Bridging the gap: Deploying AI-based Models in Real-Time Fusion Plasma Control Systems

Achieving reliable real-time control in fusion plasma experiments requires strict timing guarantees across entire control algorithms. In earlier work by Abbate et al. (2023), we demonstrated the feasibility of neural-network-based control algorithms on the DIII-D tokamak using the internally developed open-source Keras2C library for model conversion into C (Conlin et al. (2021)). However, the initial implementations relied on data buffering and branching logic outside the neural network code, causing variability in execution times. Subsequent deployments on DIII-D and KSTAR - including the RTCAKENN algorithm for kinetic profile reconstruction - proved that minimizing branching and buffering throughout the pipeline yields consistent millisecond-level cycle times under real experimental conditions (Shousha et al. (2023))). However, keeping pace with rapidly evolving AI frameworks (e.g. PyTorch) is challenging. We therefore propose a community-driven open-source effort to expand the tool, enabling real-time deployment across diverse systems that require strictly bounded execution times.