Chinese scale-up Deep Principle, a World Economic Forum Tech Pioneer, uses generative AI, quantum chemistry, and high-throughput experimentation to accelerate materials discovery, property prediction, formulation optimization, and controlled experiments. The company’s flagship offering is the AI Scientist platform, recently upgraded into an agentic system called “Agent Mira,” which unifies six algorithmic modules — ReactGen (materials generation), Reactify (precision computation), ReactControl, ReactBO (broad screening), ReactNet (synthesis navigation), and ReactHTE (high-throughput experiments).
The Hangzhou-based scale-up frames its approach around what it calls the “ECML” paradigm — a unified AI decision framework coordinating Experiment, Compute, and Machine Learning — which it positions as the successor to trial-and-error, top-down experimental chemistry.
The company’s ambition is to transform “material innovation from discovery to design,” integrating generative AI, quantum chemistry, and automated experimentation into one workflow spanning molecule generation, synthesis design, and reaction/formulation optimization, Jia Haojun, co-founder and CEO, said in an interview during the World Economic Forum’s Annual Meeting of the New Champions in Dalian, China, which took place June 23-25.
“Normally it takes five to seven years to develop a new material. With our approach we can do it within three months,” says Jia.
Deep Principle not only helps corporate clients to develop materials faster: it can also significantly increase the productivity of a product. In the case of one chemical reaction pathway, Deep Principle was able to increase the productivity of a fine chemical from 72% to almost 90%, he says.
The company already has multiple large corporate customers, including L’Oréal. After Deep Principle won a Big Bang award in 2025, a prize that recognizes creativity and innovation, the Chinese scale-up and L’Oréal began exploring ways to use Deep Principle’s Mira as an enterprise solution and are currently co-developing new active ingredients in cosmetics, says Jia.
“We are currently focusing on specialty chemicals, cosmetics and personal care, and renewable energy for batteries, and will later expand to other futuristic industries such as superconducting and fusion,” he says.
A Rapidly Growing Field
Deep Principle is among seven AI for materials discovery companies in the Forum’s current crop of active technology pioneers, says Michelle Mormont, the Forum’s Lead, Innovator Communities – Technology Pioneers. The others are: CuspAI, which in June raised a $400 million round that more than quadrupled the two-year-old AI developer’s valuation to $2.6 billion; Atinary, MatNex, Dunia, Pheno Innovations, and Albert. Collectively, this category of startups has raised over $1.3 billion in the past two years, according to Pitchbook.
“This year we selected more AI-for-materials companies than ever in our Technology Pioneers Community cohort, as their potential for exponential, positive impact is enormous,” says Mormont.
Why Breakthroughs in Advanced Materials Are Needed
As global competition for finite resources intensifies, and the urgency to address climate challenges grows, breakthroughs in advanced materials are expected to be foundational for delivering cleaner, more resilient technologies. That is why the Forum has launched a program dedicated to improving the sustainability, accessibility, and productivity of materials by bringing together industry, innovators, and policymakers, says Fernando Gómez, the program’s head. (See The Innovator’s separate interview with Gómez.)
Traditional approaches to materials development — primarily driven by experience, intuition, and trial-and-error methodologies — are increasingly insufficient to meet current challenges. Materials scientists and engineers today face several critical hurdles, according to an article published on the Forum’s website. These include:
- Identifying promising target molecules from tens of millions of possible structures.
- Designing economically efficient and environmentally sustainable synthesis pathways.
- Optimizing product formulations to meet complex market requirements.
- Scaling engineering processes effectively from laboratory to industrial production.
This is where ‘AI for science’ offers transformative potential, says Jia. Advanced machine learning models trained on extensive datasets, combined with high-throughput computational methods, can predict properties for numerous substances in minimal time, rapidly screening candidate materials against desired parameters.
Beyond models that predict material properties, generative models can directly design molecular structures based on desired properties, creating the materials and molecules needed. Deep Principle‘s latest generative model, ReactGen, for example, can propose novel and complex chemical reaction pathways by learning underlying reaction principles, enabling efficient and innovative synthesis route discovery, says Jia. These models can also generate and iteratively explore diverse synthesis pathways, incorporating chemical information and physical constraints to recommend feasible reaction routes. In product formulation, AI enables multi-objective optimization to meet complex and varying market requirements precisely, saving significant human capital, material resources, and development time, he says.
An Increasingly Crowded Field
Global technology leaders, from Microsoft and Google to Lawrence Berkeley National Laboratory, have launched initiatives such as MatterGen and GNOME that use AI to vastly augment the scale and precision of materials research.
In parallel, emerging players like Deep Principle are attracting top talent. The company was founded in late 2023 by a team that has roots in MIT. Jia earned his PhD in physical/computational chemistry at MIT, where he first conceived the idea behind Deep Principle while researching AI’s role in predicting chemical reactions. He previously worked in Dow Chemical’s core R&D division on AI-driven catalyst formulation. Co-founder Chenru Duan, the company’s CTO, is also an MIT PhD. He previously worked as a research scientist at Microsoft (Quantum) and briefly at Nvidia. Heather Kulik, Deep Principle’s chief scientist, was Jia and Chenru’s PhD advisor at MIT. Luyang Zhang, Deep Principle’s COO, is also an MIT alumnus who previously worked on autonomous driving and HPC. The wider team is drawn from MIT, Stanford, Harvard, Tsinghua, Zhejiang University, Shanghai Jiao Tong, and Fudan, plus alumni of Microsoft, Meta, Dow Chemical, BASF, Saint-Gobain, JD.com, and Tencent.
The company says it has so far raised a total of over $100 million in private capital.
One of Deep Principle’s chief competitors is Cambridge, England-based CuspAI, which has raised venture capital from Bezos Expeditions, the Amazon founder’s family office, as well as venerable Silicon Valley venture firm Kleiner Perkins. Advisors include AI pioneers Geoffrey Hinton and Yann LeCun, former vice president and chief AI scientist at Meta and founder of Paris-based startup AMI Labs (Advanced Machine Intelligence Labs).
Other competitors include: Periodic Labs, Lila Sciences, Edison Scientific, and Orbital Industries.
Deep Principle’s differentiators include:
- Full-chain, closed-loop integration (the “ECML” framework): Rather than treating experiment, computation, and machine learning as separate stages, Deep Principle’s pitch is a single AI-driven decision framework that schedules and balances all three, aiming to avoid the weaknesses of using any one method alone.
- Peer-reviewed, cover-featured generative models: The company’s diffusion-generation model OA-ReactDiff (2023) and its successor React-OT (2025) were published as cover papers in Nature Computational Science and Nature Machine Intelligence, respectively. Deep Principle states React-OT predicts transition-state structures in roughly 0.4 seconds on a single GPU — down from days or months for traditional quantum-chemistry calculations — with over 25% lower error than its predecessor.
- Dual generative architecture: The company runs diffusion-model and large-language-model generative approaches in parallel/complementary tracks, which it says lays the groundwork for agentic delivery.
- The first AI scientist platform for chemistry/materials: Its AI scientist platform Mira launched last year and is the recipient of a 2026 WAIC SAIL award, the most prestigious honor given by the World Artificial Intelligence Conference. (Anthropic just released a similar product called Claude Science.)
IP Ambiguity
New technologies require new business models, says Jia. Deep Principle works with corporate clients in two ways. Clients can use Deep Principle’s platform to get an AI agent to help with their R&D, or the scale-up can co-develop new materials with the corporate partner, in which case the client pays up front and then shares royalties after the material goes to market.
But, as AI materials and AI drug discovery companies are finding out, who owns a material co-developed by a startup and a large customer is an unsettled area of law and contracting practice.
Commercial and technology lawyers structure these collaborations around two categories:
- Background IP — IP owned or controlled by either party before the collaboration began (e.g., a startup’s pre-existing AI models, algorithms, and training data; a customer’s pre-existing proprietary molecule libraries or manufacturing know-how). This generally stays with whoever brought it.
- Foreground IP (also called “Developed IP”) — new IP created during the collaboration, such as a newly discovered material, formulation, or compound. Ownership of this is where the negotiation — and the risk of dispute — concentrates. Legal guidance identifies three common models: (a) the customer owns all foreground IP outright (most protective for the customer, and often the default customers push for since they are paying for the work); (b) the AI/materials company retains ownership and grants the customer a license (more protective of the startup’s ability to reuse the innovation with other customers or in its own pipeline); or (c) joint ownership, which lawyers widely describe as the worst outcome for both sides because of the practical complications of co-owned patents (e.g., needing mutual consent to license, litigate, or exploit).
On top of the ordinary background/foreground question, AI-assisted discovery raises two additional, less settled issues that legal commentators are actively debating:
- Patentability/inventorship: Patent offices in the US, UK, and EU (and the EPO) have consistently ruled that an AI system cannot be named as an inventor — a human must make a “significant contribution” for an AI-assisted invention to be patentable at all. This matters directly for materials discovery: if an AI platform proposes a novel compound with minimal human intervention, the resulting “invention” could in principle be unpatentable by anyone, which would be commercially disastrous for a company relying on patent protection to monetize a new material.
- Whose data trained the model that made the discovery: Because these platforms are frequently trained or fine-tuned using data supplied by industrial partners, a live question is whether a customer’s proprietary formulation data used to fine-tune a shared model creates a claim not just over the specific output for that customer, but over improvements to the underlying model itself — improvements the startup might want to reuse with other, potentially competing, customers. Legal commentary increasingly recommends explicit contract language barring a vendor from using one customer’s proprietary data to improve a model that is then sold to competitors.
“These companies are navigating a new business model and IP landscape, separating rights to the materials they discover from the methods behind them, and questioning whether to invest in fully closed-loop systems, where a company owns the entire pipeline from discovery to physical production,” says Mormont, the head of the Forum’s Technology Pioneers program. “There’s a real opportunity for policymakers, industry, and startups to shape the future of AI for materials science together.”
The Forum can play an important role as a facilitator, says Gómez. “Given our experience with the C4IRs [Fourth Industrial Revolution Centres] in informing policy and regulation for fast-paced technologies, the fact that we convene across all stakeholder groups, and that we are an impartial platform, the Forum will explore ways to work alongside innovators, experts, and corporates to support them as they navigate these challenges,” he says.
