Wu Tailin wants to use AI to accelerate nuclear fusion, and Shanghai state-owned capital has invested in the related project.

In the past, when people conducted simulations of fluids, mechanics or plasmas and designed devices based on relevant results, each iteration would take hours, days or even longer. With the empowerment of AI, this period can be shortened to the order of minutes or seconds. In the future, even if you give AI any target, it can design and manufacture machines that meet the corresponding functions, such as rockets, spacecraft, Mars rovers...

This cutting-edge field is called AI for Engineering. Enabling humans to solve energy problems with the help of AI and gradually move towards space is the dream of Wu Tailin, founder of AI for Engineering enterprise Force Engine.

ChinaVenture learned that Force Engine has announced the completion of an Angel+ round of financing of nearly 50 million yuan. This round of financing is jointly led by Shanghai Future Industry Fund under Shanghai State-owned Assets Investment and Shanghai Science and Technology Innovation Group, with Fudan Sci-Tech Innovation as the additional lead investor, Longxin Venture Capital, CGC Capital and other institutions participating in the follow-up investment, and Lighthouse Capital acting as the exclusive financial advisor. The funds will be mainly used for continuous R&D of generative simulation and intelligent control engines, device-level verification of benchmark scenarios such as nuclear fusion, and expansion of industrial customers in sectors including energy and high-end manufacturing.

The three lead investors, namely Shanghai Future Industry Fund, Shanghai Science and Technology Innovation Group and Fudan Sci-Tech Innovation, have distinct backgrounds of Shanghai state-owned assets and scientific and technological innovation. Considering that the first verification scenario of Force Engine is nuclear fusion, and Shanghai is precisely the highland of the domestic controllable nuclear fusion industry, the intention of this investment is not difficult to understand.

In the broad concept of AI for Science (AI4S), AI4Engineering can be regarded as an important branch oriented to engineering physical systems and industrial implementation. It inherits the capabilities of AI4S in rule learning and complex system modeling, and further extends AI to links such as simulation, state perception, design optimization and real-time control, finally forming a closed-loop delivery that runs through the full life cycle of engineering systems. In Wu Tailin's view, many problems in different fields of engineering are similar, such as multi-scale multi-physics field simulation and control, so a general platform should be built to serve all fields at the same time.

Multi-scale multi-physics field simulation has many technical difficulties, and this track is still in the stage of non-consensus. "Many investors are waiting and seeing, and those who agree with us recognize our ideas very much," Wu Tailin told me. When state-owned capital begins to invest real money to support the development of the industry, it also means that the AI4Engineering track, which sounds somewhat abstract, is entering a critical stage of technical scenario verification.

Shift from quantum computing to AI research

To understand what Force Engine is doing, we must first understand who Wu Tailin is.

Wu Tailin graduated from the School of Physics of Peking University with a bachelor's degree, obtained a doctorate in physics from the Massachusetts Institute of Technology, and then engaged in postdoctoral research in the research group of Professor Jure Leskovec of the Department of Computer Science of Stanford University. He is currently a distinguished researcher and assistant professor of the School of Engineering of Westlake University — this is Wu Tailin's academic resume, which can be regarded as the standard portrait of young scientists that VC institutions are keen to invest in at present.

He once put forward the new learning paradigm of "theory learning" and the method centered on AI Physicist with his doctoral supervisor Professor Max Tegmark, and made a series of pioneering achievements in enabling complex physical system simulation, control and design with generative AI.

But before becoming a scientist behind the "AI Physicist", his research direction was actually quantum computing.

In 2016, Wu Tailin, who was in the third and a half year of his PhD in the Department of Physics of MIT, suddenly faced a test: due to the sudden cut of funds, the laboratory fell into stagnation. The academic path was blocked, and most of the surrounding students chose to transfer to other laboratories to continue doing quantum computing experiments, but Wu Tailin made a different decision. It coincided with the time when AlphaGo defeated Lee Sedol, and AI entered the vision of more people. After a month of thinking, he decided to switch his research direction to AI to study AI + physics.

"It was equivalent to burning our boats behind us," he described the transition. For example, in order to carry out AI-related scientific research better, half a year before his internship at Google, he simply moved to Los Angeles and discussed with his mentor at Google every day. A year later, his research began to make progress, and by 2019, a number of his achievements were intensively released. From quantum computing to AI, Wu Tailin spent three and a half years, which was equivalent to completing a new doctorate from scratch.

His previous academic experience shaped his underlying understanding of the relationship between AI and physics. In his view, when AI wants to connect the physical world, it needs to model, simulate and control physical systems. People with a physics background can better understand what physical problems are important and what problems can be better solved by AI. During his postdoctoral period at Stanford, influenced by the campus atmosphere, Wu Tailin had already started to prepare for entrepreneurship. After joining Westlake University, he cooperated with industrial sectors such as nuclear power plants, 3C enterprises and fusion companies to learn more about the needs of the industry.

The opportunity for entrepreneurship came in 2026. The policy and industrial environment of AI4Engineering are forming a joint force: the "AI+" action of the State Council proposes to promote the integrated coordination of AI-driven technology R&D, engineering realization and product implementation, and the outline of the 15th Five-Year Plan lists nuclear fusion energy as one of the future industrial directions for key cultivation. Force Engine is exactly at the intersection of these plans.

Why does AI need to "understand" physics?

In the R&D process of engineering physical systems, Force Engine does not simply replace traditional CAE software. Instead, it takes numerical simulation and experimental data as the high-reliability training basis and integrates more physical priors, learns the evolution and conditional distribution of physical fields through methods such as neural operators and diffusion models, and completes rapid prediction and coupled generation in the reasoning stage. On this basis, it combines intelligent agents to realize the automatic completion of multi-stage complex tasks.

"Traditional solvers can achieve relatively high accuracy, but they are very slow. Simulating a physical process may take hours or even days. AI-based methods can compress the time from the order of hours or days to the order of seconds or minutes. In addition, the AI-native method we developed also has very high accuracy compared with traditional solvers," Wu Tailin explained.

At present, the company has achieved dozens to hundreds of times acceleration in complex physical simulation tasks such as fluid, oil reservoir and plasma. Some tasks can be compressed from the traditional hour or day-level calculation to the second level, and the model can be continuously calibrated through verification on real devices and industrial sites.

An ideal state is that humans only do the "head" and "tail" work, that is, define problems and provide acceptance criteria. AI completes all the intermediate work, and continuously optimizes in the process of interacting with the real world.

Verify the upper limit of technology with controllable nuclear fusion

Controllable nuclear fusion is the first high-difficulty verification scenario selected by Force Engine.

Magnetic confinement fusion involves high-fidelity simulation, state estimation and closed-loop control of complex plasma systems under extreme conditions, spanning macroscopic magnetohydrodynamics and microscopic turbulence scales. It is one of the most representative high-difficulty scenarios of AI4Engineering. In recent years, the academic community has used deep reinforcement learning to realize plasma configuration control on the TCV tokamak device, and realized active avoidance of tearing mode instability on the DIII-D device. AI is gradually entering the closed loop of actual fusion experiments.

"The problem of fusion is so important, and it is now on the verge of a breakthrough. I expect that the demonstration commercialization of fusion may arrive in 2040, and it can provide almost endless and low-cost energy for human beings including AI," Wu Tailin told me. More importantly, the strict requirements of fusion scenarios for model accuracy, real-time performance and safety can reversely temper the reusable physical modeling and control capabilities.

If complex multi-scale plasma systems can be well simulated in the fusion field, the same technology may be able to solve problems in aerospace and other fields.

This idea of "verifying the upper limit of technology with the most difficult scenarios and then migrating to complex engineering scenarios" will reuse underlying methods such as neural operators, generative modeling, safety control and multi-agent collaboration, and then carry out adaptation for the geometric structure, boundary conditions, working condition data and safety specifications of various industries. Relevant algorithms have been deployed in industrial scenarios in Saudi Aramco's 10 million grid oil reservoir simulation, and will also promote industrial cooperation in nuclear power multi-physics field simulation, 3C mechanical-thermal simulation, aerospace and other directions in the future.

In addition to cutting-edge technologies, we also talked about the choices in the process of academic research and entrepreneurship, as well as the understanding of intelligence from an AI founder with a physics background.

The following is the conversation between Liu Yanqiu and Wu Tailin:

Liu Yanqiu: Now a large number of people with physics backgrounds are entering the AI industry. When you switched from quantum computing to AI during your doctoral period, how did you make the choice at that stage?

Wu Tailin: Most other people transferred to other laboratories to continue doing experiments in the field of quantum computing. But I have done experiments for more than three years, which was equivalent to staying in a small dark room every day, adjusting many complex laser optical paths, and making various precision parts on machine tools. Most of the work was physical labor. I felt that this kind of life was not what I was good at and what I wanted.

It was 2016 at that time, and AI was just emerging. I spent a month thinking and decided to switch my research direction to AI. After making the decision, I told my supervisor, and he was also confused at first. But I was very firm, and later my supervisor gradually supported me, and introduced me to some collaborators in the AI field.

I found the associate supervisor Max Tegmark, and did AI+physics research with him. My supervisor also introduced me to a mentor at Google for an internship there. In order to do scientific research better, I moved my home to Los Angeles half a year in advance, lived next to him, and discussed with him every day to carry out scientific research. Continuous in-depth discussions put me in a state of flow, and I made a number of achievements. At that time, I really burned my boats behind me. I was already in the fourth year of my doctorate, which was equivalent to starting all over again. I spent a period of time and seized every possible opportunity.

Liu Yanqiu: So you only spent more than a year to complete the transition?

Wu Tailin: Yes, more than a year. In the first year after the transition, I didn't get any achievements. It was not until 2017 that I gradually got some results, and I also encountered various situations where my papers were rejected. It was not until 2018 and 2019 that my achievements broke out intensively.

Liu Yanqiu: From the perspective of a physicist, what is the difference between your understanding of "intelligence" and that of people with a pure AI background?

Wu Tailin: There is no unified definition of intelligence in the academic circle at present, and there are different opinions on what counts as AGI. But if I have to say it, the core of intelligence lies in: it can quickly adapt to the environment in any scenario, understand the operating mechanism of the environment, and efficiently achieve the goal after the target is given.

This is not a direct generalization, but a process. If you put an agent in a brand new environment, it will definitely not do well at the beginning, but it can quickly gain experience through continuous interaction with the environment, master the physical laws of this environment, and then complete tasks faster and better in this environment.

Liu Yanqiu: What you emphasize is the interaction with the real world.

Wu Tailin: Right. This is also the empiricist view of Rich Sutton, the father of reinforcement learning. If current large language models are only trained with static data, they will definitely hit a bottleneck. But if you can let the model continuously interact with the environment and accumulate experience from it, the amount of data is theoretically infinite.

Liu Yanqiu: For the company you founded, which positioning do you agree with more, AI for Science or AI for Engineering?

Wu Tailin: I started to do AI for Science in 2018. But now the more accurate name is AI for Engineering. They can be regarded as two circles with a large intersection, not a subordinate relationship. The common part is that AI is used for simulation, design, control and diagnosis. The difference is that AI for Science pays more attention to discovery, such as automated laboratories continuously iterating to discover new knowledge; AI for Engineering pays more attention to combining with the actual industry to solve their problems. One focuses more on scientific discovery, and the other focuses more on solving practical industrial problems.

Liu Yanqiu: Why did you choose nuclear fusion as the first verification scenario? This direction sounds very difficult.

Wu Tailin: There are several factors. The most essential one is that the problem of fusion is too important, and it is now on the verge of a breakthrough. I expect that it can be commercialized or demonstrated commercialized by 2040, and can provide almost endless and low-cost energy. Energy is the core driving force of the entire social development, and the future development of AI also needs a much larger amount of energy.

Moreover, fusion is one of the most representative high-difficulty scenarios of AI4Engineering. It spans multiple space-time scales, including the coupling of various complex physical processes such as macroscopic magnetohydrodynamics, microscopic turbulence, and fast particle dynamics, and the core and edge parts influence each other. The strict requirements of fusion scenarios for model accuracy, real-time performance and safety can reversely temper the reusable physical modeling and control capabilities. If we can simulate complex plasma systems in the fusion field, the same technology is likely to be able to solve problems in other engineering fields, such as aerospace.

Liu Yanqiu: Is the cost and expense of doing nuclear fusion research with AI high?

Wu Tailin: It is very, very low compared with building a physical device. Building a device costs at least hundreds of millions of yuan, and large devices cost billions to tens of billions of yuan. If we can carry out better and more detailed simulation in advance before or during the construction of the device, explore the interval of high-parameter stable discharge, and design the device based on the results, we can save a lot of money and speed up the progress.

Liu Yanqiu: What is still missing from the current known technical level to the future you want to reach?

Wu Tailin: There is definitely a long way to go. The key is the overall high-fidelity simulation of multi-scale and multi-physics field coupled systems. For example, for future burning plasmas, processes such as alpha particle self-heating, turbulent transport, and boundary heat removal will form a highly coupled dynamic system, and engineering systems such as magnets, heating, power supplies and diagnostics also need to operate in coordination for a long time. There is no very perfect method to achieve the overall simulation of such a complex multi-scale and multi-physics field system, but AI-based methods have greater potential to realize it.

Such multi-scale and multi-physics field problems are not only problems in the fusion field, but also important problems in other fields. For example, combustion has the multi-scale coupling of chemical reaction and fluid; materials have the cross-scale coupling from atoms and molecules to mesoscopic defects and macroscopic properties; life sciences have the multi-scale coupling from molecules, cells to tissues and organs. These are all very difficult problems.

Liu Yanqiu: I saw that you shared in a previous speech that your dream is to let human beings go to space. What is the relationship between this dream and what you are doing now?

Wu Tailin: The core lies in increment. If we only redistribute the existing stock, it is difficult to achieve sustainable development. Only a large enough increment can alleviate contradictions and bring more development opportunities. The resources and energy on the earth are limited, and the next era may be the space era. There is far more energy and more resources in space than on the earth.

Fusion can provide more efficient energy. If fusion technology is solved in the future, rockets in space can use fusion energy instead of chemical energy. The maximum speed of chemical energy rockets launched from the earth is only more than ten kilometers per second, which is too slow in the universe. For example, it takes several months to reach Mars, more than ten years to reach Pluto, and tens of thousands of years to reach the nearest star. When fusion technology is solved, it is expected to build rockets based on fusion propulsion, whose theoretical speed can reach a few percent of the speed of light, hundreds of times that of chemical energy rockets. In this case, it only takes a few hours to reach Mars, a few days to reach Pluto, and about 100 years to reach the nearest star. It is equivalent to that traveling anywhere in the solar system in the future is like taking a ship on the earth now. Human beings will gradually become a solar system civilization, bringing hundreds of millions of times of increment in the future.

Liu Yanqiu: What is the source of this space dream? Science fiction novels?

Wu Tailin: Of course,