Sensationelle Meldung in Science: KI hat einen „erfolgreichen“ Fusionszündvorgang vorhergesagt.

What? Artificial intelligence (AI) can actually predict the success rate of fusion ignition and has been published in the authoritative scientific journal Science!

This sounds a bit like a plot from a science fiction movie, but it has really happened -

This generative machine learning model, constructed by the Lawrence Livermore National Laboratory in the United States, predicted the result of a fusion ignition experiment at the National Ignition Facility (NIF) in the United States with a probability of over 70%, indicating "successful ignition." Here, ignition refers to the situation where the energy generated by fusion exceeds the laser energy used in the experiment.

Paper link: www.science.org/doi/10.1126/science.adm8201

This discovery may provide guidance for future researchers conducting inertial confinement fusion (ICF) experiments. ICF experiments use high-energy lasers to compress and heat hydrogen isotope capsules to initiate nuclear reactions that produce fusion energy, which is an efficient energy solution in the era of energy shortage. Imagine aiming the most powerful laser system on Earth at a tiny hydrogen capsule and making it release huge amounts of energy, just like mastering a miniature sun.

The researchers said that having a "successful prediction model" will inspire fusion energy researchers, help them adjust experimental designs, and determine whether increasing laser energy and other variables in the future can improve fusion output and efficiency.

Why is fusion ignition prediction important?

The ICF project aims to use the world's largest laser system, the NIF, to compress and heat a millimeter-sized capsule containing hydrogen isotopes deuterium and tritium (D-T). During the compression process, nuclear reactions in the D-T fuel will release fusion energy.

Computer simulations of ICF performance are very important in NIF experimental design. The NIF conducts about a dozen ignition experiments each year, and researchers must rely on these simulations to optimize experimental designs. To ensure the success of these efforts, it is necessary to develop prediction models that can accurately estimate observable target quantities before the experiment is carried out.

However, any actual simulation must reduce computational costs through simplified assumptions, which usually leads to a deviation between simulation predictions and experimental observations.

To reduce this deviation, researchers usually adjust the uncertain parameters in the model during the "post-processing" stage to match the experimental measurement results. The adjusted model is then used in the "pre-processing" stage for small-scale extrapolation to provide guidance for future experiments.

Although the manual "post-processing" parameter adjustment method has proven effective in small-scale extrapolation within the design space, it is still necessary to develop computer models with real predictive capabilities to support decision-making when upgrading laser systems or designing future high-yield facilities, especially when these facilities will operate under significantly different experimental conditions. Quantifying these uncertainties is particularly difficult due to the scarcity of experimental data, the vast design space, and the high cost of simulation calculations.

The prediction model proposed by the research team this time provides a promising method for ICF experimental prediction modeling and a new framework for the development of driving models for other complex systems.

How does AI achieve fusion ignition prediction?

In this work, the research team described a joint information model based on deep learning, which combines simulations and experiments and takes into account multiple sources of uncertainty.

The research team's goal is to provide quantitative predictions of fusion yield and other key diagnostic features before the experiment, while considering uncertainties, by combining previously collected NIF data, high-fidelity physical simulations, and expert knowledge. The model can adapt to modified designs, serve as a decision-making tool for future experiments, and provide robustness indicators in design optimization studies.

Figure | Workflow for predicting the variability of proposed experiments.

The prediction model combines a large simulation database, Bayesian analysis, and transfer learning techniques in machine learning to construct a statistical model based on both experimental data and simulation data. The model is first constructed based on previous NIF experiments using a statistical model based on both simulations and experiments. The model provides a series of input conditions for the model uncertainty and inter-experiment variability observed in a series of NIF experiments. Subsequently, these input conditions are applied to the proposed design of future experiments to generate a distribution of expected results based on previous experiments.

This study integrates a number of improved techniques for data-based ICF experimental prediction modeling, including a large-scale simulation database generated by an advanced high-performance computing (HPC) workflow, Bayesian post-processing analysis, and transfer learning techniques in machine learning.

As an extension of the previously published Bayesian post-processing analysis, the model quantifies the variability in a series of nearly repeated experiments conducted between 2021 and 2022. This "variability model" provides a distribution of expected variability for the on-site conditions of the NIF Hybrid-E experiment. These distributions are propagated forward through the machine learning model of the design being tested to predict the performance variability of future experiments.

In September 2022, the NIF conducted an ICF experiment using 2.05 megajoules (MJ) of laser energy for the first time, which was an increase compared to the previous 1.9 MJ of laser energy. This design achieved an output of over 1 MJ and laid the foundation for further performance improvement.

In the expected second experiment driven by 2.05 MJ of laser energy, one week before the successful ignition experiment in December 2022, the researchers used this method to predict that this design had a 74% probability of exceeding the break-even yield, significantly higher than any previous design. The actual results of the experiment were completely consistent with the predicted confidence interval, and other observable experimental quantities also met expectations.

In addition, for the subsequent repeated experiment of the design in December 2022, the results were consistent with the predicted variability distribution. The accurate prediction of this variability by the model was verified by the close consistency between the experimental results and the predicted confidence interval.

Figure | Relationship between the main yield (performance indicator) and DSR (percentage) (confinement indicator).

More than just fusion ignition prediction

The research team made a quantitative and physically meaningful prediction for a controlled fusion experiment, which achieved a target gain >1.

The prediction model takes into account the inevitable variability in on-site experimental conditions, including fluctuations in laser transmission and capsule quality, uncertainties in input conditions caused by experimental measurement uncertainties, and intentional design changes in upcoming experiments.

By combining Bayesian analysis of previous experiments with transfer learning, the surrogate model for the new design can be efficiently trained, enabling the prediction of the expected result distribution of upcoming experiments within a few days.

The prediction model takes into account the inevitable variability in on-site experimental conditions, including fluctuations in laser transmission and capsule quality, uncertainties in input conditions caused by experimental measurement uncertainties, and intentional design changes in upcoming experiments. By combining Bayesian analysis of previous experiments with transfer learning, the surrogate model for the new design can be efficiently trained. They were able to predict the expected result distribution of upcoming experiments within a few days.

This method provides an opportunity for optimizing design decisions in future NIF experiments under credible, data-driven uncertainties. It is not limited to ICF and can also be applied to other research fields that require scientifically based extrapolation to determine new configurations of complex engineering systems.

Overall, this work proposes a prediction modeling method under conditions of scarce data, and its application scope goes far beyond fusion ignition.

This article is from the WeChat official account "Academic Headlines" (ID: SciTouTiao), author: Xiaoyang. It is published by 36Kr with permission.