FEL: Musk redefines EUV and opens up space chip manufacturing Elon Musk's recent statement "FEL FTW" has brought Free Electron Laser (FEL) back into the semiconductor industry's vision. It is easy for the market to understand this as' Musk wants to create his own ASML ', but this may not be the most accurate framework. FEL is not a completely new lithography principle, but a new EUV light source. If this route eventually enters Terafab, it may truly challenge not the entire ASML EUV scanner, but a fundamental architecture formed by the EUV industry over the past two decades: each scanner is equipped with its own LPP light source. A more radical possibility for FEL is to transform EUV light sources from machine level components to fab level infrastructure. 1. From tin droplets to electron beams The core wavelength of ASML EUV today is 13.5 nm. The method for generating this light is very complex: high-power lasers hit tiny tin droplets flying at high speeds tens of thousands of times per second, instantly turning the tin into high-temperature plasma, and then collecting 13.5 nm EUV from the plasma radiation. This is LPP, Laser Produced Plasma. There is a huge engineering problem here: tin droplets become ultra-high temperature plasma before EUV can be produced. But during the process, it is necessary to deal with tin debris, collector pollution hydrogen plasma、 Mirror lifespan, cleanliness, heat load This itself is a huge technological barrier accumulated by ASML/Cymer over more than 20 years. ASML has even been collaborating with university research teams for a long time to address issues related to hydrogen plasma environment, tin cleaning, and component lifespan. FEL is completely different. The electron gun generates an electron beam, which is accelerated by an accelerator to near the speed of light, and then passed through an accelerator composed of periodic magnetic fields. Electrons constantly oscillate and gradually form microbunching, ultimately producing highly coherent electromagnetic radiation. By designing electronic energy and an inverter, 13.5 nm EUV can be directly generated. So the two routes solve the same problem: how to generate a large and stable amount of 13.5 nm photons. The answer to LPP is "laser+tin plasma"; The answer to FEL is "relativistic electron beam+accelerator", which can bypass the huge engineering technology barrier of tin plasma and increase EUV photon throughput. 2. The biggest bet for FEL is power The EUV light source power directly affects the exposure speed, photoresist dose, and scanner throughput. LPP has developed from a few watts in the early days to several hundred watts and is currently advancing towards the kilowatt level. But it is becoming increasingly difficult to continue increasing power. More laser energy means more extreme tin plasma, greater collector heat load, more severe pollution, and more complex system reliability issues. The long-term temptation of FEL lies in EUV ranging from several kilowatts to even 10 kW. Public FEL-EUV studies have already discussed this magnitude. It is necessary to distinguish between "theoretical/engineering design goals" and "semiconductor mass production capabilities" here: today there is no stable mass production chip for a 10 kW FEL in advanced wafer fabs. But if this type of light source is ultimately realized, it may not only provide "a little more light than LPP", but also a technology route of different orders of magnitude. The advantage of FEL is not only power. It can provide highly coherent, narrow spectrum, and controllable polarization radiation without the problems of tin droplets, tin debris, and plasma pollution like LPP. Some polarization modes may even improve the imaging contrast and process window of extremely small sized graphics. In other words, FEL is not just a larger light bulb, but may be a stronger, cleaner, and more controllable EUV photon engine. The cost is also very obvious: it has replaced the "tin drop hell" with the "particle accelerator hell". Electronic gun RF、 Acceleration chamber, magnet undulator、 Ultra high vacuum beam control、energy recovery、 Cooling and control systems, each of which can become a new engineering bottleneck. The biggest advantage of LPP today is not that its physical principles are beautiful, but that it has undergone billions of hours of industry learning and large-scale wafer production verification. The biggest problem with FEL is not whether it can generate EUV - the answer is already yes - but whether it can turn a particle accelerator into a semiconductor grade manufacturing infrastructure. 3. Terafab changed the economics of FEL If the goal is to manufacture a device and sell it to TSMC, the volume and complexity of FEL are huge disadvantages. But if the goal changes from 'building a lithography machine' to 'designing a super large AI wafer factory from scratch', the problem is completely different. The traditional architecture is roughly as follows: Scanner A + LPP A Scanner B + LPP B Scanner C + LPP C Scanner D + LPP D A more radical architecture for FEL may be: 10 kW Central FEL → EUV Beam Distribution → Scanner A / B / C / D / E…… This means no longer trying to fit a particle accelerator into a lithography machine, but directly building the entire wafer fab around the particle accelerator. FEL has transformed from a component of a single machine into a utility for the entire fab, similar to electricity, ultrapure water, and central gas supply systems. This may be the most imaginative aspect of the combination of Terafab and FEL. The huge accelerator is no longer a fixed cost that a single scanner must bear, but can be shared among ten or even more scanners. A set of FELs worth hundreds of millions of dollars or even higher, if able to stably provide several kilowatts of EUV to a large number of scanners, may have completely different unit exposure costs. 4. In the long run, FEL is even more suitable as a "space semiconductor factory" than traditional lithography machines FEL itself has several natural characteristics suitable for space. Firstly, high vacuum. Electron beams and EUV already require a vacuum environment, and space is highly compatible with this system at least in the external environment. Next is the scale. Ground construction of a one kilometer level accelerator means isolation of land, buildings, tunnels, foundations, and vibrations; If large-scale rail structures can be deployed in a low-cost modular manner in the future, 'being very long' may not necessarily be the most fatal problem. The third is the mechanical environment. There are no traditional earthquakes or ground vibrations in space, Therefore, rail FEL is easier to imagine than rail advanced wafer fabs. When low-cost heavy-duty transportation, autonomous robot construction and maintenance, and highly autonomous AI fabs mature simultaneously, the "Orbital FEL Terafab" is more likely to transform from a sci-fi engineering concept into a truly calculable ROI industrial solution. 5. RSI is the most noteworthy variable in FEL This may be the most important factor in the whole matter, and it could also be the ultimate factor that made Musk make up his mind. AI is helpful for both mature LPP and FEL, but the marginal value may be very different. LPP has undergone over 20 years of extreme engineering optimization. The remaining issues are increasingly focused on collector lifespan, heat load, tin contamination, material durability, photoresist, and precision manufacturing. These issues can be optimized by AI, but are increasingly limited by real-world materials and manufacturing capabilities. FEL still has a huge unexplored design space. Electron gun、beam energy、emittance、RF phase、bunch compression、magnetic optics、undulator geometry、energy recovery、beam distribution The feedback control forms an extremely complex high-dimensional system, and a significant portion of it can be simulated in a computer using Maxwell equations, particle dynamics, and accelerator physics. This happens to be the field where research AI is most likely to generate huge leverage. In the past, engineers proposed a plan, ran simulations, modified parameters, manufactured prototypes, and then tested them. In the future, AI may generate 100000 architectures, use a surrogate model for initial screening, then search for Pareto frontiers through high-precision physics simulations, and finally only manufacture the most promising dozens of designs. The value of AI is not only in design. One of the biggest industrial issues with FEL is stability: small drifts in RF phase, beam energy, magnet current, temperature, and beam position can ultimately become CD variations on the wafer. FEL is essentially a massive real-time high-dimensional control system, and this is where advanced AI control is most likely to come into play. The constantly enhancing AI capabilities will turn the FEL development and construction cycle into: AI design for FEL Gen 1 → manufacturing → automatic collection of beam/wafer data → AI finding simulation to reality gap → modification of accelerator, accelerator, and control → FEL Gen 2 → more data → FEL Gen 3. This also explains why FEL may be worth reassessing in the RSI era more than ever before. Without AI, challenging LPP with over 20 years of industrial learning curve is an extremely difficult proposition. With the increasingly powerful research AI, the vast number of unresolved problems in FEL precisely belong to the "simulation+search+optimization+control" space that AI is most prone to compress. Therefore, Musk is optimistic about FEL, and the most important underlying logic is actually: When the demand for AI chips expands by an order of magnitude, research AI begins to participate in hardware design, and wafer fabs move towards Terafab scale, the FEL architecture is more likely to be the optimal solution compared to today's EUV architecture with independent scanners and independent LPP sources as the core!