[AI Library] 20 Weather Forecasting and Nuclear Fusion
Demis Hassabis, Father of Google's Artificial Intelligence
Part 7. AI for Science
20 Weather Forecasting and Nuclear Fusion
Kim Kyung-ran, Kim Kyung-jin
When DeepMind was founded in 2010, this sentence was written on a whiteboard. Fifteen years later, it has left the lab and become reality, transforming the sky above our heads, the energy beneath the ground, and the unknown world of matter. WeatherNext 1) The roar of supercomputers and the silence of AI: compressing a hundred-year dream into one minute In the fall of 2023, the massive server room at the European Centre for Medium-Range Weather Forecasts (ECMWF) in Reading, England, was humming as usual. At this institution, home to the world's finest meteorologists, a basketball-court-sized supercomputer was running millions of processors to calculate the movements of Earth's atmosphere.
The 'Numerical Weather Prediction (NWP)' model they used had been the standard method for forecasting weather for the past hundred years. This approach breaks the atmosphere into small segments, applies the Navier-Stokes equations, a notoriously difficult problem in fluid dynamics, to each grid cell, and repeats differentiation and integration. It consumed enormous amounts of power and took several hours to predict the weather ten days out. At the same moment, a very different scene was unfolding at the Google DeepMind office in London.
Remi Lam and his research team were sitting in front of a single desktop computer. When they pressed the Enter key, a ten-day global weather forecast appeared on the screen in just one minute. The AI model was called 'GraphCast.'
This AI predicted hurricane paths more accurately than results produced by supercomputers costing hundreds of billions of won after hours of calculation. This was not a matter of speed alone. It marked a break with the old method.
'If the traditional method is a deductive approach that calculates atmospheric physics laws one by one, our AI is an inductive approach that learned atmospheric patterns on its own by studying forty years of weather data. Just as a seasoned captain can sense a storm by looking at the shape of the clouds, AI reads the flow of the Earth intuitively, without equations.' Through 2024 and 2025, this
technology evolved under the name 'WeatherNext.' 'WeatherNext 2,' unveiled in late 2025, was eight times faster than the original GraphCast and showed overwhelming performance in predicting extreme weather events such as localized heavy downpours and sudden temperature swings. The reason this innovation matters is that it intersects with the defining crisis of our time: climate change. In an era when extreme weather has become routine, accurate flood warnings that ring like alarms can save thousands of lives.
Traditional supercomputer models leave a carbon footprint by consuming massive amounts of energy. WeatherNext, by contrast, can produce national-scale forecasts using only the power of a home computer. Hassabis has pointed to this as an example of DeepMind's engineering philosophy: 'Technology built to save the planet must not destroy it in the process.' There was pushback from the meteorology community. The criticism that 'How can we trust predictions from a black-box AI that knows nothing about the laws of physics?' remains valid.
But DeepMind responded with a hybrid model called 'NeuralGCM.' Large-scale atmospheric flows are calculated using traditional physics models, while complex, fine-grained domains like clouds and turbulence are handled by AI. This embodies what Hassabis has always emphasized: 'the combination of human knowledge and AI intuition.'
WeatherNext's technology has been integrated into Google Search, becoming an everyday tool that helps billions of people around the world decide whether to bring an umbrella. The dream of Lewis Fry Richardson, who tried and failed a hundred years ago to calculate the atmosphere by hand, has finally been realized through AI. AI in Nuclear Fusion Plasma Reactors 1) How to hold jelly at 100 million degrees with your fingers If weather forecasting is about reading patterns in the sky, nuclear fusion is about recreating the heart of the sun here on Earth.
Hassabis had dreamed of traveling through space since childhood, and he knew that nuclear fusion was the only energy source that could make that dream a reality. But nuclear fusion had a fatal challenge: plasma hotter than 100 million degrees had to be suspended in midair inside a reactor called a tokamak, never touching the walls.
At the Swiss Plasma Center (SPC) at the Swiss Federal Institute of Technology in Lausanne (EPFL), researchers had struggled for decades to control the plasma dancing inside the tokamak. Plasma is as unstable as a living snake; even the slightest miscalibration of the magnets' fields causes it to collapse in an instant. The conventional control method required humans to pre-calculate and input thousands of variables, making it practically impossible to attempt complex new plasma configurations. Hassabis dispatched DeepMind's reinforcement learning team to the facility.
Their mission was straightforward: 'Teach the AI how to play with plasma.' DeepMind's AI went through millions of rounds of trial and error in simulation. At first, it kept crashing the plasma into the walls and extinguishing it. But just as AlphaGo taught itself brilliant moves in Go, the AI learned to stabilize the plasma by making fine adjustments to nineteen magnetic field coils ten thousand times per second.
Then came the decisive day: the AI was given control of the actual tokamak. The results arrived. The AI shaped the plasma freely into not just the standard round doughnut shape but also geometrically complex forms such as a snowflake shape and a droplet shape, exactly as the researchers had requested.
What stunned the physics community was that the AI stably maintained a plasma configuration with 'negative triangularity.' This shape was a kind of Holy Grail that had existed only in theory, because it could reduce heat damage to the reactor walls while simultaneously increasing energy efficiency.
Inside a nuclear fusion reactor, there is an ultra-hot gas exceeding 100 million degrees, known as plasma. This plasma is too hot to be contained by any material, so it is suspended in midair and confined by a powerful magnetic field. The cross-sectional shape of the plasma, as seen from the side, is the critical factor that determines the reactor's performance. Until now, nearly every fusion device in the world has shaped its plasma cross-section like an uppercase letter D. This outward-bulging shape is called 'positive triangularity.' But this shape comes with a problem.
Energy periodically erupts from the plasma's edge in explosive bursts, repeatedly hammering and eroding the reactor's inner walls. This was a challenge that had to be solved before commercial fusion could become a power plant. Physicists had long known of one theoretical solution.
Flip the D-shape sideways so that the convex side faces inward. This is called 'negative triangularity.' In calculations, the energy burst phenomenon naturally disappeared in this configuration, and
at the same time, the efficiency of confining energy remained high. It was the answer that caught both rabbits at once: reducing wall damage while also improving performance. But this was purely theoretical.
Stably maintaining this inverted plasma shape inside an actual reactor was a problem of an entirely different order, and no one had succeeded in decades. That is why this configuration had been called the 'Holy Grail' of nuclear fusion. Everyone knew it existed, but no one had been able to grasp it.
And now, that Holy Grail had been realized in the laboratory. The reason the physics community was shocked by the news that a plasma shape previously existing only in theory had been stably created and maintained lies precisely here. A thread had finally been seized that could overcome one of the greatest engineering barriers blocking the commercialization of fusion.
This achievement graced the cover of Nature in 2022, and in December 2025, it led to a partnership with the American fusion startup Commonwealth Fusion Systems (CFS). DeepMind agreed to transplant its AI control technology into SPARC, the commercial fusion reactor that CFS is building. It was the decisive moment when scientific curiosity in the laboratory crossed over into the industrial reality of producing actual electricity.
For Hassabis, this project is the perfect example of how 'intelligence' can break through the limits of the 'physical world.' AI accomplishes control at 0.0001-second intervals that are impossible for human reflexes. This is not just a technology for generating electricity; it is the key for humanity to move beyond energy scarcity toward 'Radical Abundance.' Hassabis often says:
'If we are to solve climate change and venture into space, we need infinite, clean energy. AI will be the key that opens that door.' The Possibility of an Autonomous Scientist (AI Scientist) 1) The sleepless laboratory: in 2026, the pace of science changes On one side of Hassabis's office, biographies of great scientists are always on the shelf.
Darwin, Einstein, Turing... He feels awe at the discoveries they devoted their lifetimes to, yet at the same time, a sense of frustration. 'A human scientist needs to sleep, needs to eat, and can be trapped by bias. What if there were a scientist who never tired, held no biases, and had read every paper ever published?' This question
took concrete form in 2025 when Google DeepMind announced 'AI Co-Scientist.' This is not a simple assistant that answers questions like ChatGPT. When a researcher sets a goal such as 'Find out how this protein interacts with cancer cells,' AI Co-Scientist independently searches tens of thousands of related papers, formulates hypotheses, and designs experiments to test those hypotheses. Hassabis's vision does not stop there.
He announced plans to open the world's first 'Automated AI Science Lab' in the United Kingdom in 2026. There will be no one in a white lab coat in this facility. Instead, robotic arms will ceaselessly move pipettes, synthesize materials, and analyze results under microscopes.
It sounds like a scene from a science fiction movie. The 'Gemini' model serves as the brain, issuing commands such as 'Let's synthesize a new superconductor candidate material,' and robotic arms immediately carry out experiments across hundreds of different mixing ratios. When results come in, the AI analyzes the data in real time and makes judgments like 'Failed. Adjust the mixing ratio by 0.5 percent and retry.' Experiments that would take a human researcher a week are completed here in a day, or even a few hours. This project is part of a five-billion-pound partnership with the British government, focused on discovering next-generation battery materials, high-efficiency solar cells, and drug candidate compounds.
Isomorphic Labs, the drug development subsidiary that Hassabis founded, is using this automated system to accelerate the era of 'digital biology.' At this point, Hassabis faces a peculiar tension: 'Then what should human scientists do?'
He argues that AI will not replace humans but promote them to the role of 'conductor.' 'AI will take on the hard labor of experimentation and data analysis, and humans will focus on the creative role of deciding what questions to ask and what research is worth pursuing.' The automated laboratory of 2026 is the grandest and most tangible answer to the thought Hassabis had at age twelve, sitting before a chessboard: 'Can't this intelligence be used for something better?'
Just as AlphaGo broke the conventions of Go, the AI scientist is now poised to shatter the limit that has governed humanity for centuries: the speed of discovery. This may be the democratization of scientific discovery, and the beginning of a 'singularity' in which the total sum of human knowledge explodes exponentially. Interior of a nuclear fusion tokamak reactor
Kim Kyung-jin
Attorney · Former Member of the National Assembly · AI Policy Researcher
© 2026 Kim Kyung-jin. All rights reserved.