Starting with a core team of just two full-time members, Suyuan Matrix has secured tens of millions of yuan in orders and now serves six major materials companies in just over a year. In the cement industry, its AI system has achieved closed-loop control over key production processes at leading building materials firms, with a single production line generating an estimated annualized direct economic benefit of over 12 million yuan. More importantly, AI is no longer just offering recommendations; it has entered the decision-making loop of real industrial production.

Recently, Na Rongyu, co-founder and COO of Suyuan Matrix, discussed how this AI for Materials company connects capabilities that were previously scattered across laboratories, software teams, and factory floors to bridge the "last mile" from fundamental principles and models to industrial value. In his view, AI's most profound impact on deep-tech entrepreneurship isn't efficiency, but the entrepreneurial path itself. "AI redefines the boundaries of an individual, and the value boundaries achievable through human-AI collaboration," Na stated.

Building an Industry's Cognitive Map in Three Days

During the interview, Na revealed a surprising figure: in its first year, Suyuan Matrix secured nearly 30 million yuan in orders with a team of just 10 people, having started with only two. How do two people support the business volume of serving six large materials companies? Na explained the team's "hexagonal matrix" of six core roles: AI4S (AI for Science), materials/chemistry, process engineering, physics/mathematics, computer science/automation, and product/project management.

"The initial two-person team didn't do less work; they used AI to compress multidisciplinary capabilities into a smaller team," Na said. The team later grew to six, with some members spanning three to four disciplines while others focused on a single area. These six expert roles are supported by AI-driven functions, including AI-assisted product delivery, visualization, and automated R&D—where AI itself proposes hypotheses, synthesizes materials, validates them, and builds a high-content closed-loop laboratory. The physics, mathematics, and computer science capabilities form a stable "general-purpose engine" for the team. When entering different industries, vertical domain experts in cement, chemicals, and metallurgy are brought in. "Industry experts fill in the on-site knowledge that public data lacks—for instance, why data might be under-recorded or mis-recorded, which values are influenced by operational habits, and which metrics can't be directly compared. This experience is crucial for the model to correctly assess data credibility," Na elaborated.

Suyuan Matrix chose cement as its first application scenario, not because of prior industry experience, but precisely because they had none. "Cement is primarily based on silicates and inorganic non-metallic materials, sharing some fundamental physicochemical laws with ceramics and glass. This can further extend to thin-film and interface challenges in semiconductor manufacturing," Na explained. "If AI can't understand a basic materials production line with significant raw material fluctuations and complex operating conditions, it can't truly claim to understand the materials world. We aim to start with cement and progress toward more sophisticated material systems."

To enter an unfamiliar industry, Suyuan Matrix uses a "three-day deep dive" methodology. The team forms a learning group with clear divisions: one person researches the upstream and downstream supply chain and cost relationships; another studies production mechanisms, technical routes, equipment differences, and personnel structures—including each role's KPIs, concerns, and industry jargon; another focuses on modeling core reaction mechanisms into simulators, predictors, and controllers; and another investigates industry regulations. These three days aren't meant to replace industry experience but to quickly build a problem map for validation by experts and on-site teams. Using large language models and custom-built intelligent agents, the team completes an initial structured analysis to form a foundational knowledge framework, which is then refined through expert interviews and real-world data.

Na noted that a significant amount of common industry information is publicly available. "Research paper topics often reflect industry problems, and topics discussed by industry associations provide important clues. AI helps retrieve and summarize this information, which we then cross-validate with papers and expert interviews." What truly requires on-site presence is each production line's unique equipment history, raw material fluctuations, data quality, and operational habits.

From Offering Recommendations to Entering the Production Loop

When this approach is brought to clients, the young team's initial impression is quickly replaced by their ability to delve into process details and discuss the quality, cost, and production stability that clients care about most. "Clients are initially surprised by our youth, but once discussions move into process mechanisms, on-site constraints, and value goals, we quickly shift to a shared professional language," Na said.

This rapid learning ability enables efficient expansion into new industries. Na revealed that the cycle from starting in a new field to securing a contract is about six months—significantly faster than the typical two-year cycle for traditional new technologies in industrial settings. "The bottleneck is often the client's time, as they aren't available daily. If they could dedicate time to us daily, we're confident we could close a deal in a month," he said.

Given the cement industry's high sensitivity to production safety, Suyuan Matrix launched its first project as a joint validation effort, with the first client not paying any fees. "The client took on some production safety risks by being the first to try, and that was their contribution," Na explained. For new technologies, the first client—especially one in a key position—is typically cultivated as a "co-built certification" benchmark.

The implementation was rolled out in four phases. Initially, the system offered recommendations: clients manually input data, AI simulated and provided production decisions—such as raw mix, burning, and control plans—which were then reviewed by humans before execution. After execution, actual and predicted values were compared, and the team issued validation reports. By the second month, the client opened their data interface, eliminating manual input. By the fifth month, after six months of successful validation, the client said, "We don't need to review anymore; you close the loop, analyze, and issue commands directly," as all previous recommendations had been adopted without error. The final phase achieved full closed-loop control: AI automatically analyzes, predicts, and sends instructions to the DCS execution system, then iterates for the next prediction and execution cycle. Client staff now only monitor via large screens. "Previously, four people were needed to monitor; now one is sufficient," Na said.

Na compares the system's maturity to autonomous driving levels L1 through L5. "Depending on a factory's digital foundation, data quality, and on-site management, different plants may be at different stages. Some of our scenarios are close to L4, where the system operates autonomously with staff monitoring and handling anomalies." For safety redundancy, the team uses mature engineering methods: gray releases, canary deployments, and gradual scaling. The model's capability is rolled out in stages—from 10% to 30% to 60% and 80%—with immediate rollback if issues arise. Real-time analysis for supervisory review provides a second layer of protection.

Unlocking Existing Value and Creating New Materials

Na describes Suyuan Matrix's business as a closed-loop materials value chain: "unlocking value" from existing production lines on one end, and "creating value" through new materials on the other. On unlocking value, he emphasizes, "We're not just adding a dashboard to the factory; we're making AI truly understand the material being produced." The cement and concrete projects fall into this category. Even in highly automated industries, raw material fluctuations, equipment conditions, and process coupling can lead to inconsistent material properties. Traditional technical expertise and parameter control struggle to capture these changes in real time. Suyuan Matrix aims to shift control from "equipment parameters" to "material state," improving stability, quality, and cost efficiency within safety boundaries.

On creating value, he states, "Discovering a material is only the first half; manufacturing it stably and scaling it to mass production is the second half." Suyuan Matrix uses AI to design new materials, connecting formulation, experimental validation, pilot-scale testing, and mass production. For example, the team is developing radiation-protection materials with research institutions and exploring thin-film deposition materials and process designs for high-aspect-ratio structures in semiconductors. Na reveals that the marginal efficiency of "unlocking value" is currently very high due to the similar reaction mechanisms across materials industries. The team is increasingly shifting focus toward "creating new materials," which he believes showcases technical strength, generates higher profits, and captures a larger share of the future new materials market.

In the semiconductor sector, for instance, the AI system is used to design deposition film materials with an aspect ratio of 60:1 (depth 60, width 1), requiring stable atomic growth in extremely deep trenches. Na calls this "atomic-level manufacturing": controlling layer-by-layer atomic growth for better uniformity and precision. "Better product consistency and yield are hallmarks of the most advanced processes," he said.

Regarding the company's moat, Na emphasizes that the true barrier isn't a single model but a stack of three capabilities. The first layer is the underlying model. Suyuan Matrix's focus—scientific intelligence—is one of AI's three major directions, alongside large language models and embodied intelligence. He notes that a key origin of scientific intelligence is the Lawrence Berkeley National Laboratory, with global leaders including U.S. national labs and MIT. Core team members of Suyuan Matrix hail from these important research institutions.

The second layer is the engineering capability to take models to the field, and the iteration speed that results. This requires not only model accuracy but also understanding process constraints, integrating with systems like DCS, APC, MES, and LIMS, and using gray releases, human review, and rollback mechanisms to safely turn R&D results into operational products. The third layer is the speed of innovation and iteration. Na calls this "maximizing R&D and innovation within limited time and resources, because entrepreneurship isn't about polishing details locally or jumping aimlessly, but about expanding the exploration stride while controlling risk, using successive validations to approach a better solution."

The underlying logic comes from an AI algorithm principle: stochastic gradient descent. "It's like being in a high-dimensional mathematical space where you're at the highest point and the optimal solution is at the lowest, but the space is full of bumps," he explains. "This tells you not to make tiny, localized steps or jump around randomly, but to take slightly larger strides, try, adjust, and repeat." Na translates this into a management philosophy: "Step out of your comfort zone, explore new technologies, summarize and communicate, and continuously do these things."

Discussing the company's positioning, Na sets a clear goal: "We want to become the AlphaFold of materials." This isn't about benchmarking a single model but about reconstructing how material value is created. "We can provide production optimization systems for factories, assist in analysis, prediction, and decision-making; participate in new material design and commercialize through licensing or industrial partnerships; and drive traditional materials factories toward becoming AI-native factories."

In his view, no single company currently covers all these areas. "Everyone has their own strengths and weaknesses. We recognized this from the start and are using AI to truly integrate these capabilities while taking responsibility for whether materials ultimately generate industrial value." Na uses a comparison to illustrate this integration: traditional materials research institutes excel in processes, materials, and computation; industrial internet teams excel in software and products; university labs excel in physics, materials, and AI. But these capabilities have long been scattered across different organizations and departments, lacking a unified goal centered on final material value. "What we need to do is truly connect the full spectrum of capabilities required to take a material from discovery to mass production." He specifically highlights two roles within the team: forward-deployed scientists and forward-deployed engineers, who "bring cutting-edge AI4S talent to the most complex real-world production lines, completing research, validation, and delivery on-site, so the most advanced technology confronts the most real problems."

AI, Making Expertise a Product Faster

Na believes that the fundamental reason for such a lightweight organizational structure in the AI era is a qualitative change in information access and collaboration. "In the past, learning something new required finding many experts, who weren't always patient in explaining. In the AI era, what you don't know, AI knows, and AI helps you expand." He positions AI as the team's "second brain" and "second avatar." The team maintains a central document, and members can query AI to check for deviations in their understanding, asking for three possibilities, quickly correcting cognitive biases. "AI offers powerful, 24/7 intelligence. If used well, it's like having Einstein on hand; if not, it's just a smart person," Na said.

Reflecting on the entrepreneurial journey, Na believes AI's greatest change to deep-tech startups isn't efficiency itself, but the path—it changes how expertise becomes a product. "In the past, the scarcity was in knowing how to do something," he says. "But today, it's more about rapid execution, rapid trial-and-error, and rapid translation and understanding of needs to turn client problems into executable, solvable solutions. AI hasn't replaced expertise, but it allows us to quickly turn our expertise into products and take them to the world."

Now, over a year since its founding, Suyuan Matrix serves six large materials companies, has secured tens of millions of yuan in actual orders, and expects to cover more than 20 production lines this year. The team has grown from 2 to 10 people, but the underlying logic remains unchanged: use scientific intelligence to understand the materials world, use AI tools to accelerate delivery from model to scenario, and leverage the smallest organization to unlock the greatest value. "AI truly enables a team or individual with deep professional barriers to cross from research and mechanisms to the 'last mile' of model-to-scenario delivery quickly," Na concluded.