杜克大学团队造AI闭环造芯系统 提振二维半导体良率30%

杜克大学团队造AI闭环造芯系统 提振二维半导体良率30% 通过让模型避开材料缺陷定制晶体管并自主制造物理芯片，AI与硬件将形成软硬一体的“自生智能”闭环演化。 Harry Wang从晶体管微缩的物理瓶颈讲起，指出硅材料在5纳米以下因漏电和电子迁移率暴跌面临极限，而二硫化钼等超薄二维材料虽具备优异特性却受制于极低的制备良率。为此，其团队结合Meta SAM大视觉模型与Google DeepMind的Co-Scientist智能体架构，实现了让AI自主文献检索、合成配方生成及微观材料缺陷检测的端到端自动化。更进一步，团队颠覆传统均一化光刻思路，利用AI规避材料缺陷定制非均质芯片，使晶体管阵列良率提升约30%，并由此延伸出由模型与实体硬件闭环演进的“自生智能”（Autogenic Intelligence）乃至软硬件是否会走向“向蟹化趋同演化”的前沿猜想。 传统硅基材料在厚度降至5纳米以下时面临严重漏电与电子迁移率断崖式下滑，当前产业界的“2纳米/1纳米”等指标本质上是通过NanoSheet堆叠实现的公称节点，而非单一物理沟道尺寸。 (5:00) 杜克大学团队开发了由大模型与计算机视觉驱动的智能体系统，不仅能以99.7%的良率识别二维材料缺陷，还与Google DeepMind合作实现了根据实验室硬件自适应生成合成配方的闭环科研。 (15:00) 针对二维半导体缺陷密集的落地瓶颈，团队打破统一光刻范式，让AI在纳米级健康区域定制非均质晶体管，将原始材料仅65.6%的制造良率提升了约30%。 (25:00) 演讲者提出“自生智能”（Autogenic Intelligence）概念，探讨由AI设计并制造承载自身算力的芯片闭环，并借生物学“向蟹化”（Carcinization）设问软硬件协同演化是否终将趋同。 (26:30)

杜克大学助理教授 Harry Wang 在硅谷101大会上指出，传统硅基芯片正逼近物理微缩极限，而具有极薄优势的二维半导体受制于材料缺陷与良率瓶颈；通过引入 AI 接管材料表征、实验合成与非均质芯片制造，能够打破传统光刻均一性限制并大幅提升良率，推动软硬件共同进化的“自生智能”闭环。 硅基物理极限与二维材料的微缩优势 (5:00) 芯片演进依赖在晶体管内堆叠开关，如英伟达 Blackwell 架构已集成突破 2000 亿颗晶体管。当硅材料通道微缩至 5 纳米以下时，材料底层易发生严重漏电（leakage）并导致电子迁移率断崖式下滑。Harry Wang 澄清业界所谓的“2纳米”或“1纳米”仅为依赖 NanoSheet 立体堆叠换算出的公称节点（nominal nodes），并非物理栅极真实尺寸。为此，单层厚度仅约 0.65 纳米的二硫化钼（MoS2）等二维过渡金属硫族化合物，凭借其极薄沟道能够杜绝关断漏电，在 5 纳米以下展现出超越传统硅基单器件的物理性能。 良率之壁与 AI 驱动的自动化微观表征 (10:00) 台积电路线图虽规划迈向互补场效应晶体管（CFAT）并引入二维材料，但其落地面临致命的良率问题：传统 12 英寸硅晶圆纯度可达 10 至 12 个 9，全片仅一两处堆叠缺陷；而气相沉积（CVD）生长的二维材料缺陷密度极高。为此，团队开发了名为 Atomic 的工作流，利用大语言模型调度底层显微镜，结合 Meta 的 Segment Anything Model（SAM）等视觉模型进行逐像素分析。该系统在无需数据中心级专门重训的前提下，将材料与缺陷分类准确率提升至 99.7%，敏锐捕捉肉眼易漏检的畴界缺陷，推动技术走向“AI for Engineering”。 人机协同实验与颠覆均一化制造范式 (20:00) 针对跨实验室硬件差异导致实验难以复现的问题，杜克大学团队联合 Google DeepMind，将 Co-Scientist 架构接入物理机台，依据设备具体几何结构直接生成定制配方，实现了一次试样跑通目标材料合成。在芯片制造端，团队打破了传统均质化光刻要求晶体管整齐划一的规则，让 AI 逐片扫描缺陷密布的“非理想”合成衬底，自主挑选纳米健康区域按需定制每个晶体管。实验表明，传统二维材料初始制备良率仅约 65.6%，而经过 AI 闭环制造后晶体管阵列良率提升约 30%，团队已就此成立初创公司探索 AI 原生半导体制造路线。 软硬件一体化构筑“自生智能”飞轮 (26:30) 针对递归自我改进（RSI）忽视硬件制造的现状，Harry Wang 提出软件模型（Mind）与物理芯片（Brain）相融合的“自生智能”（Autogenic Intelligence）概念：模型利用数据设计并指导仪器制造芯片，新芯片反过来提供算力训练更强大的模型，形成物理实体的闭环迭代。关于这种硅基生命的未来形态，他借用生物学中多谱系物种演化收敛为蟹类的“趋同演化”（Carcinization）现象，提出了 AI 与芯片协同进化最终会走向无限发散还是收敛至某种特定形态的开放性思考。 该演讲勾勒了从二维材料微观表征到非均质芯片制造的 AI 闭环路径，并指出了约 30% 的良率提升成效；但其整体方案仍处于从小规模先导试验线向产业化过渡的早期阶段，人机协同环节尚未完全由机械臂取代，软硬件一体化进化的最终收敛形态也仍是有待验证的猜想。

[5:00] Harry Wang: Silicon currently has some problem. The reason is if we basically cannot shrink silicon infinitely.

[0:00] 过去几十年,芯片性能的提升很大程度上依赖于晶体管不断缩小,但随着传统硅基芯片逐渐逼近物理极限,半导体行业正在寻找下一代材料和制造技术。而有意思的是,AI不仅是这场技术变革的需求推动者,也可能成为下一代芯片的设计者。那么当AI开始参与制造下一代AI芯片智能的发展会走向什么方向接下来让我们回到Alignment 2026大会现场观看这场Tino的完整演讲或者是中文化的工程和工程。所以,我再次介绍一下我一部分。我叫 Harry Wang。我现在是室长的教学。我达了我的高科学学习课在上海教通学习课然后我学了一年在 MIT 我学了一年在高tech的科学and I started my lab in Duke about three years ago今天是Valley101,我先开始了我会叫做Siticon101, 和你基本上的知识, 如何是其实是用的。在最少的地方,在最少的地方,我会说,我们也看到一些新的商品AIor to boost the design of chips and everything like that. So I will say this can become a loop.AI needs chips and chips needs AI. But what are the chips and what are the basic elements of chips?所以基本上, 全部都是所以,即使所有的大量的模式和其他模式的模式变成很 complex在某些年代的一种, 但仍然是在某些年代的一种, 但仍然是在某些年代的一种。所以,如果你看到某些人的晶片, 基本的一种, 它叫做"transistor",它是比较像一个"faucet", OK?如果我们转它上,你可以看到这里, 彩虹能浮动through the source to drain and this is the very basic element and turn from zero to one one to zero and realize everything now based on the chips and based on the AIAll right, so again, we talked about the scaling law, but the scaling law is already there in silicon or in the chips.Okay, so in semiconductor, so we do believe like if we stack more transistor in a single chip, it's just be more powerful and has higher performance.Okay, here shows like from a single transistor, if we accumulate it to billions of transistors, it can become a chip and become very functional.如果我们看这个例子,如果我们看 NVIDIA Blackwell 300 AI托子,它其实有200BDNCswitches or transistors in a single chip.OKSo in that caseLike basicallyIf you haveYou want to have better chipsYou need to have moreTransistors in a singleDevice所以在那些事件中,那是很普通的问题,如何能够进入在某些股票OK?So in that case, here shows just a look back of how transistor在50年代的50年代的一年代的一年代的一年代的一年代的一年代的一年代transistor,and then shrink it down to from a micron scale to nanoscale.然后从一个比较的旅行to the we have a thin fat and then we have like nano sheet structure basically we stack everything together所以,最近有些问题,所以我们有贫贫,然后我们先放进去,但后来先放进去 [5:00] There'll be other problems such as the thermal problem, the heat. generated at these interfaces, generate some problems. Okay? So in that case, like... So I want to introduce a little bit of common failure mode of the transistor. So transistor, when you shrink it in a small scale, so sometimes there's a problem as it also happens in the faucet in your home. OK, so if it's very thick, so basically the electron sometimes not just flowing, if you turn it off. It's not flowing directly from here, but also can flow from the... underneath. So we call it leakage. So this leakage actually, that's why I consume a lot of power and waste a lot of power and generate the heat. So in that case, if you use a thick material, like very thick, like this, sometimes like the electrons can go underneath it. In that case, you have to make the material as thin as possible. So in that case, if it's thin, then we can close the faucet in the off state. No electrons can flow here. In the on state, the electrons can flow well. So in that case, I tried to give you some common concept in silicon 101 or semiconductor 101. So smaller is better for each element and thinner is better. for each element or semiconductor materials. So we already see this trend, but I want to convince you that currently the silicon has a big problem. Okay, I really value like in the previous panel, we talk about the finesse, maybe some problem or It's still going, right? But I want to convince you like the... Silicon currently has some problem. The reason is if we basically cannot shrink silicon infinitely. So if silicon goes from like pretty large scale, to shrink it to below 5 nanometer. So electrons become, the mobility and speed of electrons become slow. Or finally, like you can see, it's not flowing well. In that case, in the on state, so no electrons can flow if you basically have a really small scale of silicon. So in that case, like as a scientist, so we think it's possible to have a next generation semiconductor. We call it two-dimensional semiconductor represented by graphene, mollabium, disulfide, Nobel Prize in 2018 10? Right? So basically-- so but comparing with a two-dimensional semiconductor So two-dimensional semiconduct can keep its electron mobility or the capability of flowing electrons, even if it's below 5 nanometers. OK, so you might be feel confused that you look at the news. Actually, the semiconductor industry always tell you, oh, we now have a two nanometer node. In the future, we have one nanometer node. Why is not like what I just said is five nanometer is a limit. Right. So I want to say like every like the node numbers showing here is it's currently is all nominal nodes. That basically means like if we check like like the the structure and the history of semiconductor and transistors. So currently, we are using the nano sheet structure. Nanoseed structure, you stack the silicons. So in that case, even though you have a 7 nanometer, you have three layers. Then you have 7 over 3. That's around 2 nanometer. Okay, so similar thing we actually already seen in the Intel's original, we have Pantom, Pantom 1, Pantom 2, Pantom 3. And now we have like 12 cores or like 14 cores. So because it's hard to shrink a single element. So basically industry decided to stack more. OK, so in that case, like I want to convince you like like silicon still fundamentally has a problem. You can make it you cannot make it extremely small like like other two dimensional materials. So but this materials, as I mentioned, is only has one molecular layer. Here shows two layers. So you have one moleculum atom and two sulfur atoms on. And in total, this material has only the thickness Below one nanometer is about 0.65, 65 nanometer. So, but still as a semiconductor, it has been showed in academia in last 10 years a single device. built upon this material is much better than a single device of using the silicon to build it. okay so if we see like when this material will enter your phone into your computer become really electronics in product so this shows [10:00] The roadmap by TSMC especially shows like, so now we are at between the nanosheet structure, moving to the CFAT as I just showed to you. But in the very near future, this will go two-dimensional transition metal dichotronites, or I just showed you as MOS2. OK? So in that case-- So you might think like if this material just the material, like I tell you, it's very thin and can perform as a channel material of a transistor. So basically, so the problem is like how to build it efficiently. So even though we have some small flakes And we demonstrate single device is great, but we always need A lot of, as mentioned, hundreds of billions of transistors in a single chip. So by building this material, basically currently in the university lab, and I know there's some industry partners are working on these similar things. We use a method called chemical vapor deposition, have some two precursors. The precursor can carry with the gas and grow. this. extremely thin semiconductor on top of substrate. Okay? So the problem is like this one layer can perform really well as a channel in the transistor, but it's not like silicon. Silicon, we use a physical way to build it, so have really high purity like 10 knives or 12 knives. on a 12-inch wafer, if you check every item, So only one or two stacking forts on 12 inch wafer. However, this material is still struggling about like getting as perfect as a silicon. So if it's not as perfect as a silicon, basically the current industrial flow should be changed. So in that case, like, so I want to also to mention that this material not just can be built as semiconductor. In my lab, we also developed something. We found if we print this material on top of some soft substrate, and this material kind of responds to... to the light, you can see it's like an inch warm, can just move one direction. I'm joking with my friends, this is doing PhD, when the advisor talking with a student, they just bend a little bit. If I don't talk with them, just laying down, right? So, but we are moving, we are always moving. So, you cannot always bend, right? So, So maybe the managers are laughing, so you know what's going on. So then you're questioning, like, this material seems very magical, but why is it still not in your phone, not in your daily life? The problem is, like, I will say just one word, it's yelled, period. So TSMC is great because you can manage Yale the best in the world. Okay, and the Intel and Samsung also very good on yield management. That means like any semiconductor industry, if you want to like get a share in the industry, so the yield management is very important. That means you have to build a chip about hundreds of billion. transistor on one chip, but how do you make sure everyone is working? Everything is working, every transistor is working. So TSMC is doing great, but for this new 2D semiconductor, yield management needs some new logic, but still something like you need to solve. Basically, I will say in the whole, go back a little bit, in the whole semiconductor industry, so we still rely on human. to do the yield management and at least to see these materials, see the defects, see it's good or see it's not good. In that case, about two years ago, my lab, we started to to look at whether we can use AI to see these materials or to see the defects of these materials. So by doing this yesterday, there was a really good talk about telescope, like to see the stars and everything. This goes through a macro scale of seeing things. So in our lab, if we want to engineer in nanoscale, we have to see these small things as small as possible. So therefore, like we published paper last year, we call it atomic. We use AI, basically large language model to control a microscope and let the microscope to see these nanomaterials and characterize these nanomaterials and understand these macro materials. So the work is done two years ago. At that time, even agents as a concept was still early. So we call it a workflow. And basically, we can use a large the microscope and to segment the image captured, and we use a meta-segment anything model [15:00] And then we can use some supervisor clustering, which is a very basic Machine Learning. Don't forget machine learning is the beginning of AI. So we can teach the AI to understand what materials are there. Is that clean? Is there defects? Or is there something, just nothing there? OK? So in that case, we can see it's very interesting. This is not a response now. Maybe we should replace it too. not silicon based thing. So... I think we're frozen. But again, like basically what we did is like we can let the microscope to do the task research talks by itself. So by doing this, basically, it's possibly to we can see this is more like in the category of AI for science. Or we can say like it's actually like more like a structure of like how to use AI to actually create. see the manufacturing of the semiconductors. Okay? So, and the most importantly, we notice it's possible to So at that time, people are still questioning, is Chesapeake GPT or still have a lot of hallucination? This word recently not heard by many of the research. But at that time, the hallucination is a problem. But we validated that it's possibly to use AI to identify the material and also to finally classify the results by just seeing the material. or what are the nothing over there. OK. So basically, we found this model works really well and it's working now. Yeah, so this model works really well. So we can see like comparing the human eyes, this yield of the model can reach to 99.7, even though the latency is still a problem. The human just look at it and then you can get results. But we still need six seconds. So recently we are lowering the latency, it's working pretty well. But again, like in some tasks, it's even better than human. So, if I show the left image showing here, because I'm telling you, like asking you to label all the possible like MS2 flakes, sometimes the human will just label like this in the red, okay? But the model cannot just see. the whole flakes, but also identify some We call it domain boundaries. It's a boundary of different domains. And these boundaries are actually a kind of defect. So there's no tricks here, but because a human, we just use intuition. by our eyes, but this model, they have pixel by pixel analysis. So in some certain task, I try to convince you that AI can work definitely better than human. So we also try different kinds of materials in semiconductor. For example, showing here, like different kinds of, for example, graphene, which is a Nobel prize-winning material and all kinds of 2D semiconductors, this model works really well. And most importantly, we even don't need to train it. Can you believe it? We don't need data center. We can just build an agentic system and do the prompts, and we can manage it to harness it for our application. actually how to manufacture it, how to get it. So I will say in the last two years, AI for science is very important. And I will say soon we'll change to, I'm thinking maybe the AI for engineering will be very important. So in that case, we think for this novel semiconductors, it's possible to use AI. to speed up the discovery and speed up the manufacturing. Okay, so actually we already see the AI is very good on hypothesizing Ideas. So, Google's co-scientists actually published last year and they realized using some structure like alpha evolved. So AI can give more ideas. OK? And also, the problem is, as mentioned yesterday in the AI4Science panel-- Actual physical experiment and also the actual real-world validation is really difficult. OK? So in that case, like what we did, this paper just released a lot-- four weeks ago. So my lab collaborated with Google DeepMind. And we tried to use this co-scientist structure and add it into the physical world and the physical experiments to do the end-to-end validation and the discoveries. Okay? So basically, the idea is I want to introduce the work a little bit. [20:00] So the co-scientists now... cannot just do the research review, it can give ideas and also can do the, give the, experimental or the execution plans And also we connect our our instruments and a self-built experimental setup and let AI to control it. And now it can do the experiment by itself. Okay. So in that case, like in about one year ago, people are thinking in the future, Harry, you don't need a PhD student. Okay. Then after I saw this work, I'm saying like, maybe you don't need a professor because AI can give ideas better than mine. So why need a professor? And also it can write paper by itself. So maybe we should all work on data center. So basically, this work can do the reads literature, design experiment, validate end to end. And more importantly here, this video shows how it works. As a co-scientist geometry, you can-- can predict the precursor and also generate a protocol to actually execute it. And also like controlling some panel in the software and the human is still in the loop right now. So, so, so we have to use hands to weigh in the precursor. put it into the tube and also seal it. But maybe soon we can use a robot arm in that hallway to replace this stuff. But again, it's generated a really good recipe and even in the paper, we highly… truly believe it's possible to grow some new semiconductor never grown before. OK, even theoretically predicted, but it's very hard to grow. And for the conventional, 2D semiconductors as I showed you, MS2 and everything. So it also shows really good results. We just let AI to control the instrument. It looks at the literature and get the results. The most importantly we found I will say like in academia, many of the results are very hard to repeat because different labs have different geometry. So my lab have this furnace, have that tube, but the published result has a different setup. So in that case, like we found like AI can... work really good on how to repeat the results. That means that if we have the geometry of our lab and give it to AI, and let it look at the literature and try to see like, can you have a customized recipe using my customers instruments? It works really well. So it's actually showing here like the in is always prompt and the task description. And alt can be the directly get from the customized instruments. So basically here shows a demo. Like I try to push my student to grow. My lab is growing this material for a long time. But this material, I try to push my student to, can you grow it? Can you grow it? And my students say, oh, Dr. Wang, I have to look at literature. It takes a few weeks or something. Then one day I just say, "Can you just try to use large angle model and try a recipe, and let's see whether it works?" Then it's working. I joke with a student in the future that you don't need to even look at literature. You can just follow what models tells you. So that saves me a few months. So I try to push my students. There's a lot of mental torture for myself. But again, here shows like it's the AI can... do the, try the industrial scale manufacturing. just with one shot, or we say there's just first trial. Okay, this is also something if you want to repeat something and you have just a new setup, and AI can save your time and speed up the discovery and your research. So, basically, here shows if we go one step beyond. We already see the material. And also we can grow the material. How about we directly build the actual chip So here image showing like not published yet. So we can combine this microscope. You can see the microscope seeing some materials. we grow on which is not ideal. We on purpose chosen on non-ideal material. You can see a lot of black dots actually defects. If I give this material to the industry partner, they will say this is crap. Okay, you can never grow any chips on top of that. But if we use our AI and let AI to build it, so what AI will do is just choosing the good area. So you can see a conventional way, we have uniform transistors, billions. Okay. But if we use AI, we can see every transistor are different. And now we are connecting AI with all the cleanroom manufacturing instruments, and we can build, like, it's actually... like actually build a chip So, yeah. [25:00] champion is not beating at scale okay So we can still have thousands of transistors. But we think it's a very good beginning. So basically, we also see by taking the LED AI control, the whole semiconductor manufacturing, it's increased yield. So originally this material can only have 65.6% which is crap, but build upon this low yield material, we can have a high yield, like transistor arrays, which is very basic chips. So, this 30%, around 30% increase, we think it will wholly, like, will change and disrupt the whole manufacturing, semiconductor manufacturing in the future. That's my belief. Okay, so in that case, So in that case, basically, we recently built a startup about a month ago, and we think using AI, to interrupt the semiconductor industry will be very interesting. We think AI native semiconductor manufacturing will be more and more important. We think we will try the Yale intelligence or lab to fab solutions, or we think of like maybe we want to build a small pilot line for the new 2D semiconductors. But again, like for everyone talking about today, so all your money pay to TSMC in the end, right? So in the end, it's how to build these chips actually takes the most, even though we talk about AI, we talk a data center and everything. Now, I'll give you an alternative solution, which is cheaper, okay? So again, like I will say that AI has already rise and models. Recently, a lot of people talk about RSI, recursive self-improvement. And then like we think like building the missing link. how to use AI to manufacture its own brain is very important. So this shows like a flywheel, like if AI can design the chip, can build the chip, can train the AI, chips can train the AI with the data in the middle. So it's actually become a, more efficient like flywheel. Or I will say like the model and the chips can finally combine to each other, similar to our biological brain. So for example, our minds and our brain are never separated. But currently, the model and the chips are kind of separated. So if we combine these chips and model by using model to build chips and using a chip to run the model, so I will say this is a new format of life. Maybe we can say it's like a new format of the silicon live format, whatever you call it. Okay? And I will say like this could be the next step if we have a conventional AI, And then we have RSI. And then this RSI, maybe we can call it a super intelligence. And recently, it seems like SI is more favorable than AI. But again, if we realize put the manufacturing into the building of the chips, and I will say this could be new intelligence. And because nobody has given a name of that, maybe I can try to give a name to bring it back to AI. I will call it autogenic intelligence. Let's go back to AI again to finish this closed loop. So again, like we think if we combine the minds and the brain or the chips and models together, it's a new. And the question is, where it goes? As a professor, I always think. So basically, I want to borrow some concept discovered in 1916, about more than 100 years ago. This concept is called constant sensation. So what does it mean is like people find no matter what kind of ocean animals are after life. Millions of years, it's become a crab. So what does it mean? Like if you have a fish, millions of years... crab, maybe like king crab. So if you have a turtle, medium of years it becomes like a blue crab, for example. Right. So it's a constantization is observing our nature, mother nature for already there for billions, millions of years. Okay. So in that case, I'm thinking like I give a question for you. So do you think the AI plus chips will be... diverge or converge. Oh, well, that would be a crab shape. of this silicon life format. All right? So, but maybe next time when you eat the crab, you should remind us, think about it, and maybe you are eating a very intelligent animal. So with that, I conclude my talk. Thank you very much.