在化成前构筑经设计的负极界面层，使界面由随机生成转为可控设计。 An engineered anode interface, built before formation, so the interface is designed rather than left to form at random.

硅提高了负极容量，也使界面在每次循环中重新生成。Silicon raises anode capacity, and it also forces the interface to re-form on every cycle.

硅碳负极已进入旗舰智能手机电芯量产。硅的理论比容量约为石墨的十倍，但锂化时体积变化接近 300%。颗粒反复胀缩使固体电解质界面膜（SEI）破裂、再生，持续消耗活性锂与电解液，表现为首次库仑效率下降、循环容量衰减和产气。 Silicon–carbon anodes are in volume production for flagship smartphone cells. Silicon's theoretical specific capacity is roughly ten times that of graphite, but its volume changes by close to 300% on lithiation. Repeated swelling cracks the solid–electrolyte interphase (SEI), which then re-forms, consuming active lithium and electrolyte. The result is lower initial coulombic efficiency, capacity fade and gas generation.

消费电芯对产气导致的厚度膨胀有严格的限值；快充时负极表面的浓差极化又会提高析锂风险。电解液添加剂能改善 SEI 的组成，但不能控制界面的结构与离子输运。 Consumer cells have tight limits on thickness growth from gassing, and under fast charge, concentration polarization at the anode surface raises the risk of lithium plating. Electrolyte additives change SEI composition; they do not control the structure of the interface or ion transport through it.

我们按结构与性质的对应关系设计包覆层聚合物。We design the coating polymer from its structure–property relationships.

电池负极包覆通常沿用陶瓷或通用粘结剂类材料。我们的方法是先确定界面需要的性质，再从链结构上逐项实现：主链的刚性与缠结决定涂层能否随颗粒胀缩而不开裂，侧链上的固定电荷基团通过 Donnan 排斥降低阴离子在层内的浓度，使层内锂离子迁移数接近 1，从而减弱快充时的浓差极化。 Anode coatings are usually ceramic or general-purpose binder chemistries. We start from the properties the interface needs and build each one into the chain structure. Backbone stiffness and entanglement determine whether the coating follows particle swelling without cracking. Fixed-charge groups on the side chains lower the anion concentration inside the layer through Donnan exclusion, which pushes the local Li⁺ transference number toward 1 and reduces concentration polarization under fast charge.

同一设计框架通过调整侧链化学可以面向不同的界面问题。现阶段我们只做消费电芯硅碳负极。 By changing the side-chain chemistry, the same design framework can address other interface problems. For now we work only on silicon–carbon anodes for consumer cells.

在纽扣电池里表现好的涂层，常在浆料、辊压和软包电芯中失效。我们按产线的顺序验证。Coatings that work in coin cells often fail in slurry, calendering or pouch cells. We validate in the order the production line sees them.

1. 浆料相容性与流变Slurry compatibility and rheology与负极浆料体系共混，无凝胶化与相分离，涂布窗口不变。Blends with the anode slurry without gelation or phase separation; coating window unchanged.
2. 涂层完整性Coating integrity辊压后覆盖连续性、电解液溶胀与剥离强度。Coverage continuity after calendering, electrolyte swelling and peel strength.
3. 电化学稳定性Electrochemical stability0.05–0.5 V（vs. Li/Li⁺）区间的不可逆还原、首次库仑效率、阻抗。Irreversible reduction between 0.05 and 0.5 V vs. Li/Li⁺, initial coulombic efficiency, impedance.
4. 软包电芯Pouch cells与含 FEC/VC 的优化电解液基线对比：循环、快充、产气厚度膨胀。Compared against an optimized FEC/VC electrolyte baseline: cycling, fast charge, gassing-induced swelling.
5. 电芯厂产线验证Line qualification at a cell maker在客户电芯与产线条件下确认性能与工艺窗口。Performance and process window confirmed in the customer's cells and on its line.

当前Now与一家头部消费电芯企业开展联合开发。In joint development with a leading consumer-cell manufacturer.

推理台把一个合成或材料问题转化为可核对的路线、条件和电子结构数据。The bench turns a synthesis or materials question into routes, conditions and electronic-structure data that can be checked.

这是我们自建、本地部署的研发系统。语言模型、逆合成与正向预测模型、量子化学计算都运行在自有服务器上，专有结构不离开内网。每个输出的数值都标明来源与状态：文献实测、模型推断或待标定。 This is our own system, deployed on-premises. The language model, the retrosynthesis and forward-prediction models and the quantum-chemistry jobs all run on our servers, so proprietary structures stay on the internal network. Every number it outputs carries a source and a status: measured in the literature, inferred by a model, or awaiting calibration.

1. 理解需求Parse the request本地大模型把中文或英文提问解析为目标分子与化学问题；分子名称经结构库核对后再进入计算。A local LLM turns a question in Chinese or English into a target molecule and a chemical task; names are checked against a structure database before any computation.
2. 路线搜索与文献检索并行Route search and literature search in parallel蒙特卡洛树搜索生成完整逆合成路线，每一步用正向预测模型复核；同时检索文献。Monte Carlo tree search builds complete retrosynthetic routes and a forward-prediction model re-checks every step, while the literature search runs alongside.
3. 与文献比对排序Rank against the literature模型路线中的步骤若在文献中有对应报道，该路线排在最前。Routes whose steps are reported in the literature rank first.
4. 综合工艺条件Merged process conditions文献抽取的参数与模型推荐条件按槽位合并，输出结构化操作序列。只有全部必需数值均为文献实测时，才允许下发到自动化设备。Parameters extracted from papers are merged slot by slot with model-recommended conditions into a structured operation sequence. It can be sent to automated equipment only when every required value is literature-measured.
5. 电子结构Electronic structureGFN2-xTB 几何优化，B3LYP/def2-SVP 计算前线轨道、偶极矩与静电势表面。GFN2-xTB geometry optimization; B3LYP/def2-SVP frontier orbitals, dipole moment and electrostatic-potential surface.

{
  "reaction": "二氢青蒿素 + 琥珀酸酐 → 青蒿琥酯",
  "solvent":     { "value": "DCM",   "status": "inferred" },
  "temperature": { "value": "14 °C", "status": "inferred" },
  "time":        { "value": null,    "status": "to_be_calibrated" },
  "forward_check": "consistent",
  "dispatch_allowed": false
}

• 推理台以 MCP（Model Context Protocol）服务器形式提供工具接口，智能体与实验编排系统可直接调用路线规划、工艺条件与电子结构查询。The bench exposes its tools as an MCP (Model Context Protocol) server, so agents and lab-orchestration systems can call route planning, process conditions and electronic-structure queries directly.
• 在示例中，青蒿琥酯的首选路线与文献一致；反应时间缺少文献实测值，因此系统不允许下发。In the example, the top artesunate route matches the literature; the reaction time has no literature-measured value, so the system blocks dispatch.

我们和电芯企业一起，从具体的界面问题出发确定材料方案。We work with cell makers, starting from a specific interface problem.

Jeremy Gao

创始人。高分子物理博士，负责材料设计与推理台架构。Founder. PhD in polymer physics; leads materials design and the architecture of the bench.

联系我们Contact

请在邮件中注明来意：电芯合作、投资或其他。Please say what your email concerns: cell partnership, investment or other.