Controlled nuclear fusion is one of the most ambitious engineering goals. In tokamak reactors, the physics of the edge plasma and divertor determines how heat and particles are managed, setting thermal fluxes, target conditions, and the onset of detachment. Accurately predicting these quantities is crucial for operating current and future devices, but edge simulations based on codes like SOLPS-ITER require hours or days, making them impractical for parameter scans, optimization, or real-time control.
To overcome this limitation, neural surrogates have emerged as an ultrafast alternative. A promising approach is the cycle-consistent model, which combines a conditional U-Net forward model with an optimization-based inverse method that uses the frozen forward network. The forward model maps five control parameters (gas puff rate, heating power, etc.) to two-dimensional plasma fields (density, temperature, fluxes). The inverse method recovers the control parameters from observations, and cycle consistency —verifying that applying forward and inverse models successively returns the original values— acts as a self-supervised quality check, needing no ground-truth labels during inference.
Results show normalized root-mean-square errors below 2.6% and Pearson correlations above 0.95 for all fields. Additionally, an ensemble of multilayer perceptrons predicts electron temperature and density profiles at the outboard midplane and divertor targets, with uncertainty estimates that flag where more simulations are needed. Cycle-consistency regularization raises the average cyclical R² from 0.59 to 0.99 without degrading forward accuracy, and enables recovery of the gas puff rate with a Pearson correlation of at least 0.97. With approximately 4 million parameters, the model produces full 2D predictions in milliseconds, five to six orders of magnitude faster than SOLPS-ITER.
This machine learning paradigm, where forward and inverse models reinforce each other, has direct applications in the business world. At Q2BSTUDIO, a software and technology development company, we apply similar principles of artificial intelligence and machine learning to create solutions that turn data into decisions. For example, we develop custom applications that integrate cyclical predictive models, enabling companies to automatically validate the consistency of their predictions and recover input variables from observed outcomes — essential in industrial and monitoring environments.
The infrastructure supporting these models must be scalable, secure, and efficient. That is why we deploy our systems on cloud AWS/Azure, ensuring elasticity and high availability for AI workloads. Cybersecurity is equally critical: our cybersecurity services protect both sensitive simulated plasma data and machine learning models, preventing information leaks and adversarial attacks. Furthermore, we integrate Power BI to visualize predictions and uncertainties in real time, facilitating data-driven decision-making.
One of the most exciting innovations in this field is the use of AI agents that, based on the frozen inverse model with a k-d tree warm start, achieve database completion rates above 95%, compared to roughly 30% outright failures when cold-started. This warm start concept is directly transferable to industrial process optimization systems, where our AI agents can efficiently explore parameter spaces to find optimal configurations without restarting full simulations.
In summary, cycle-consistent neural surrogates not only revolutionize tokamak plasma simulation but also illustrate an AI design pattern that can be generalized to any domain requiring speed, reliability, and inversion capability. At Q2BSTUDIO, we are ready to help companies implement these architectures, combining expertise in custom software development, cloud computing, cybersecurity, business intelligence, and autonomous agents. The fusion of plasma physics knowledge with cutting-edge software engineering opens possibilities that previously seemed unattainable, both in energy research and in corporate digital transformation.
