A Chinese startup has identified the most expensive bottleneck in fusion energy, thanks to AI.

      For many years, fusion energy has been hailed as the ultimate clean energy solution. This process, which fuels the sun, has the potential to generate vast amounts of energy without the carbon emissions linked to fossil fuels. Scientists have dedicated generations to attempting to replicate it on Earth, convinced that achieving large-scale fusion could significantly transform the future of global energy. However, the challenge lies in the fact that fusion is incredibly complex, both scientifically and economically. The construction and testing of experimental reactors require immense financial resources, and advancements often come through a frustrating cycle of experimentation. Researchers formulate a theory, develop the necessary hardware, collect data, modify the design, and repeat this process, which can take years. A Chinese startup named VeloAlpha now believes that artificial intelligence might help disrupt this cycle.

      Founded this year by fusion scientist Xie Huasheng, the Beijing-based VeloAlpha is creating FusionAlpha, a simulation platform that enables researchers to test fusion reactor designs digitally before investing in costly physical experiments. While it may not have the thrill of a large reactor producing endless clean energy, if VeloAlpha's technology fulfills its potential, it could address one of the fusion industry's most expensive and persistent challenges.

      The fusion industry's impossible triangle

      Xie notes that fusion researchers have long faced an uncomfortable dilemma. The most advanced simulation software currently available can accurately model plasma behavior. Plasma, the superheated, electrically charged gas that powers fusion reactions, is notoriously challenging to manage, and understanding its behavior is vital for designing a functional reactor. The downside is that these simulations require significant computing power and can take considerable time to execute.

      Conversely, newer AI-driven systems can conduct calculations much more quickly. However, researchers often remain cautious with these tools because of their potential reliability issues and their limitations in extrapolating beyond the data used for training. Additionally, there are simplified physics models that are computationally efficient but often too basic to effectively guide the design of next-generation reactors. Xie describes this as the "impossible triangle" of fusion software: speed, accuracy, and predictive capability. Historically, researchers have had to compromise one aspect to gain another. VeloAlpha's business model is centered around the belief that this trade-off can be eliminated.

      The company asserts that advancements in artificial intelligence, along with new mathematical approaches, can significantly speed up simulations without compromising the underlying physical principles. Xie claims that some elements of FusionAlpha can operate anywhere from 100 to 10,000 times faster than current leading fusion codes while keeping benchmark errors below 5%. These claims still require independent verification, but if confirmed, they could represent a substantial advancement for the sector.

      The high costs of building a star

      To grasp the importance of this software, it's necessary to understand the goals of fusion researchers. Fusion occurs when the nuclei of light atoms collide and combine, releasing vast amounts of energy. This is exactly what occurs inside stars. Replicating such conditions on Earth necessitates heating fuel to temperatures exceeding those of the sun’s core, producing plasma that must then be contained and stabilized long enough for fusion reactions to happen. Most researchers pursue this using machines known as tokamaks—massive, doughnut-shaped devices that utilize powerful magnetic fields to contain plasma. Others are exploring alternative methods such as stellarators, linear devices, and laser-driven fusion systems.

      Each design presents its own engineering challenges. Researchers need to determine how to sustain reactions, endure extreme heat, manage radiation, secure fuel supplies, and ultimately produce electricity affordably enough to compete with current energy sources. Addressing these issues is not cheap; a single experimental facility can cost hundreds of millions or even billions of dollars. Even minor design modifications often necessitate extensive testing and validation. This is why simulation software has become increasingly crucial. The more accurately researchers can predict outcomes prior to constructing hardware, the less financial waste they incur on unsuccessful endeavors.

      Fusion’s EDA moment

      Xie likens FusionAlpha to electronic design automation (EDA) software, which revolutionized the semiconductor sector. Modern chip manufacturers do not build a physical processor every time they wish to explore a new concept. Instead, they rely on advanced software tools to model, simulate, and optimize designs before sending them to fabrication facilities. Without EDA software, the pace of innovation in semiconductors would be significantly slower.

      VeloAlpha believes the fusion field is nearing a similar inflection point. Instead of relying predominantly on physical experimentation, future fusion enterprises could utilize advanced simulation platforms to virtually test thousands of design variations, identify promising approaches, and substantially cut development costs. Thus, the next generation of fusion reactors might be developed twice: first in software, then in metal.

      Why timing is crucial

      The emergence of this startup is particularly noteworthy for China's fusion sector. For years, fusion research was primarily driven by governments and national labs, but this is beginning to shift. China has recognized nuclear fusion as a critical future industry, placing it alongside areas such as quantum computing, embodied AI, biomanufacturing, brain-com