Deep
Principle

A robot pipettes. A model thinks. A reaction gets shorter.

Somewhere in Cambridge, a piece of software is deciding which experiment to run next. Not a scientist - the software. It reads what the last reaction produced, computes what should happen if it tweaks a catalyst, queues the test, and waits for the answer to teach it something. Then it does it again. This is Deep Principle on an ordinary Tuesday, and the ordinariness is the whole point.

The company is small - around fifteen people - and young, founded in 2024. It does not make a consumer app. It makes ReactiveAI, a platform that fuses generative AI, quantum simulation, and high-throughput experiments into a single closed loop. The pitch is unglamorous and enormous at the same time: most of modern industry runs on chemistry, and chemistry is slow. Deep Principle wants to make it fast.

"Deep Principle aims to revolutionize material discovery, drastically reducing the time and cost of traditional experimentation." - Deep Principle, on the record

It is a tidy mission statement. It is also a quiet accusation - because the way most labs work today looks a lot like the way they worked decades ago.

Discovery, priced like a moon landing.

Here is the part nobody puts on a brochure. A working materials lab might run one or two experiments a day. Getting a new material from the bench to production can take about ten years. A single lab-scale success can cost somewhere between two and twenty million dollars. And when the work is done, roughly 70% of researchers report trouble reproducing results, partly because more than half the data they generate is unstructured - notebooks, screenshots, intuition that never made it into a database.

Four numbers that explain why your phone battery did not get twice as good last year.

Catalysis sits at the center of this. It is involved in roughly 90% of all chemicals produced globally - fuels, fertilizers, plastics, medicines. Improve how chemists find catalysts and you nudge a startling share of the physical economy. The trouble is that the math behind it, density functional theory, is accurate and brutally slow. Run enough of it and you wait. A lot.

The bottleneck in science was never a shortage of ideas. It was the speed at which you could test them. - The thesis, stated plainly

Two chemists who decided to skip the waiting.

Haojun Jia and Chenru Duan met where a lot of improbable companies begin: a single research group at MIT. Both earned PhDs studying how to point AI and high-throughput quantum chemistry at the design of catalytic materials. They were, in other words, professionally annoyed by the exact problem they would later quit to solve.

The bet was simple to say and hard to do: stop treating computation, experiment, and machine learning as three separate departments. Wire them together so each one feeds the others. Let the model propose, let the simulation check, let the robot test, and let the result retrain the model. A loop, not a relay race.

They did not invent a faster pipette. They invented a faster way to decide which pipette to pick up. - On what "closed loop" actually means

In a charming twist, the founders later hired their own former PhD advisor. In March 2026, MIT's Heather J. Kulik - the Lammot du Pont Professor and a recognized pioneer at the intersection of AI and science - joined as Chief Scientist. When your old supervisor signs on to grade your homework in public, the homework is probably good.

ReactiveAI: one loop to close them all.

ReactiveAI is built on five in-house algorithm modules. On their own they would each be a respectable research project. Together they form a closed system that runs from "what should we try?" all the way to "did it work?" - covering discovery, property prediction, formulation, and experimental validation.

Most AI tools answer a question. ReactiveAI is designed to decide which question to ask next. - The difference between a calculator and a colleague

Around the platform sits a small constellation of work: AQCat25, a quantum-chemistry dataset that stretches into transition states and spin polarization; SAGA, a generalist research agent demonstrated across chemistry, biology, and materials; and SciClaw, launched in 2026 and pitched, with admirable nerve, as an "AI Research Partner." Not a tool. A partner.

Money, professors, and a published list of failures.

Skepticism is the correct response to any startup claiming a 20,000x speedup. So here is what can be pointed at. In March 2025, Deep Principle announced a Pre-A round of 100 million yuan - roughly the same neighborhood as its reported ~$13.75M total - backed by funds affiliated with Lenovo and Baidu. The stated use of funds: push ReactiveAI into new materials, energy, and pharmaceuticals. Coverage noted the company had already secured orders from industry leaders, including interest from the battery world.

There is also the academic trail, which is harder to fake than a pitch deck. The founders publish - in the Journal of the American Chemical Society and Nature Machine Intelligence, including work on using optimal transport to generate the elusive transition states of chemical reactions. And in a stunt that was also a result, the company helped build an evaluation framework that made some of the world's top large language models fail at scientific reasoning. The write-up reportedly drew over two million views. Publishing a test your competitors flunk is one way to make a point.

A 20,000x claim is a marketing line. A peer-reviewed paper is a receipt. Deep Principle keeps producing receipts. - For the skeptics in the back

The short, busy life of Deep Principle

First principles, on purpose.

The name is not decoration. In quantum chemistry, a "first-principles" calculation is one built up from the underlying physics rather than fitted to past data - the hard, honest way to predict how matter behaves. Deep Principle took that idea and made it a business: compute chemistry from the ground up, then let AI make the grind survivable.

The goal is to shrink the three taxes on discovery at once - time, cost, and irreproducibility. Less waiting on a single DFT run. Fewer million-dollar dead ends. And structured, machine-readable records so the next experiment can stand on the last one instead of guessing at it.

Make discovery cheap enough and you do not just find better catalysts. You change who gets to look for them. - The mission, said quietly

Their customers are not consumers; they are the R&D teams behind the materials, the batteries, and the medicines. If Deep Principle is right, those teams stop spending a decade and eight figures to learn that idea number forty-seven does not work - and start learning it by Thursday.

The Tuesday that gets shorter.

Return to that Cambridge lab. The robot still pipettes. The model still thinks. But notice what changed. The reaction that would have taken a graduate student a year of careful, lonely repetition now closes its loop overnight. The next one is already queued. The work that used to depend on a single brilliant person's intuition now compounds, structured and reusable, into the next experiment and the one after that.

That is the bet Deep Principle is making with about fifteen people and a platform named after the hardest, most honest way to do chemistry. Most of the physical world - the fuels, the fertilizers, the medicines, the batteries - is waiting on the answer.

If chemistry is the slow engine under modern industry, Deep Principle is trying to swap in a faster one - without anybody having to stop the car. - Where this goes next

Five things worth knowing

• The name nods to "first principles" - computing chemistry from physics, not from guesswork.
• Both founders did their PhDs in the same MIT lab, then hired their old advisor as Chief Scientist.
• HQ at 25 Ames St sits right inside MIT's campus - a short walk from where the idea started.
• Catalysis, the core focus, is involved in roughly 90% of all chemicals produced globally.
• The company helped build a science test that top LLMs failed - and the write-up cleared 2M views.