He asked whether a machine could predict a reaction it had never seen. Then he built a company to find out.
Haojun Jia runs Deep Principle, a company with an unusual employee at its center: an AI scientist. Its job is to invent - to propose new materials and chemical reactions, coordinate the experiments that test them, and shave months off work that used to crawl. Jia is the co-founder and CEO. His pitch fits on a lab coat pocket: deep learning plus first-principles physics, pointed at the messiest problem in industry, which is finding the next useful molecule before someone else does.
The idea did not arrive in a boardroom. It arrived at MIT, where Jia was deep into a PhD in physical chemistry, studying how single atoms behave when you ask them to do the work of a catalyst. Somewhere between the quantum calculations and the coffee, a question stuck: if the physics is knowable, could a model learn to predict chemistry it had never actually seen? He turned that question into Deep Principle rather than into another paper.
Most founders sell software. Jia is selling something closer to discovery itself. The company's platform - marketed first as ReactiveAI and now as the agentic tool Agent Mira - lets a researcher describe a goal in plain language and have the system orchestrate wet-lab experiments, simulations, and AI models to chase it. The target markets are the ones where a better material changes the math for everyone downstream: batteries, polymers, biodegradable plastics.
Before chemistry, before Cambridge, there was physics - and a lot of frequent-flyer distance. Jia earned dual bachelor's degrees in 2019, one from Jilin University in China, the other from National Research Tomsk Polytechnic University in Russia. It is the kind of resume line that sounds like a rounding error until you sit with it: he learned the same laws of nature twice, in two systems, in two languages.
In November 2019 he joined the Kulik Research Group at MIT as a PhD student in physical chemistry. His research homed in on the spin-state-dependent properties of single-atom catalysts - the smallest functional units you can build a reaction around. His undergraduate work had already wandered across 2D piezoelectric materials, high-pressure phase transitions, surface science, and gas-phase chemistry. A generalist's curiosity, aimed at very small things.
Deep Principle did not appear from thin air. During his PhD, Jia authored or co-authored roughly eight publications on subjects like codoped single-atom catalysts for turning methane into methanol, iron and ruthenium catalysts for oxygen reduction, and new techniques for modeling the fleeting transition states where reactions actually happen. The company is, in a sense, that research grown a size too big for a journal.
Traditional materials discovery is slow because reality is stubborn. You hypothesize, you synthesize, you wait, you measure, you mostly fail, you repeat. Deep Principle's answer is to put a coordinating intelligence on top of the whole loop. You tell it what you want. It reasons across first-principles calculations and learned models, proposes candidates, and helps drive the experiments that confirm or kill them.
The bet underneath all of it is a philosophical one. Pure data-driven AI can pattern-match. Physics-grounded AI can reason about why a molecule behaves the way it does. Jia's wager is that the second kind wins in chemistry, where the training data is expensive, scarce, and often the result of an experiment nobody can afford to run twice.
Sources: Forbes, MIT Kulik Research Group, Google Scholar, Deep Principle, China Beat, World Economic Forum, XtalPi. Figures such as total funding are drawn from public reporting and may change. Illustrative charts are labeled as such.