AI 化工全景图 · 一个分子的放大 · 2026-07 · 全产业链 + 硬骨头Chemicals × AI · The Scale-up of One Molecule · Jul 2026 · Full chain + hard bones

从一克,到一万吨 From one gram to ten thousand tonnes

手机、汽车、衣服、药品、房子里的每一克物质,都在某个反应釜里被合成过一次——化工是万亿级产业,却几乎没有消费者认知:它是一切实体产品的隐藏 BOM。本图的最小单元是一次放大(one scale-up):把实验室烧瓶里的克,放大到工厂里的万吨——这既是这个行业的核心手艺,也是核心风险。AI 正分层吃进化工:过程控制层已经真金白银地赚钱,配方层渗透最快,「AI 发现新材料」的声量最大、证据最薄,而从克到万吨的那一跳,AI 至今完全无法代劳。 Every gram of matter in your phone, car, clothes, medicine and house was once synthesised in a reactor — chemicals are a trillion-dollar industry with almost no consumer awareness: the hidden BOM of every physical product. This map’s smallest unit is one scale-up: turning laboratory grams into factory kilotonnes — the industry’s core craft and its core risk in a single act. AI eats chemicals in layers: process control already earns real money, formulation moves fastest, «AI discovers new materials» is the loudest claim with the thinnest evidence — and the leap from gram to kilotonne remains the one act AI cannot perform.

本图相信:流程工业 AI 是真生意——APC 节能 3–15%、回收期 6 个月–3 年;横河的强化学习控制器已在真实化工厂正式商用;Emerson 以 168 亿美元企业价值全资收购 AspenTech;中控技术拿着国内 DCS 40.4% 的份额发布时序大模型;配方化工的组合爆炸(一万种组合起步)是 AI 的天然试验场(B)。This map believes: process-industry AI is real business — APC saves 3–15% energy with 6-month-to-3-year paybacks; Yokogawa’s reinforcement-learning controller runs a real plant commercially; Emerson bought AspenTech at a $16.8B enterprise value; Zhongkong ships time-series models atop a 40.4% domestic DCS share; and formulation’s combinatorial explosion (ten thousand combos to start) is AI’s natural playground (B).
本图不相信:「AI 已独立发现新材料」。GNoME 的漏斗:220 万预测→38 万「稳定」→736 个被外部合成(0.03%)→同行评审的结论是「新颖-可信-有用三重验证的证据极少」;A-Lab 论文 2026 年被《自然》发布作者更正;明星公司 Orbital 转型去卖数据中心冷却硬件。也不相信「国产化突破」公告=替代完成——送样验证、小批量、规模量产三态之间,通常还差 2–3 年(A/B)。It does not believe «AI has independently discovered a new material». The GNoME funnel: 2.2M predictions → 380k «stable» → 736 externally synthesised (0.03%) → peer review finding «scant evidence» of novelty, credibility and utility together; the A-Lab paper drew a 2026 Nature author correction; star startup Orbital pivoted to selling data-centre cooling hardware. Nor that a «breakthrough» notice equals substitution — sampling, small-batch and mass production sit 2–3 years apart (A/B).

主脊:一个分子的七站(分子发现→小试中试→工艺放大→连续化生产→配方化工→下游认证与国产替代→绿色与地缘),④⑤标红;三条结构带:装置经济与大出清、精细化工与隐形冠军、国产替代三态诚实账;判断层:AI 越强,隐性知识越值钱——公开知识的处理成本趋近于零时,竞争的全部重量压到没被公开的那部分上。姊妹:同构的老师傅手感→factory,上游材料清单的下游→chip,电池材料接口→renew,AI 能否做科学→sci The spine: one molecule, seven stations (discovery → pilot → scale-up → continuous production → formulation → qualification & substitution → green & geopolitics), ④⑤ in red; three bands: plant economics and the great clear-out, fine chemicals and hidden champions, the three-state honesty of substitution; the judgment: the stronger AI gets, the more tacit knowledge is worth — when processing public knowledge costs nothing, all competitive weight lands on what was never published. Sisters: factory, chip, renew, sci.

17 年
万华引进的 1 万吨 MDI 装置从投产到达产用了近十七年——图纸齐全、设备完整、专利清晰,它就是开不起来:纸面工艺与万吨稳定运行之间,隔着的不是文档,是隐性知识(A/B)Wanhua’s imported 10kt MDI plant took ~17 years from start-up to full rate — complete drawings, equipment and patents, and it still would not run: between paper process and stable kilotonnes lies not documentation but tacit knowledge (A/B)
0.03%
GNoME 漏斗:220 万预测晶体→38 万「稳定」→736 个被外部实验室合成;同行评审复核的原话是「新颖、可信、有用三者兼备的证据极少」(A)The GNoME funnel: 2.2M predicted crystals → 380k «stable» → 736 synthesised externally; peer review’s verdict — «scant evidence» for novelty, credibility and utility combined (A)
−11%
2026 年中国乙烯-HDPE 一体化装置 ROI 预计跌至历史低点,2032 年后才转正;全球乙烯开工率降到 84%——史上最深的一轮过剩(B)China’s integrated ethylene-HDPE ROI is forecast to hit a record −11% in 2026, positive only after 2032; global ethylene utilisation fell to 84% — the deepest glut on record (B)
45% vs 60–70%
中国精细化工率 vs 北美西欧日本——20 个百分点的差距,正是 C&EN 全球 50 强里中国 10 席全部是炼化一体化型的直接映射(B)China’s fine-chemicals ratio vs North America, Western Europe and Japan — a 20-point gap mapped directly onto the C&EN Top 50, where all ten Chinese seats are integrated refiners (B)
口径裁判:精细化工率 45% 与「近五成」并存,均无国家统计局原始出处;化工新材料自给率 79%/60% 与「仅 14%」相差 5 倍——「自给率」在中文语境几乎没有统一定义,引用必须问口径;电子特气国产化率 25% vs 55%、12 英寸硅片「不足 1%」vs 数十万片/月产能,均为口径冲突并列呈现;万华 MDI 380 万吨/33% 与 320 万吨/32% 为不同时点;AI 配方加速倍数(10 倍/4 倍/90%)全部为厂商口径,无一获第三方量化验证;「某云实验室 2025 年关闭」的流传说法查无实据。Refereeing the numbers: fine-chemicals ratios of 45% and «near 50%» coexist with no statistical-bureau source; new-materials self-sufficiency spans 79%/60% vs «only 14%» — a 5× spread, so always ask the denominator; specialty-gas localisation runs 25% vs 55% and 12-inch wafers «under 1%» vs hundreds of thousands of wafers monthly — conflicts listed, not reconciled; Wanhua’s 3.8Mt/33% and 3.2Mt/32% are different vintages; every AI formulation speed-up (10×, 4×, 90%) is vendor-claimed with zero third-party verification; the rumoured 2025 cloud-lab shutdown has no reliable source.
中心装置 · 万华 MDI 三十年:同一种分子,两种命运The central device · Thirty years of Wanhua MDI: one molecule, two fates
这不是「引进-消化-吸收」的成功故事,而是一个引进彻底失败、逼出自主研发的故事——它把本图的核心命题演示得比任何理论都好。Not a licence-digest-absorb success story but one of total import failure forcing indigenous R&D — it demonstrates this map’s core thesis better than any theory.
上半场:买来的十七年(1978–1995)First half: the seventeen bought years (1978–1995)
1978 年从日本引进 1 万吨/年 MDI 装置,总投资 5 亿元,当时国内最大轻工项目。日方未转让核心技术,装置 1983/84 年投产后长期开不稳,近十年未能满产,直到 1995 年才达产。图纸齐全、设备完整、专利清晰——它就是开不起来。你可以读完所有 MDI 专利,仍然造不出稳定运行的装置:这是化工版的隐性知识A 10kt/yr MDI plant imported from Japan in 1978 for ¥500M — then China’s largest light-industry project. Core technology was withheld; started up in 1983/84, the plant ran unstably for nearly a decade, reaching full rate only in 1995. Complete drawings, equipment and patents — and it would not run. You can read every MDI patent and still not build a stable plant: tacit knowledge, chemicals edition.
下半场:逼出来的世界第一(1993–2026)Second half: the forced world №1 (1993–2026)
1993 年起自主攻克缩合、光气化、分离精制三大核心技术,1996 年 1.5 万吨装置试产一次成功;2000 年中国成为全球第五个拥有 MDI 自主知识产权的国家;2011 年逆势收购匈牙利 BorsodChem。截至 2024 年末产能 380 万吨/年、全球份额约 33%,全球第一;TDI 144 万吨全球第一;2024 年研发投入 45.5 亿元、研发人员 4763 人。全球至今仅 5 家企业拥有 MDI 自主知识产权,核心反应器设计数十年未外泄。From 1993 the team cracked condensation, phosgenation and separation on its own; a 15kt plant started up first-try in 1996; by 2000 China was the fifth country holding sovereign MDI technology; in 2011 it bought Hungary’s BorsodChem at the bottom. By end-2024: 3.8Mt/yr, ~33% global share, world №1; TDI at 1.44Mt also №1; ¥4.55B R&D and 4,763 researchers in 2024. Only five firms on earth hold sovereign MDI technology; the reactor designs have not leaked in decades.
判词Verdict工艺 know-how ≫ 专利文本。而 AI 时代给了这条老规律一个新推论:当 AI 把公开文献的重组成本打到接近零,专利反而更容易被判「显而易见」——AI 越强,公开专利越贬值,技术秘密越增值。这不是 AI 消灭了老师傅,是 AI 让老师傅变得更贵。Process know-how ≫ patent text. And the AI era adds a corollary: when AI reduces the cost of recombining published literature to near zero, patents become easier to rule «obvious» — the stronger AI gets, the more public patents depreciate and trade secrets appreciate. AI didn’t eliminate the master craftsman; it made him more expensive.
Reading the Map

从这张图带走的五条规律Five patterns to take away

立场声明:本图为化工产业结构分析,不构成任何投资建议;全文不含任何危险化学品的工艺配方、合成路线或安全防护规避细节;事故数据全部来自官方调查报告,对每一位遇难者保持敬意——安全不是化工的成本项,是第一约束;冲突数据一律并列呈现并标注口径,厂商自述与第三方核查严格分级;市场地图 ≠ 推荐背书;仅供研究与批判性参考。Stance: a structural read of the chemicals industry — not investment advice; nothing here contains process recipes, synthesis routes or safety-circumvention detail for hazardous chemicals; accident data comes from official investigations, with respect for every victim — safety is chemistry’s first constraint, not a cost line; conflicting data are listed with their denominators, vendor claims strictly separated from third-party checks; a market map is not an endorsement; for research and critical reference only.