[논문 리뷰] Machine learning prediction of plasma behavior from discharge configurations on WEST

Chenguang Wan, Feda Almuhisen외 11명|arXiv (Cornell University)|2026. 02. 22.

Magnetic confinement fusion research인용 수 0

한 줄 요약

논문은 WEST에서 비방전 신호로부터 여섯 가지 글로벌 플라즈마 매개변수를 예측하는 트랜스포머 기반 ML 대리모형을 개발하여 높은 정확도와 빠른 추론 속도를 달성합니다.

ABSTRACT

Accurately predicting plasma behavior based on discharge configurations is essential for the safe and efficient operation of tokamak experiments. While physics-based integrated modeling codes provide valuable insights, their high computational cost limits their applicability for fast scenario design and control optimization. In this study, we propose a transformer-based machine learning model to predict key global plasma parameters on the WEST tokamak, including the normalized beta ($β_{n}$), toroidal beta ($β_{t}$), poloidal beta ($β_{p}$), plasma stored energy ($W_{\mathrm{mhd}}$), safety factor at the magnetic axis ($q_{0}$), and safety factor at the 95% flux surface ($q_{95}$). The model uses only signals that can be defined before the discharge, such as magnetic coil currents, auxiliary heating power, plasma current reference, and line-averaged plasma density. Trained on 550 discharges from the WEST campaigns, the model demonstrates an average mean square error (MSE) loss of 0.026, an average coefficient of determination $R^{2}$ of 0.94, and achieves inference times on the order of 0.1 seconds. These results highlight the potential of data-driven surrogate models for assisting in discharge planning, scenario evaluation, and real-time control of tokamak plasmas.

연구 동기 및 목표

• WEST의 비방전 신호로부터 여섯 가지 핵심 글로벌 플라즈마 매개변수(정규화된 베타 βn, 토로이드 베타 βt, 폴도이드 베타 βp, 플라즈마 저장 에너지 Wmhd, q0, 그리고 q95)를 예측한다.
• 방전 전 사용 가능한 신호(자기코일 전류, 보조 가열 전력, Ip 기준치, 라인-평균 밀도)를 사용하여 빠른 대리모형을 구축한다.
• 방전 예측에 대해 최상의 성능을 보이는 ML 아키텍처를 식별하기 위해 여러 ML 구조를 비교한다.
• 550개의 WEST 방전에 걸쳐 모델 일반화 성능을 평가하고 q-매개변수 예측의 한계를 분석한다.]
• method 마크다운은 제외하고 자연어로 작성

제안 방법

• 데이터를 스무딩하고 고정 크기의 슬라이딩 윈도우(길이 1024, 간격 512) 및 중첩 평균화를 통해 안정성을 확보한다.
• 550개의 방전을 학습/검증/테스트(train/validation/test; 60/20/20)로 구분한다.
• MLP, LSTM, Transformer Encoder, Transformer Decoder, 및 Bayesian 하이퍼파라미터 최적화를 사용하는 Transformer 기반 모델을 비교한다.
• 검증 MSE 기준으로 최적 모델을 선택한 뒤 테스트 세트 성능을 보고한다.
• 추론 시간과 방전 계획 및 실시간 제어의 실용적 적용 가능성에 대해 논의한다.]
• research_questions[0] Can a pre-discharge, multi-signal input set predict key zero-dimensional plasma parameters for WEST?,
• research_questions[1] Which ML architecture best captures temporal and cross-channel dependencies in discharge data?,
• research_questions[2] What is the predictive accuracy (MSE, R²) and inference speed achievable for these predictions?,
• research_questions[3] What are the limitations in predicting q0 and q95 from pre-discharge signals, and why?,
• research_questions[4] How well does the model generalize across a diverse set of WEST discharges?]
• key_findings:["Average test-set MSE is 0.026 for the six outputs.","Average R² is 0.94 on the test set.","Transformer-based model achieves the lowest validation MSE among tested architectures (0.0096).","All input signals show significant Granger causality and Pearson correlations with outputs, justifying feature relevance.","q0 and q95 predictions are less accurate due to weak identifiability from pre-discharge inputs and lack of pressure/Kinetic profiles in inputs.","Inference time is about 0.1 seconds on both RTX 3090 and A100 GPUs."]
• table_headers:["모델 타입","평균 제곱 오차(MSE) 손실"]
• table_rows:[ ["Multilayer perceptron (MLP)","0.0224"], ["Long Short-Term Memory (LSTM)","0.015"], ["Transformer Encoder","0.011"], ["Transformer Decoder","0.011"], ["Our transformer-based model","0.0096"]]} }

Figure 1: The Granger causality and Pearson correlation coefficient between input and output signals. For Granger causality, a smaller coefficient indicates a stronger causal relationship, whereas for the Pearson correlation coefficient, a larger value reflects a stronger linear correlation. PowerLH

실험 결과

연구 질문

• RQ1 Can a pre-discharge, multi-signal input set predict key zero-dimensional plasma parameters for WEST?
• RQ2 Which ML architecture best captures temporal and cross-channel dependencies in discharge data?
• RQ3 What is the predictive accuracy (MSE, R²) and inference speed achievable for these predictions?
• RQ4 What are the limitations in predicting q0 and q95 from pre-discharge signals, and why?
• RQ5 How well does the model generalize across a diverse set of WEST discharges?

주요 결과

모델 타입	평균 제곱 오차(MSE) 손실
Multilayer perceptron (MLP)	0.0224
Long Short-Term Memory (LSTM)	0.015
Transformer Encoder	0.011
Transformer Decoder	0.011
Our transformer-based model	0.0096

• Average test-set MSE is 0.026 for the six outputs.
• Average R² is 0.94 on the test set.
• Transformer-based model achieves the lowest validation MSE among tested architectures (0.0096).
• All input signals show significant Granger causality and Pearson correlations with outputs, justifying feature relevance.
• q0 and q95 predictions are less accurate due to weak identifiability from pre-discharge inputs and lack of pressure/Kinetic profiles in inputs.
• Inference time is about 0.1 seconds on both RTX 3090 and A100 GPUs.

Figure 2: The workflow of machine learning model in the present work