FusionMAE, a self-supervised pretrained model to optimize and simplify diagnostic and control of fusion plasma

In magnetically confined fusion device, the complex, multiscale, and nonlinear dynamics of plasmas necessitate the integration of extensive diagnostic systems to effectively monitor and control plasma behaviour. The complexity and uncertainty arising from these extensive systems and their tangled interrelations have long posed a significant obstacle to fusion energy development. In this work, a self-supervised model, fusion masked auto-encoder (FusionMAE) is pre-trained to compress the information from 88 diagnostic signals into a concise embedding, to provide a unified interface between diagnostic systems and control actuators. Two mechanisms are proposed to ensure a meaningful embedding: compression-reconstruction and missing-signal reconstruction. Upon completion of pre-training, the model acquires the capability for ‘virtual backup diagnosis’, enabling the inference of missing diagnostic data with 97.2% accuracy. Furthermore, the model demonstrates multiple downstream applications: automatic data analysis, universal control-diagnosis interface, and enhancement of control performance on multiple tasks. This work pioneers the integration of large-scale artificial intelligence (AI) models into the field of fusion energy. It demonstrates that pre-trained embeddings can simplify the system interface, reduce diagnostic redundancy and optimize operation performance, paving the way for high-performance future fusion reactors. This study proposes a self-supervised pre-trained model for experimental data from fusion devices. Experiments conducted on HL-3 demonstrate that this model can serve as a virtual backup diagnostic, reconstructing missing diagnostic data with 97.2% accuracy, and comprehensively enhance the performance of other downstream AI4Fusion algorithms as a device-specific foundation model.