TokaMark: A Comprehensive Benchmark for MAST Tokamak Plasma Models

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Science & Research — Physical Sciences & Chemistry, Engineering & Applied Sciences, Research Methodology & Innovation · Depth: Expert, extended

Summary

TokaMark is a new, open-source benchmark designed to evaluate AI models for predicting plasma dynamics in fusion energy reactors, specifically using real experimental data from the Mega Ampere Spherical Tokamak (MAST). It addresses the lack of curated datasets and standardized benchmarks in fusion AI. TokaMark unifies access to multi-modal, heterogeneous fusion data, harmonizes formats, and provides 14 tasks across four groups: equilibrium reconstruction, magnetics dynamics, profile dynamics, and MHD activity. The benchmark includes a multi-branch convolutional encoder-decoder baseline model, trained on 11,573 shots from the FAIR-MAST dataset with an 80%/10%/10% split. It uses a hierarchical evaluation protocol and provides Python tools for data loading and processing.

Key takeaway

For AI Scientists and Machine Learning Engineers developing models for fusion energy, TokaMark offers a critical standardized platform. You should use this open benchmark to rigorously compare your models against established baselines. Focus on tasks like profile dynamics and MHD activity. The baseline shows significant room for improvement in these areas, with NRMSE scores exceeding 0.17; task 4-5 even exceeds unity. This will accelerate the development of robust, data-driven plasma models essential for commercially viable fusion.

Key insights

TokaMark provides a unified, open benchmark with 14 tasks and a baseline for AI models in fusion plasma modeling.

Principles

• Fusion data is multi-modal, multi-rate, and often incomplete.
• AI models can learn latent plasma representations from raw data.
• Standardized benchmarks accelerate progress and ensure reproducibility.

Method

TokaMark defines 14 tasks with input/output windows, uses a sliding-window approach with a 0.001-second stride, and employs a hierarchical evaluation protocol (samples -> windows -> signals -> tasks -> shots) with NRMSE.

In practice

• Evaluate AI models on 14 diverse fusion plasma tasks.
• Utilize the provided multi-branch convolutional baseline.
• Develop models robust to multi-fidelity and missing data.

Topics

• TokaMark
• Fusion Energy
• Tokamak Plasma Modeling
• AI Benchmarking
• Multi-modal Data
• MAST Tokamak
• Plasma Diagnostics

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Offline Reinforcement Learning for Plasma Control in Nuclear Fusion: Codebase and Benchmark — RL4F provides a standardized offline RL benchmark for multi-actuator, long-horizon nuclear fusion plasma control.
• Science Needs AI Data Stocktakes — Scientific progress requires structured "AI data stocktakes" to identify and address data obstacles for AI model training.
• Accelerate Clean, Modular, Nuclear Reactor Design with AI Physics — AI-augmented simulations with surrogate models significantly accelerate nuclear reactor design by overcoming computational bottlenecks.
• PRL-Bench: A Comprehensive Benchmark Evaluating LLMs' Capabilities in Frontier Physics Research — PRL-Bench evaluates LLMs' end-to-end physics research capabilities, revealing significant performance gaps.
• Meta AI - AI at Meta — Advancements span LLM reasoning, vision models, and material science, driven by new benchmarks, datasets, and training methods.
• Commonwealth Fusion Systems installs reactor magnet, lands deal with Nvidia — Fusion power development is accelerating through advanced magnet technology and digital twin simulations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.