MaHui Entrepreneur | DeepMaterial's Xuanze Wang: AI Is Reconstructing Materials R&D, and 'Wish-Based Development' May Soon Be Possible

王轩泽： 大模型负责调度和提出假设，小模型负责确定性执行，这是分工。

大模型生成配方时，有时会有幻觉，会生成一些不符合物理规律的东西。我们用小模型做校验，把不符合的先筛掉，再交给实验验证。这个架构我们内部叫"大模型调度，小模型执行"。

ZP：这套"大模型+小模型+自动化实验"的闭环，具体是怎么运转的？

王轩泽： 我们的流程是：大模型生成配方假设，小模型和计算工具做初筛，然后进自动化实验室做高通量制备和表征，数据回传再训练模型，形成闭环。

一个配方从生成到拿到完整数据，传统实验室可能要几个月，我们现在能压缩到几天。而且数据维度完全不同——传统材料数据是写给人看的，一条记录可能就十几个数值；我们的自动化实验室产出的是720维数据，数据量约10GB一条，累计已有数万条。

ZP：这些高维数据具体包含什么？为什么对人不可读，却对AI至关重要？

王轩泽： 包含成分、工艺参数、微观结构表征、性能测试等全链条信息。比如微观结构，我们用扫描电镜、透射电镜、同步辐射等手段，获取的是图像、谱图、三维重构数据，维度极高，人没法直接看明白，但AI能从中学习规律。

传统材料数据库为什么撑不起AI？因为那些数据是给人做参考的，维度低、噪声大、格式不统一。我们从头设计数据采集流程，让数据天生就是为AI准备的。

03

自建实验室、自产数据，走最重的路换最深的护城河

ZP：2021年就遇到数据瓶颈，当时怎么判断必须自建实验室，而不是找外部合作？

王轩泽： 我们试过合作，但发现不行。

第一，外部数据格式不统一，清洗成本极高，而且很多关键维度缺失。第二，材料研发的核心数据是工艺参数，这属于企业核心机密，不可能共享。第三，数据生成速度跟不上，我们需要的是每天都能产出、能持续迭代的数据流。

所以2022年我们决定自建，现在迭代到3.0版本。实验室完全按我们的数据标准设计，从样品制备到检测表征，全流程自动化，数据直接入库。

ZP：3.0版本相比之前有什么质变？

王轩泽： 通量和稳定性。1.0版本很多环节还要人工干预，2.0实现了全流程自动化，3.0是高通量并行——同时跑多个配方，设备利用率大幅提升。现在一条产线一天能做几十个配方，一年就是上万条数据。

而且我们不止一台设备，是集群化部署。材料制备、热处理、机械加工、表征检测，每个环节都有自动化单元，用调度系统串联。

ZP：这套系统的投入和运营成本应该很高，怎么说服投资人支持这条"重资产"路线？

王轩泽： 我们找的是真正理解技术本质的投资人。**源码资本从天使轮就投我们，之后Pre-A轮加注，A+轮又联合领投，累计投了三轮。**他们看懂了一件事：材料AI的护城河不在算法，在数据；而数据没有现成来源，必须自己造。

我跟曹毅聊过，他问过一个问题：如果腾讯或者字节进来做这件事，你们怎么办？我说他们缺的不是钱，是时间。材料数据的积累是线性的，一天就是一天，没法跳步。我们现在领先两三年，这个窗口期足够建立壁垒。

ZP：目前的数据积累和材料产出到了什么阶段？

王轩泽： 累计数万条高维数据，研发出近20款金属新材料，部分已经量产。客户包括航空航天、新能源、消费电子等领域的头部企业。

我们的判断是，数据规模到数十万条时，金属材料的GPT时刻就会到来——模型能泛化到全新材料体系，提出人类想不到的配方。时间可能在明后年。

ZP：这个"GPT时刻"具体指什么？材料领域会出现自己的ChatGPT吗？

王轩泽： 不是聊天机器人，是研发范式的质变。

现在的AI还是在人类已知框架内优化，比如我们知道高熵合金这个方向，AI帮我们更快找到好配方。GPT时刻意味着，AI能提出全新的材料体系，人类之前没想过、甚至跟现有理论矛盾，但实验证明它是对的。就像AlphaFold之于蛋白质结构预测，不是更快，是解决了人解决不了的问题。

到那一步，材料研发就从"经验驱动"彻底转向"系统探索"，九成未开发的成分空间会被打开。

04

不做卖水人，要做挖金人：AI材料公司的终局是大型材料企业

ZP：你们从day one就定位自己做材料、卖材料，而不是卖算法或软件服务。这个选择在当下AI for Science创业中似乎很少见。

王轩泽： 这是由行业特性决定的。

材料行业的客户要的是材料本身，不是算法。你卖软件，客户会问：能保证我做出这个材料吗？不能，那我不买单。而且材料一旦量产，客户关系极深，替换成本极高，这是最好的商业模式。

更深层的原因是，**金属材料是数十万亿级的市场，而研发是其中成本最高、周期最长、结果最不确定的环节。**谁解决了研发问题，谁就有机会整合全产业链，成为一家大型材料企业。我们不想只做环节中的工具提供商。

ZP：这个终局设想中，创材深造会成为什么样的公司？

王轩泽： 一家用AI整合研发、生产、销售的材料平台企业。

上游对接矿山和冶炼厂，中游是AI驱动的研发和生产，下游直接服务终端客户。现在我们的模式是，根据客户需求定制材料，收取研发费加材料销售分成。未来随着材料牌号极大丰富，会形成标准产品库，客户像选SKU一样选材料。

ZP：这个路径上最大的风险是什么？

王轩泽： 两件事。一是数据积累的进度，能不能在资金烧完前达到临界规模。二是量产转化的能力，实验室做出来和工厂稳定生产之间，还有很长的工程化路要走。

我们在第二个风险上已经做了布局，跟几家大型制造企业共建了产线，不是纯代工，是联合调试工艺参数，确保从实验室到工厂的平滑过渡。

ZP：如果明后年真的迎来金属材料的GPT时刻，行业格局会发生什么变化？

王轩泽： 材料牌号会爆炸式增长。现在全球注册的金属材料牌号大概几十万种，理论上应该有几千万种。研发瓶颈打开后，这个数字会快速膨胀。

同时会出现平台级企业。传统材料行业非常分散，有做特种钢的、做铝合金的、做钛合金的，各管一摊。AI能跨体系泛化后，一家公司可以覆盖多类材料，规模效应会很强。

ZP：最后一个问题，如果给同样想做AI for Materials的创业者一个建议，你会说什么？

王轩泽： 想清楚你要做"卖水人"还是"挖金人"。卖工具短期轻松，但天花板低、壁垒薄；做材料很重、很慢，但一旦跑通，护城河极深。

另外，团队必须有双重理解——懂AI的不懂材料，会觉得数据随便采；懂材料的不懂AI，不知道什么数据对模型有用。我们团队核心成员要么有AI+材料交叉背景，要么两拨人深度协作。这是最难复制的部分。

---

本文首发于Z Potentials，经授权转载。

Xuanze Wang: We're firm believers that a single large model isn't enough. Large models can only read papers, and the structural data in papers is poor: low dimensionality and poor consistency. Combine data from Paper A and Paper B and you pollute the dataset; switch equipment and the error spikes. There's also the difference between dry and wet methods. Of course, if you use a large model to develop a small model and establish a new theoretical framework — that's exactly our approach.

The human body is the most intelligent system we know, and it works this way. The brain handles overall coordination, but it doesn't directly control the heart pumping blood or the stomach digesting food. If the brain had to manage every contraction of the heart, it would stop beating the moment you fell asleep. The brain only does macro-level regulation — when you need to run, it tells the heart to beat faster.

Materials R&D works the same way. Many vertical small models and accurate computational tools outperform large models on specific problems because small models have determinism, while large models inevitably hallucinate. So we use a framework where the large model serves as the brain and small models serve as the organs.

ZP: In this framework, what new roles does the large model specifically take on? How far along is it now?

Xuanze Wang: There are two core functions. First is overall orchestration — previously done by humans, now delegated to the large model, beyond just reading literature. Second is RSI, recursive self-improvement. We're still some distance from true RSI, but we can already have the large model generate hypotheses and validate them through our dry and wet experiments. I don't just want materials to keep getting better; I want to distill new theoretical frameworks in the process. Validated hypotheses get accumulated, continuously improving R&D capability itself.

As for how far along we are, I'll only speak to our own measurements. Currently, materials scientists using our small models for R&D still outperform large models doing the calling. But we've recently invested heavily in large models, even planning to train our own vertical large model. Because we're seeing that within the agent architecture, language models' causal reasoning and logical capabilities are deterministically and rapidly strengthening. It likely won't be long before large models calling small models and generating theoretical hypotheses outperform materials experts. The results will show up as fewer experiments, better material properties, and broader search boundaries.

ZP: Across the full R&D workflow, which steps have been engineered, and which still depend on veteran expertise?

Xuanze Wang: At this stage, many steps still rely on experienced hands. Before R&D begins, you need to bound the problem — there are too many elements to tune each one; clients have requirements too, wanting aluminum alloys, light weight, or nickel-based. With eight elements to co-optimize, in a search space of 10^8, how do you set boundaries and place sampling points? Much of this depends on experience. Uniform distribution works, but we have strong metals R&D people who are already well-adapted to AI-assisted R&D. Having them frame the boundaries speeds things up considerably.

In the future, large models can take this over. But I want to emphasize: the bottleneck isn't large model capability, it's data — we're severely short on data right now.

---

03

The bottleneck in materials R&D has never been algorithms, but data

ZP: You've repeatedly emphasized data as the biggest constraint. What's wrong with existing real-world data — factory production data, paper data, public databases?

Xuanze Wang: Data is the biggest bottleneck across the entire AI for Science industry, whether for government, research institutes, or startups — everyone is talking about this. There are two types of data: simulation data and real experimental data. Let me start with real data, which has three main sources: production companies like state-owned steel mills, academic papers, and foreign databases — foreign ones are significantly better than domestic, though some people have been scraping them lately.

All these sources share a common problem: they're written for human researchers, not for AI, so the dimensionality is extremely low. One data point is a few tenths of a kilobyte, just a dozen or two numbers — composition, process, process parameters, corresponding properties. Humans can read this fine, but AI trying to understand physical changes in the macro world from so few numbers won't converge. Simple data distributions that can be drawn with one stroke can converge; add more elements, more coupling complexity, or switch to new materials, and it won't converge.

Each source has its own problems too. Production data: companies don't like sharing it, and the composition and process variation is small, so its R&D value is limited. Paper data is the most accessible and easiest to obtain, but has the worst consistency — different research groups, different equipment, dry vs. wet methods produce very different results, with no unified database. Combine a dozen or several dozen papers and you severely pollute the data pool. This is the worst data. Of course, I'm talking here about structured data for training small models; for large models, papers remain the most core training material — they learn scientific thinking and logic, which is a completely different matter. Foreign databases have similar problems, and the volume is only around two thousand entries.

ZP: Can simulation data fill the gap? Why must metals R&D depend on real experimental data?

Xuanze Wang: If the goal is end-to-end development of metals, or macro mechanical performance materials in general, not just publishing a paper, you absolutely need real experimental data.

Because simulation error is too large. DFT calculates at the microscopic level, atomic scale of 10^n, but metals, ceramics, composites, polymers — these macro materials change enormously at the mesoscale. It's a chaotic problem, not a linear superposition of atomic-scale behavior.

Take metals as an example. The biggest source of performance difference isn't what crystal it is, but between grain and grain. When you pull a material until it breaks, the fracture always occurs at grain boundaries; inside the grains it's very dense. Grain boundaries, plus inclusions and dislocations, are among the biggest factors affecting macro performance. Many papers, including those by academicians, have noted this. Studies in authoritative materials science journals like Physical Review Letters show that DFT-calculated mechanical properties typically differ from experimental measurements by several-fold to several dozen-fold. In short: it can't calculate accurately, so we can only rely on real experimental data.

ZP: What kind of data is actually useful for AI? How far have you accumulated?

Xuanze Wang: We identified the data problem back in 2021. The problems the industry is hitting today, we encountered five or six years ago, and started addressing them long ago. Around 2022–2023, we built out our automated lab, dramatically reducing data costs and increasing speed. We're now on version 3.0.

We've added extensive sensors throughout our lab. Each self-produced data point has 720 dimensions and 10 GB, covering the full process of preparation, post-processing, testing, and characterization — process parameters and process variables throughout, and it's multimodal: numerical, image, time-domain, vibration, optical. In the metals field, this kind of high-dimensional, high-consistency data is currently extremely scarce; we were among the earlier ones to systematically pursue this.

We've now accumulated tens of thousands of entries — enough for materials R&D, and nearly 20 materials have been developed this way. But generalization capability is still relatively weak. We estimate we'll need at least several hundred thousand entries for models to approach human-level performance, with clear algorithmic improvement as well. We'll reach that scale very soon. A major portion of this year's funding goes toward deploying more production lines to generate this type of data. Mass-producing high-quality, high-throughput metals data is one of the company's most core priorities.

ZP: Building your own lab to generate data is a very heavy path. Is this the only way to obtain data?

Xuanze Wang: It's the only way. Real data can only be obtained this way.

Data has to come from experiments. Normally one data point costs tens of thousands of RMB; we've reduced costs by several orders of magnitude. The specific numbers are too sensitive to share.

The key is that we use these 10 GB, 720-dimensional intermediate representations and process parameters to connect the few dozen dimensions of composition and process to final properties, so AI can truly learn the mapping from composition and process to performance. And it's all from our own automated experiments, with extremely high consistency.

ZP: You've argued that synthetic data is essentially informationally closed. Where does this judgment come from?

Xuanze Wang: Informational closure holds under specific conditions.

For example, using a deterministic formula to generate data, like DFT, then having a model learn from it. In the ideal case, the model can at best learn to be as good as that formula. AI's role is acceleration — from calculating 10^3 to calculating 10^8.

But our ultimate goal is to develop new materials. Theoretical simulation itself has large error and poor generalization; if the formula is inaccurate, synthetic data generated from it is inaccurate too, and training on it at best gets you to the formula's level. That's theoretically speaking; in practice, some technical methods can create variation.

So synthetic data must be deterministically synthesized. Another case is large model-generated data, especially with human intervention, adjustment, and judgment — that can be useful.

ZP: Language models had their ChatGPT moment, protein structure prediction had its AlphaFold moment. Will metals see a similar breakthrough? Under what conditions, and when?

Xuanze Wang: It will happen, and the condition is simply accumulating enough data. Everyone is short on data right now; with this little data, you can't get such obvious effects. When you have several hundred thousand entries of truly AI-suitable data, it could happen — I expect next year or the year after.

Some might find this timeline aggressive. But what I mean is, once the data is in place, there won't be major algorithmic obstacles. I'm confident saying this because data — the biggest bottleneck — is exactly where we have the most conviction.

The industry may not have consensus on when the GPT moment arrives, but there's consensus that the bottleneck is data. We're already using this data to make significantly better new materials than before, with industrial applications. When data reaches several hundred thousand or millions of entries, the 90% of materials humans haven't explored may quickly be discovered.

---

04

No intermediate deliverables, only mass-producible materials: How four puzzle pieces enable wish-fulfillment R&D

ZP: Among the many directions in AI for Science, why did you choose metals as your landing domain? How much change can AI bring to this field?

Xuanze Wang: Several reasons. First, the market is large enough. Metals is one of the largest materials domains, comparable to or larger than biomedicine — a tens-of-trillions-level market. Second, commercialization is relatively easier. Biomedicine is hard to develop and hard to validate; a new drug takes over a decade to market. For metals, you can test nearly all properties in two weeks; once performance meets spec and batch consistency is there, you can sell. Our fastest material went from receiving the request to mass production in about half a year. Third, my background fits — I have some accumulation of industry resources, industry understanding, and materials science knowledge.

As for the change, I believe it's a paradigm shift. We can work with any metal. Among nearly 20 materials, we have aluminum, titanium, steel, nickel, high-entropy alloys, copper — including high-entropy alloys humans haven't explored before, with dramatically reduced costs and increased speed. The entire industry is undergoing visible change, because the hardest part of metals has always been R&D: high investment, long cycles, uncertain outcomes. And this can now be almost entirely handed over to AI.

ZP: What exactly does DeepMaterial do today? When customers come to you, what do they walk away with?

Xuanze Wang: We're an end-to-end, full-chain, all-in-one company. We don't deliver intermediate results — we deliver materials that can be deployed. From the customer's perspective, they only need to state requirements; they don't even need to provide data. We directly deliver a formula that can be mass-produced on their line, tune it on their production line, and the customer takes it to make money, sharing a portion with us. In some areas, customers say your material is so good, why don't you just make it? Then we'll find OEM partners to produce and supply the material to them.

ZP: How was the technical system supporting this end-to-end delivery built?

Xuanze Wang: We're now operating in an agent R&D mode. The entire system is called M-Loop, with four core components: a large foundation model called M-Cortex; domain-specific small models and computational tools called M-Science; wet-lab experiments under M-Lab; and the underlying M-Data infrastructure.

Earlier I mentioned the large model handles orchestration — essentially, it's functioning as an intelligent agent. M-Cortex is the core of that agent, surrounded by a harness architecture where the small models and other components sit. Beyond general harness components, there are AI for Science-specific elements: the lab itself, knowledge graphs, M-Data, and the overall large model. The large model calls upon small models and these modules, and self-improvement is also driven by the large model.

ZP: Why build all four components in-house?

Xuanze Wang: Because we're thinking about this from the end goal backward.

The future of materials R&D will be wish-based: you make a wish to the AI, and a few weeks later the AI delivers the material, wet experiments and industrialization included — you can either start production or publish a paper directly. We're not there yet; self-improvement hasn't been fully realized either. But to reach that point, all four pieces — large model, small models, lab, and data — are indispensable. We're one of the few teams working on all four. Combined with our 720-dimensional high-dimensional data as the foundation, each step can be executed solidly, and materials can actually be developed.

ZP: Can you give a concrete example that quantifies the value AI-driven R&D creates?

Xuanze Wang: For one of our materials, compared to a patented material from a major international aerospace company from 2017, our core fatigue performance is more than double, at one-tenth the cost, because we eliminated expensive elemental additions. For example, a leading consumer electronics brand uses titanium alloys; our new material achieves 1.7x the tensile strength of conventional TC4 at equivalent cost, with comparable elongation. These are already delivered — nearly 20 materials, all at roughly this level.

ZP: What results has this methodology produced so far? How far from actual product integration?

Xuanze Wang: We've developed nearly 20 new metallic materials, likely exceeding 20 by year-end, with world-leading performance, all independently developed IP with core formulation patents held by the company. Historically, aluminum alloys ran from the 1xxx to 7xxx series, each with different compositions. We've completely broken outside that framework, tuning eight elements simultaneously.

All performance data has been tested by CNAS-certified third-party labs, not ourselves, and verified by end customers as well. We have extensive test reports, and everything is moving toward industrial delivery.

Several materials are already in mass production. In the next year or two, the phones, watches, and golf equipment people use may incorporate our materials. Golf has already landed this year and is entering mass production.

ZP: You describe M-Loop as capable of recursive self-improvement. Where specifically does "getting smarter with use" show up, and how does it differ from pure language model self-improvement?

Xuanze Wang: It shows up not just in material performance improving, but in R&D capability itself improving.

M-Loop contains scientific research capabilities, knowledge graphs, and tool-class algorithms. On one hand, materials get better and better; on the other, the system itself proposes theoretical hypotheses and develops new tool-class algorithms, validating them through dry and wet experiments. What survives extensive validation gets deposited into the knowledge graph or scientific toolkit — that's a capability upgrade. We're already seeing some effects, but we're still distant from the endgame of wish-based R&D.

So it's not just model RSI. The core large model evolves, and the harness around it evolves too. Like a child growing up — it's not just the brain developing, the heart pumping stronger, the stomach graduating from milk to meat; the entire body grows. I believe this is true self-improvement: from elementary school, middle school, university, all the way to PhD, ultimately making the agent the most capable researcher possible.

ZP: What role do human materials scientists play in this system?

Xuanze Wang: Two roles: designing this loop, while simultaneously transferring their own capabilities to the large model to learn.

Many of our scientists are already doing R&D in an AI-native way, and all of this gets recorded. Our data comes in two types: structured real experimental data — from physically breaking bars in tests; and the R&D process itself — recordings of R&D meetings, dry and wet experiments at every stage, all fed to the large model for understanding.

ZP: What's the current commercial progress? Which application directions are you prioritizing?

Xuanze Wang: We reached eight-figure RMB revenue scale last year. Materials have not only achieved mass production but completed real commercial delivery with ongoing repeat purchases. In terms of applications, the higher-certainty, shorter-cycle directions are consumer electronics, sporting goods, aerospace, and thermal management materials, with consumer electronics industrializing soon. Longer-term directions like controlled fusion materials and liquid-cooled server materials we're also positioning for early.

ZP: Beyond materials themselves, you're also promoting the One Person Lab platform. What problem does it aim to solve, and how broadly applicable is it?

Xuanze Wang: OPL, One Person Lab, is an all-in-one service platform we've built. Leading research institutions domestically and internationally are buying our equipment; we've federated the sold equipment into a distributed laboratory network, which became OPL. The ideal form is: one person at home communicates with our large model via computer and gets materials developed, with wet experiments handled by us. Of course, actual users won't be individuals but R&D teams, or R&D leaders with budgets inside companies.

This platform's distinguishing features are closed-loop dry/wet experiments, closed-loop industrial chain, and pilot-scale capabilities — it's result delivery, not process delivery. You can publish papers directly, with figures and charts auto-generated; or go straight to industrialization, since production line conditions are already factored in, requiring only minor tuning with limited production data for mass production.

The reason we can open this up is that this capability applies beyond metals to macroscopic performance materials generally — inorganic non-metallic ceramics, composites, polymers — all using the same underlying logic.

05

Moats, Competitors, and Endgames: When Materials Are No Longer Scarce, Where Will the AI for Science Giants Emerge?

ZP: If leading materials companies or major steel mills decide to build their own AI teams and high-throughput labs, what's hardest to replicate about you?

Xuanze Wang: Honestly, the hardest thing to replicate is our DNA: genuinely understanding both materials and AI.

Large enterprises face real difficulties here. They've talked about it for years without much progress. One of our early investors decided to back us for an interesting reason. He'd visited research institution labs with excellent equipment — tens of millions per machine — but lacked people who could bridge materials and AI. After that, he became more convinced this belongs with startups. Large companies are hard to turn; spend millions hiring an AI person, processes don't push through easily, reporting structures get awkward. Harder still is finding someone who truly understands both sides, knows what to invest in, and at what stage. I often say my dollar goes ten times further because I understand both — I know exactly what to invest in at each node.

Beyond DNA, there are hard barriers built up over years. On equipment, our automated lab has iterated to 3.0. On data, we have 720-dimensional, 10GB-per-record high-dimensional data, tens of thousands accumulated, and this is the only way to obtain real data. On algorithms, we've been building domain-specific models since the neural network era, and now have the M-Loop agent. On formulations, nearly 20 materials are all independently patented, and among materials humanity hasn't yet explored, we hope to capture a substantial share of the best formulations. We've kept extremely low-key, almost no PR, but the problems the industry faces today are ones we encountered and began solving five or six years ago.

What ties this together is determinism. We've transformed uncertain R&D into a deterministic materials production line, meaning that with resource investment, materials can be rapidly developed. The most powerful thing about large models is their capabilities deterministically strengthening — we're the same. This is a land-grab phase; once the closed loop and flywheel spin up, the time cost for latecomers to catch up grows increasingly steep.

ZP: Overseas companies like Periodic Labs have also emerged as AI autonomous lab companies. How do you see the similarities and differences in approach between Chinese and international players?

Xuanze Wang: The biggest difference is that we're doing truly end-to-end materials R&D, an all-in-one platform.

From what I understand, Periodic is also building high-throughput labs, oriented toward controlled fusion, superconductors, also powder metallurgy — there's technical commonality at the foundation. But from data accumulation to industrial delivery closed loop, we started earlier and have gone deeper.

We have a thesis: actually making materials happen will occur between the US and China, because AI is primarily these two; going further, it will happen in China, because China has the complete industrial chain for rapid落地. The difference lies here. Domestically, all-in-one makes sense; internationally it's more partnership-based, clients are willing to pay and pay more than domestically, so service models suffice. So I believe the largest AI for Science company will definitely be in China, because here you can do end-to-end, you can do production.

ZP: AI for Science business models roughly include selling models, selling experimental platforms, doing R&D services, doing IP licensing, and doing end-to-end yourself. How do you see their evolution? Why did Chuangcai choose its path?

Xuanze Wang: They can be divided into two categories — the first several are one category, end-to-end is another.

Materials ultimately need to be industrialized and profitable to create value. Service models provide services to production companies, collecting service fees. The problem is, like how the last wave of AI was too open-source, this wave of large models is somewhat better but still intensely competitive — whether value gets compressed by competition is a major question.

What truly creates greater value is making things. Customer demand is simple: I want to produce these materials, I want to make money. And once you reach end-to-end, full-chain closed loop, you discover that large models, small models, and lab capabilities are mutually reinforcing — improvement in each is far more significant than doing any single piece alone. For example, doing only labs makes it hard to know what data to collect or how to add sensors.

In such an intense competitive environment, if you don't directly help customers make money, they won't pay — at least not repeatedly, maybe just a taste. So our model is helping customers make money first, then sharing a portion of profits.

ZP: If R&D costs and timelines are compressed by an order of magnitude, how does the materials industry landscape change?

Xuanze Wang: I have a prediction, not sure if it'll hold: materials become more fragmented.

There may be a material for every niche scenario, material grades proliferate enormously, costs drop very low. People no longer need to develop materials, just select from existing options. Meanwhile, a large materials company built with AI will definitely emerge, integrating traditional materials companies and the entire supply chain — we hope to move toward that goal.

It should be noted that what we tune are composition and process parameters; production equipment remains the same equipment. Having AI develop equipment is still very difficult; wish-based R&D must be achieved first before AI can develop new process categories. Until then, new processes still rely on human advancement, on veteran craftsmen.

ZP: What form might such a new-type materials company take? What will Chuangcai itself do?

Xuanze Wang: You'll probably see a lot of companies or consortiums emerge. There's nothing new under the sun — the Chinese Academy of Sciences system and the CISRI (Central Iron and Steel Research Institute) system have incubated plenty of listed companies before. CISRI is a state-owned R&D institute that can't go public itself, but it can develop technologies and then spin off manufacturing enterprises. In China, setting up a production company is the easiest path — whether through independent fundraising and IPO or incubation within the system. The difference is, they relied on formulas accumulated over decades or even centuries to incubate companies, whereas we mass-produce formulas. Whichever customer has demand, we co-found a subsidiary with them to handle production.

In the near term, we're still a very light company. Because we share profits with customers, we've outsourced all production steps except the final additive — that's our formula protection, essentially mixing in a bit of our own proprietary material.

06

AI and Materials People as Each Other's "Half-Experts": The Leader Must Be Both Factory Director and Perpetual Student

ZP: People who truly understand both AI and materials are rare. How did you build this interdisciplinary team? And what are the two sides learning from each other?

Xuanze Wang: At first I debated whether to find hybrid talent or experts from each side. Hybrid talent is extremely hard to find, and when we did bring some in, the results weren't as good as having two groups of experts learn from each other. So now one side of our team is AI talent from top schools and major tech companies, and the other side is materials experts with high academic h-indexes and impact factors.

After a few years, both sides have become "half-experts" in the other's domain. The AI people tell the materials people what models exist, what problems they can solve, and what the input data should look like. The materials people tell the AI people where a material's core performance lies and what crystal phase means. Only with mutual understanding can an "all-in-one" platform truly be built.

The most typical example is those 720 dimensions. Early on, materials researchers would look at certain dimensions and say, this one is definitely useless, don't collect it. Our view was: maybe it's useful, so collect as much as possible. When we fed them into the model, performance improved significantly. And through gradient backpropagation, the model could to some extent judge which dimensions mattered more — and it turned out the supposedly "useless" ones were actually more important. Expert intuition sets the starting point; data tells you where the boundaries really are.

ZP: Two teams with different mindsets — what mechanisms do you use to get them to actually collaborate?

Xuanze Wang: Build trust through results, then institutionalize the collaboration.

At first, the AI people learned from the materials people, and the materials people thought AI was unreliable. But AI quickly delivered results. Everyone's young, so mindsets shift fast — they started actively embracing AI. Young teams have a real advantage here over many veteran experts.

On the institutional side, we've done several things. First, A/B testing — let the results speak. Second, end-to-end assignments where both sides deliver together, with weighted KPIs: if one side underperforms, the other's coefficient drops too. Third, some hard requirements, like mandatory weekly meetings and regular in-person gatherings in Suzhou or elsewhere. Once collaboration became habit, these强制措施 gradually phased out. Now it's mostly spontaneous.

ZP: Let's talk about the team itself. How is your core team structured?

Xuanze Wang: Our core team covers four segments: AI, materials, engineering, and commercialization. Each segment's leader can operate independently.

On the technical side, our algorithms lead graduated from Shanghai Jiao Tong University, with core members from companies like ArcSoft and ByteDance who have led multiple algorithm projects worth tens of millions of RMB. The materials team is led by a PhD in materials science from UNSW, with nearly ten members holding materials science PhDs who have published over 20 papers in journals including Nature and contributed to national materials database construction. Engineering team members come from leading companies in additive manufacturing and high-end metal materials, averaging about ten years of experience in aerospace and 3D printing — enough to support rapid industrialization.

On the commercialization side, our operations lead previously worked in investment, managing a portfolio worth hundreds of millions of RMB, before actively joining us. Our sales lead spent 16 years combined at Huawei and Goertek, rising to head of key accounts and handling annual sales in the billions of RMB. We've also brought on the former executive deputy director of a national key laboratory and a veteran entrepreneur from industrial-grade 3D printing as advisors. This structure lets us maintain very high per-capita efficiency with a relatively small team. We're now rapidly expanding.