Topological Is Making Physics Simulations 1,930x Faster and Engineers Are Paying Attention

The Macro: Hardware Engineering Is Stuck in a Simulation Bottleneck

I watched a mechanical engineer run a topology optimization last year. She set up the constraints, defined the load cases, specified the material properties, hit run, and then went to lunch. When she came back an hour later, it was still running. She went home. The next morning, it had finished. One design iteration, overnight.

This is normal in hardware engineering. Topology optimization, the process of finding the best possible material layout within a given design space to meet physical constraints, is computationally brutal. The math involves solving partial differential equations across millions of mesh elements, iterating hundreds of times to converge on an optimal structure. Traditional finite element methods are thorough but slow. A single simulation can take hours. A design exploration involving dozens of variations can take weeks.

Software engineers iterate in seconds. Push code, see results, adjust, push again. Hardware engineers iterate in days. Run simulation, wait overnight, review results, modify CAD, run again. This speed asymmetry means hardware products take longer to develop, cost more to design, and explore fewer design alternatives than they should. The best design is often the one the engineer never had time to try.

The simulation software market is dominated by Ansys, Siemens (with Simcenter), Altair, and COMSOL. These are excellent tools with decades of development behind them. They are also built on computational methods that are fundamentally limited by physics solver speed. You can throw more compute at the problem, but the underlying algorithms have diminishing returns. This is not a cloud compute problem. It is an algorithmic problem.

Foundation models changed NLP and computer vision by learning general patterns from massive datasets and applying them to specific tasks without solving each problem from scratch. The same approach applied to physics simulations is one of the most interesting ideas in deep tech right now.

The Micro: Two Stanford Engineers Teaching AI to Think About Physics

Nicole Ma and Zeyneb Kaya met at Stanford, where Ma studied computer science and mechanical engineering and Kaya studied computer science, physics, and math. Ma has experience running physics simulations at NASA. Kaya’s research spans reinforcement learning, pretraining methods, efficient algorithms, and synthetic data generation. They founded Topological in San Francisco and came through Y Combinator’s Summer 2025 batch. The team is two people.

Their first model is called UToP-v1. It performs topology optimization with less than 5% compliance error at 1,930 times the speed of conventional finite element methods. I want to make sure those numbers land properly. A simulation that takes 10 hours with traditional methods takes roughly 19 seconds with UToP. A design exploration that takes two weeks compresses into an afternoon.

The model works by learning the underlying physics of how materials behave under stress, rather than solving the equations from scratch every time. It understands geometry, load conditions, and manufacturing constraints as learned patterns, not as problems to be computed through brute force. This is the foundation model approach applied to physical engineering: train on enough examples and the model develops intuition about how structures should look, the same way a language model develops intuition about how sentences should read.

The manufacturing constraints piece is critical and often overlooked. A topologically optimal structure is useless if it cannot actually be manufactured. Traditional topology optimization frequently produces organic, tree-like structures that look beautiful in simulation and are impossible to machine, cast, or 3D-print within budget. If UToP bakes manufacturability into the optimization itself, that eliminates an entire round of post-processing that engineers currently do by hand.

The website is minimal. No pricing, no customer logos, no case studies. This is clearly an early-stage company that is talking to potential customers through direct outreach rather than inbound marketing. The Calendly booking link and founders@ email address are the only calls to action. For deep tech, this is normal. The sales cycle for engineering simulation tools is long and technical, built on demonstrations and pilot programs, not landing pages.

I want to see how UToP handles complex multi-physics problems. Topology optimization for a single load case in a single material is one thing. Real engineering problems involve thermal stress, vibration, fatigue, multiple materials, and load cases that change over time. The 5% error claim needs context about what types of problems it applies to. If it only works for simple 2D structural optimization, the 1,930x speedup is interesting but limited. If it generalizes to 3D multi-physics problems, this is a paradigm shift.

The Verdict

Topological is the kind of deep tech bet that either changes an entire industry or stays in the research phase for years. The speed improvement is dramatic enough that, if validated across complex real-world problems, every engineering team in aerospace, automotive, consumer electronics, and industrial equipment would want it. That is a large market.

Ansys is the elephant in the room. They have deep pockets, existing customer relationships, and their own AI initiatives. Siemens and Altair are investing heavily in simulation acceleration too. The startup advantage is focus and speed. The incumbents are adding AI features to 30-year-old software architectures. Topological is building from scratch with modern ML infrastructure. nTopology (now nTop) plays in the design-for-additive-manufacturing space and could be a complementary tool or a competitor depending on how Topological’s roadmap evolves.

At 30 days, I want to see a published benchmark comparing UToP against Ansys or COMSOL on a standardized test case. Claims need validation from people outside the company. At 60 days, the question is whether any engineering team has used UToP for a production design, meaning a part that actually got manufactured and tested. At 90 days, I need to understand the business model. Is this a SaaS subscription, a per-simulation charge, or an API for integration into existing CAD workflows? The answer to that question determines whether Topological is a tool, a platform, or a feature that gets acquired by one of the incumbents. All three outcomes could be good. But they require very different companies.