coating combinations for complex thermal components, compressing months of work into hours.

Looking Forward

CompLabs has raised $2.65M in pre-seed funding from Alt Capital, Cory Levy and Joris Poort (CEO of Rescale). The team is now working with engineering leaders across aerospace, automotive, and materials to accelerate their complex simulations.

Mechanical design today is slow: weeks of meshing, simplifying models to make them computationally tractable, and runtimes of days to months. These delays break the flow of design, making it hard to stay in a creative mindset.

Our goal is to pull mechanical engineers out of the weeds and empower them as designers.

With an AI model that understands geometry and physics, engineers will describe what they want—a part that can withstand certain loads, stay within thermal limits, or reduce drag—and the system will generate optimized designs.

Design will become more goal-oriented, fast, and creative.