Transforming the Industrial Landscape of Chemical Materials with AI | Linear Capital Portfolio Interview Series: "Deep Principle Technology"

Today, Deep Principle, a company focused on AI-driven scientific research in chemistry, announced the completion of a nearly $10 million seed round led by Linear Capital, with participation from Zhenzhi Venture Capital and Taihill Venture. XtalPi and DP Technology also joined as strategic industry investors. Founded in 2024, Deep Principle's founding team members are all MIT graduates. The company's vision is to apply artificial intelligence, quantum chemistry, and high-throughput experimental technology to the field of chemical materials, improving the workflow of materials innovation and accelerating the efficiency of R&D in chemical materials. Recently, Linear Capital Investment Director Tianyi Zhou sat down with Deep Principle co-founders Haojun Jia and Chenru Duan for an in-depth discussion about their entrepreneurial journey and perspectives on the industry.

Linear Capital Portfolio Interview Series is Linear Capital's interview column, featuring regular deep-dive conversations with founders of portfolio companies about their entrepreneurial stories and industry viewpoints, sharing the stories of tech investment and tech entrepreneurship with a broader audience. For more interview content, please click "Read More" at the end of this article.

Tianyi Zhou: Please briefly introduce Deep Principle.

Haojun Jia: Our company is called Deep Principle, which signifies the combination of Deep Learning and First Principle thinking. It also represents our commitment to using AI to deconstruct the operating principles and laws of the molecular world.

In terms of specific business, as an AI for Science startup, we are dedicated to optimizing and revolutionizing chemical materials through AI to significantly improve energy conversion efficiency and catalyze the emergence of new productive forces. The origin of this endeavor is deeply connected to our experiences and backgrounds. My partner Chenru Duan was among the earliest PhDs in the AI for Science field. Regarding entrepreneurial ideas, I actually started thinking about this slightly earlier, around two years ago, focusing on AI for Science-related ventures—this was also my research direction from undergraduate through PhD studies. I observed that within just a few years, with the emergence of various AI models, computational speed and accuracy have improved dramatically. Previously, computational costs were more tied to time factors, but now calculations that might have taken months can be completed in seconds. Under such circumstances, quantitative changes have led to qualitative transformations. I believe that in future industrial materials discovery, AI and computation-driven approaches will gradually replace experimental trial and error. More people will use AI and computation to discover materials, followed by wet lab validation, which will massively improve the overall R&D efficiency in materials chemistry.

Chenru Duan: I previously participated in some AI for Science community organizations and machine learning conferences, and I was deeply impressed by the rapid development of the entire AI for Science field. The most intuitive signal was attendance numbers. In 2021, our NeurIPS workshop had about 150 participants; in 2022, it was 500; and by December 2023, over a thousand people attended. The growth of the AI for Science community has brought in fresh perspectives, with more researchers joining, expanding our view beyond just computer science (CS) to more scientific research angles, using AI as a tool to solve problems. However, we also clearly felt that there remains a significant gap between academic research and practical industrial applications.

In my previous two years at Microsoft, my work involved integrating chemical computation tools—specifically first-principles calculation tools—with machine learning tools into a platform, delivered as Platform as a Service (PaaS). We also developed a relatively well-known product in the industry. Through this experience, I realized that in the chemical materials field, there's a substantial generational gap between R&D departments at end-user applications and the most advanced machine learning/computational technologies in academia. This creates a problem: if a platform wants to drive adoption among major chemical materials companies, it needs to invest enormous effort in user interfaces—essentially spending time on engineering and packaging. Meanwhile, because AI for Science is such a new field with increasing disciplinary depth, the difficulty of balancing both aspects in platform development becomes extremely high. Therefore, Haojun and I shared the same thinking: given our scientific research backgrounds, what we want to do is take AI for Science deeper, rather than simply doing platform integration. This also relates to Deep Principle's entire product roadmap and business model, which was the catalyst for my decision to start this company.

Tianyi Zhou: How did you two meet, and how did you evolve from classmates to business partners? I heard that initially, Chenru wanted to pursue an academic or industry career, while Haojun's original entrepreneurial direction was different?

Haojun Jia: We were both under the same advisor, Prof. Heather Kulik, working in the same lab. The first person I formally arranged to meet and chat with when I arrived at MIT was Chenru. I had a great impression—he was genuinely nice, passionate about research, and had strong insights. Later, we worked together in Prof. Heather Kulik's lab for a long time, essentially throughout my entire PhD.

I did initially want to pursue a different entrepreneurial direction before eventually focusing on AI + catalysis or AI + chemical materials. At first, Chenru didn't join because I knew he was looking for a faculty position. Once both of our plans aligned around AI + chemical materials entrepreneurship, we immediately clicked and started the company together.

I believe the most important elements of co-founding a company are, first, trust, and second, complementarity. In the early stages with few people, you need tremendous trust to make things work. Chenru and I had worked together for several years and had very strong understanding of each other's capabilities and character. Second, although we were in the same group, our research directions differed—I mainly worked on quantum chemistry calculations, while Chenru focused on AI and chemical algorithms. Additionally, in terms of personality, I lean toward organization and coordination, while he has strong academic leadership.

Chenru Duan: The shift in my own mindset primarily came from my time at Microsoft. I saw how what I built could be useful in industry, though the barriers to adoption were somewhat high. This pushed me toward the idea of digging deeper into this industry, and I sensed that the timing was right to deeply explore AI for Science applications in chemical materials. Entrepreneurship, in some ways, aligns with my original intention of pursuing academia. What I love is the process of research and problem-solving, and a startup is an excellent vehicle for that. Additionally, engaging with chemical materials companies in industry, seeing their changes and needs, revealed this window of opportunity to me.

Tianyi Zhou: When you disagree, how do you resolve it?

Haojun Jia: Honestly, conflicts and even falling out among founders are quite common. We discussed a series of solutions for handling disputes in our early days. But having known each other for so many years before and after starting the company, I don't think we've ever even come close to getting angry or upset with each other. I believe arguments often stem from differences in perception, which come from thinking patterns and information processing. Both of us have STEM backgrounds and emphasize logic, so we believe that when we disagree on something, it's most likely due to information gaps. My personal solution is to communicate more, get everyone's mindset aligned, and that resolves the vast majority of issues.

Tianyi Zhou: You should be among the first batch of startups putting GenAI into chemical workflows. What are your thoughts on this idea, and what outcomes are you expecting?

Chenru Duan: At that time, people had already applied Generative AI, particularly Diffusion Models, to some pure small molecule generation and biology applications. But these only involved generating a single molecule. When I was organizing the AI for Science workshop in 2022, I discussed this with a colleague and realized that nobody seemed to be going in the direction that chemistry truly cares about. The reason, as Haojun mentioned, is that computer science has a tendency to chase hot topics much more than other fields. Biology has been heavily worked on in AI, so many people went in that direction, while in chemistry, almost nobody was doing it.

We then asked ourselves: What are the most important problems in chemistry? Essentially, first, all our materials—from the titanium alloy casing of phones to plastic water cups—are stable materials. Stability in chemistry is defined as a minimum point on a potential energy surface. The process of chemical research, or chemical reactions, is essentially a process from one minimum point to another. So we wondered how to use diffusion models to generate chemical reactions. Of course, there were numerous technical challenges, especially since previous Diffusion Models and graph neural networks couldn't guarantee certain symmetries unique to chemical reactions. We made significant improvements in this area, developing for the first time a graph neural network that ensures the symmetry of entire chemical reactions. Second, we integrated this into a Diffusion Model architecture to generate complete chemical reactions. Finally, we focused on transition state search, improving its speed by approximately 1000x—from originally taking hours or even more than a day to under ten seconds.

Initially, I thought this was something worth doing, but I didn't expect the results to be this good or the accuracy this high. The transition states we currently generate can basically be compared with experimental reaction rates. Although the reaction rates still differ by an order of magnitude, it has already shown us the possibility of "computation and AI guiding experiments."

Tianyi Zhou: In Deep Principle's field of entrepreneurship, what is the general landscape abroad and domestically? Is the domestic scene mainly university professors? Does abroad have more prominent players like Microsoft? In this regard, does what Deep Principle is doing have some gap-filling significance when brought back to China?

Haojun Jia: The international landscape was initially shaped by some large tech companies, with quite a few doing well. The first would be Microsoft. Microsoft has approximately four AI for Science research institutes worldwide, with about 200-300 people across several locations in Europe and Beijing. On the Google side, there's DeepMind. Before the merger of DeepMind and Google Brain, each had several groups working on this, particularly in drug and materials discovery. Nvidia also has two groups, one working on AI for drug discovery and another on AI for materials discovery. Nvidia's approach stems from their conviction that scientific computing is an important future direction, developing algorithm platforms to further advance GPU R&D—essentially creating GPUs adapted for scientific computing. Meta has a group called Open Catalyst, which may be more relevant to what we do. These are the situations at several major companies.

On the startup side, the most directly relevant is Isomorphic Labs, which spun out from DeepMind, and another heavily funded company called SandBox AQ, backed by former Google CEO Eric Schmidt, which raised $500 million in its first funding round. In China, earlier there was a series of companies doing AI for drug discovery. Additionally, at the technical level in the AI for Science field, DP Technology has done quite well.

Chenru Duan: Let me add a few points. The AI for Science field has had quite a few major news items in recent months. For example, Isomorphic Labs recently reached $3 billion in partnerships with Novartis and Eli Lilly. For a startup to collaborate on joint R&D with major pharmaceutical companies is very exciting. This indicates that large companies are considering shrinking their own R&D and opening up more R&D resources and opportunities to small startups for collaboration. This is positive for the economic efficiency of large companies and also improves the growth environment for startups.

Additionally, in biology, Xaira Therapeutics was founded by National Academy member and University of Washington professor David Baker along with the former Stanford president, focused on biology and using Generative AI for more complex protein drug design. In the current market, their first funding round was publicly disclosed as $1 billion. This funding amount is unprecedented in biotech and pharma entrepreneurship, and represents a statement from US VCs about this entire direction.

There are also a few startups similar to ours emerging, with slightly different directions, indicating that everyone is optimistic about this field.

Tianyi Zhou: From a technical perspective, what difficulties have you overcome since starting the company?

Chenru Duan: Many people have commented that quite a few people doing AI for Science just take whatever new model comes out and copy it to apply to chemistry problems. I think there's some truth to that, because doing AI for Science is fundamentally not about advocating for model and architecture design. Currently, the initial applications of Transformer and Diffusion Model are in CV or NLP fields, because computer science (CS) people are the ones doing AI method development, while AI for Science itself is not positioned as AI method development.

Especially in chemistry, the symmetry of the problems we deal with is completely different from theirs. For example, neither NLP nor CV needs to consider symmetry at all. A cat, whether its photo is right-side up or upside down, is still the same cat. But in computer science, symmetry is generally not directly built into model architecture. Instead, they rotate the cat repeatedly and put it in the training set, letting the model learn through massively increased training that the right-side-up cat and upside-down cat are the same.

When we work with 3D molecules, because many problems in chemistry and materials are three-dimensional, and the granularity of results we need is much finer—for example, if an atom is carbon, it must be carbon; it cannot be nitrogen, even though carbon and nitrogen differ by only one element. Or if a carbon-hydrogen bond is 1.1 angstroms, a prediction of 1.0 angstroms is wrong. Therefore, the demand for precision in turn requires us to consider symmetry when designing models. This is what I see as a characteristic and also a difficulty in AI for Science technology, one that requires continuous攻克.

Tianyi Zhou: For startups, overcoming challenges one by one has a positive effect on team building. You just discussed technical aspects; on the operations side, can you share a recent positive feedback?

Haojun Jia: Early recruitment is extremely important—you need excellent talent to build a creative team. Let me share a recruiting example. Everyone knows that hiring for early teams is very difficult. Some entrepreneur friends even told me to be mentally prepared, that it would be tough. To convince someone to join, they first have to be willing to forego short-term benefits and accept greater work intensity. One of my principles is that the core team must be stronger than me in some area—whether it's a technical branch, BD, or operations, they must be stronger than me in some domain. I've always believed that the best way to attract A-level talent is to invite them to work alongside a series of A-level talents and move forward together. Therefore, we've invested tremendous effort in recruiting, maintaining very high standards in this area without ever compromising. The recruiting post we published may look like just an article, but the thinking and discussion behind it was actually quite extensive. The results—from article views to resumes received to people eventually hired—have so far exceeded our expectations. We are currently completing this round of recruiting and welcome anyone interested to reach out to us anytime.

Tianyi Zhou: If everything goes smoothly, what kind of company do you want to build in the future?

Chenru Duan: If everything goes well, broadly speaking, I hope to change the existing industrial landscape of chemical materials, shifting toward a model more similar to biotech and pharma, where chemical materials companies delegate more materials innovation to specialized materials innovation companies. This model is only possible when our R&D capabilities become strong enough. The core is having an internal chemical materials R&D platform that can continuously generate a pipeline of innovations that currently cannot be industrialized but would bring tremendous value once commercialized. These pipelines, after laboratory and preliminary industrial validation, can lead to deep partnerships with chemical materials companies. What we ultimately want to build is not a Software as a Service or Platform as a Service company. We hope to directly transform the landscape of chemical materials R&D through current AI for Science methodologies.

Haojun Jia: Materials chemistry and biotechnology are almost equally large fields, but due to various issues, materials chemistry has never developed to the same degree of specialization as biotechnology. However, due to opportunities brought by AI technology, I believe a new paradigm is likely to emerge, using AI or materials computation and materials discovery to drive transformative changes in the overall R&D landscape.

Chenru Duan: Traditional R&D has always been done by large companies. Biotechnology recognized earlier that its most important aspects are distribution, productization, control of industrial chains, and R&D and production of new drugs, selling pipelines after cultivation. But this model hasn't been cultivated in materials chemistry. Recently has been a good opportunity: due to the global economic downturn, more and more materials chemistry companies have recognized their massive R&D investments without achieving expected results. This relates to internal models and mindsets, but currently their trend is to gradually reduce pipeline R&D and move toward a biotech/pharma model. This is the opportunity for new companies like us.

Tianyi Zhou: How did you get to know Linear, and what were your impressions during your interactions?

Haojun Jia: When people ask me about my story with Linear, this is how I tell it. Part one: one day I woke up to see several people adding me on WeChat. Among them was someone named Yingzhe Zeng who wrote "Linear Capital" in the request—I thought it was some intern, but upon checking found out he was a partner. At that time, we weren't even planning to start fundraising, but after one conversation, the feeling was excellent. He was very approachable and had strong industry knowledge. I had been a student union president before and had met quite a few fund partners, but this was the first time I met a fund partner who was actually so hands-on with such specific work. I found this quite rare. For early-stage projects, without much financial data or clear products, knowledge and insight are extremely important. Later, meeting Tianyi, I found him to be among the young investors I've met, the one who understands both technology and market best—knowing both AI and chemical materials.

On another note, because we were still abroad at the time and couldn't meet the team offline, we could see Linear's team trying to overcome these difficulties. Whether in communication, driving decision-making, or providing assistance before investment, they put in tremendous effort and gave us great support.

To connect with the Deep Principle team, please follow the "Deep Principle" WeChat official account.

About Linear Capital

Linear Capital is an early-stage investment institution focused on "frontier technology + industry" investments, namely frontier technologies represented by data intelligence, digital new infrastructure, next-generation robotics technology, and new technological transformations in traditional fields (such as biomedicine, materials, energy, etc.), applied across various vertical industries to substantially improve industrial efficiency, empower them to solve pain points, and complete industrial upgrading—achieving excess returns in commercial value through significant increases in industrial value. Currently managing ten funds with total assets under management of approximately $2 billion.

Our investment stage focuses primarily on leading angel to Series A rounds, with individual investments ranging from $1 million to $10 million (or equivalent RMB).

To date, we have invested in over 120 entrepreneurial teams at early stages, including Horizon Robotics, Kujiale, Sensors Data, Tezign, Rokid, Guandata, Agile Robots, and others. The combined valuation of Linear's portfolio companies is approximately $20 billion.

In the near term, Linear Capital is striving to become the best "Data Intelligence Technology Fund," and in the long term, gradually build itself into the most influential "Frontier Technology Application Fund."