Trim Is Building a Foundation Model for the Physical World

The Macro: Physics Simulation Is Computationally Brutal

Here’s a problem most people never think about. Simulating how physical systems evolve over time, waves on a beach, airflow over a wing, heat distribution through a material, requires solving partial differential equations. These equations don’t have neat closed-form solutions for real-world geometries. So engineers use numerical methods like finite element analysis, which discretizes the problem into millions of tiny cells and solves each one iteratively.

This works. It’s also painfully slow and scales badly. Double the resolution and you don’t just double the compute. You might 8x it or worse, depending on the problem. ANSYS and COMSOL dominate this space, and their customers routinely wait hours or days for a single simulation run. For some problems in astrophysics, the compute requirements are measured in centuries.

The AI-for-science space has been heating up. DeepMind’s AlphaFold proved that neural networks can solve problems that traditional computational methods struggle with. Companies like Orbital Materials (AI for materials science) and PhysicsX (AI for engineering simulation) are attacking adjacent problems. NVIDIA’s Modulus framework lets researchers build physics-informed neural networks. But nobody has shipped a general-purpose foundation model for physics simulation that works across domains. That’s the gap Trim is going after.

The Micro: A Transformer That Thinks in Physics

Trim built what they call the “Trim Transformer,” a custom architecture based on Galerkin-type attention. I’ll spare you the linear algebra, but the key insight is this: traditional attention mechanisms in transformers scale quadratically with sequence length. Galerkin attention reformulates the problem so it scales linearly. For physics simulation, where the “sequence” is the spatial and temporal state of a system, this is the difference between feasible and impossible.

The practical claim is that Trim’s model acts as a “constant-time lossy lookup table” for physics. You give it initial conditions, it gives you the future state. Not by solving the equations step by step, but by pattern-matching against the physics it learned during training. The tradeoff is accuracy. It’s lossy. But for many applications, getting a 95% accurate answer in seconds beats getting a 99.99% accurate answer in three days.

Emanuel Gordis, the founder, has the exact background you’d want for this. He’s NRC-licensed (that’s nuclear reactor operator, for the uninitiated), published quantum physics research with Princeton and astrophysics work with Lawrence Livermore National Laboratory. Cornell alum. The company came through YC’s Winter 2025 batch with a three-person team in San Francisco.

They’re planning to open-source the Trim Transformer and initial models, which is a bold move. It signals confidence in their execution speed. Open-sourcing the architecture means anyone can verify the claims and build on top of them. It also means competitors can too.

The Verdict

This is one of the most technically ambitious companies I’ve seen come through YC recently. A foundation model for physics is either a massive breakthrough or a massive overreach, and the line between those two outcomes is execution.

The applications Trim highlights are deliberately varied: autonomous vehicle path planning, gravitational wave detection for the upcoming LISA space-based detector, and general engineering simulation. That breadth is both the opportunity and the risk. A model that simulates waves on a beach and also predicts autonomous vehicle trajectories needs to generalize across very different physical regimes. That’s hard.

I’m cautiously optimistic. The founder’s credentials are strong and specific. Published physics research at top institutions, nuclear operator licensing, and the kind of deep technical background that makes the “foundation model for physics” pitch credible rather than hand-wavy. The decision to open-source early is smart for building trust in the scientific computing community, where proprietary black boxes get ignored.

At 30 days, I want to see benchmarks against ANSYS and COMSOL on standard test problems. Not marketing benchmarks. Published, reproducible ones. At 60 days, the open-source release needs to happen. Promises to open-source “in the coming weeks” have a way of slipping. At 90 days, the question is whether any real engineering team has replaced a production simulation workflow with Trim. If the answer is yes, even for a narrow use case, this company is going to be very interesting.