Machine Learning-Based Surrogate Modeling for Enhancing the Magnet Design of Future Tokamaks

Abstract

This paper presents a surrogate modeling approach based on Machine Learning (ML) techniques to enhance the magnet design of the European DEMOnstration Fusion Power Plant (EU-DEMO). Despite efforts to minimize the reactor size, its dimensions remain considerable, necessitating advanced tools for efficient design exploration. To support the design process, pre-sizing codes such as the MAgnet Design Explorer Algorithm (MADE) were developed at different detail levels. MADE systematically scans all possible configurations, identifying those that satisfy electro-magneto-mechanical constraints and providing a dataset consisting of a cloud of points, each corresponding to a feasible magnet design and related layout. This work builds upon MADE-generated data to train Artificial Neural Networks creating a surrogate model capable of directly identifying optimal configurations for identifying the optimal solution within the cloud. The flexibility of the proposed ML approach allows not only for solving the forward problem (predicting outputs from given inputs) but also for addressing the inverse problem (deducing the input parameters needed to achieve desired output characteristics) or freely mixing input and output parameters. This flexibility enhances the adaptability of the method, enabling a more targeted exploration of the design space. The model has been applied to the design of the central solenoid modules, but can be extended straightforwardly to other magnet systems, such as toroidal and poloidal field coils. Furthermore, the surrogate model significantly accelerates the pre-dimensioning phase, reducing the range of configurations to be investigated and facilitating rapid iterations between engineering and physics teams. By providing fast and accurate predictions, the ML approach streamlines the design process, making it a valuable tool for the design of superconducting magnets for tokamak reactors like EU-DEMO.