Disruption Warning Database Development and Exploratory ...€¦ · •‘non-disrupted’ = sample from shot with no disruption •‘far from disr’ = sample from disrupted shot

• Avoid processing signals with non-causal filtering; this can introduce post-disruption effects into pre-disruption data

• Pre-processed signals in database to avoid excessive smoothing and interpolation

• Analyzed 7/40 dimensionless or machine-independent parameters from database using a machine learning algorithm

• Difference in timescales on DIII-D and C-Mod evident when comparing design points and time evolution of parameters

• Poorer predictive capability on Alcator C-Mod compared to DIII-D may be due to faster disruption-relevant timescales

• At present data acquisition rate, difficult to predict disrupts• Compare performance of other ML algorithms and study

dependence on new features as the database is updated

Disruption Warning Database Development and Exploratory

Machine Learning Studies on Alcator C-ModK. Montes, C. Rea, R. Granetz

Plasma Science and Fusion Center, Massachusetts Institute of Technology

Introduction

Conclusions and Future Work

References

Disruption Warning Database• SQL database of > 40 parameters from 1821 shots (~160k time

slices) from 2015 C-Mod campaign• Only time slices in Ip flattop included; composed of non-disruptive

discharges and discharges that disrupted during the flattop• Ignored intentional massive gas injection (MGI) disruptions

• Each database record consists of all parameter values at one time slice, recorded every 20 ms; for each disruption, take additional time slices every 1 ms during the 20 ms period before disruption

C-Mod and DIII-D Comparison• Given input parameters Ԧ𝑥 and historical knowledge of disrupted shots

𝑌, how can we find patterns to distinguish disruptions in our database?

• Random forest for classification using 3 different labeling schemes

AXUV diode channel (no smoothing)

𝑃𝑟𝑎𝑑 with non-causal smoothing (not ok near disruptions) from blackened bolometer

Non-causal filtering example: 𝑃𝑟𝑎𝑑calculation on C-Mod taken from AXUV diode to avoid non-causal filter

Total # of Shots 1821

Non-Disruptive Flattop Shots 1160

Disruptions in Flattop 206

Intentional MGI Disruptions 17

𝐼𝑃 Flattop

Shot # 1150501010

[1] C. Rea et al. APS (Oct. 2017)[2] O. Sauter and Y. Martin Nuclear Fusion 40 (2000) 955[3] C. M. Greenfield et al. Plasma Physics and Controlled Fusion 46 (2004) 12B[4] G.M. Wallace et al. IAEA Conference (2012)[5] J. Vega et al. Fusion Engineering and Design 88 (2013)[6] E. Alpaydin, “Introduction to Machine Learning”, 2nd Edition, MIT Press[7] L. Breiman, “Random Forests”, Machine Learning, 45(1), 5-32, 2001

Major Radius

Minor Radius

Toroidal Field

Plasma Current

Confinement Time [2]

Current Relaxation Time [3,4]

DIII-D 1.67 m 67 cm ~2 T 3.0 MA ≈ 0.1 𝑠 ~ 1 𝑠

C-Mod 68 cm 22 cm 3-8 T 0.4-2 MA ≈ 0.04 𝑠 ~ 0.2 𝑠

• Common cause of C-Mod disruptions is radiative collapse from high-Z first wall molybdenum (1-2 ms timescale)

• In contrast, DIII-D has low-Z carbon wall; most disruptions due to MHD instabilities

Supervised Learning for Classification

Binary Phase Classification:• ‘stable’ = non-disrupted or > 40

ms from disruption• ‘disruptive’ = < 40 ms from

disruption

Classification Accuracy:• Disruptive: 48.5 % • Stable: 99.3 %• Overall: 97.3 %

Binary Classification:• ‘non-disrupted’ = sample from

shot with no disruption• ‘disrupted’ = sample from

disrupted shot

Classification Accuracy:• Disrupted: 52.6 % • Non-Disrupted: 97.0 %• Overall: 91.2 %

• Predicting and mitigating disruptions in tokamaks is critical to the mission of sustaining a fusion plasma

• To understand what causes disruptions, we want to answer:• Which parameters are correlated with the approach of a

disruption? What are their threshold levels?• Are the thresholds reached with significant warning time?• Are there combinations of parameters that are useful?• Are the same parameters useful on different tokamaks?

• Goal: Develop a disruption warning algorithm that works in near real-time, embedded in the plasma control system

Yes No

Yes No

𝑥1 > −0.55

𝑥2 > 0.3

branchesR1

R3R2

Minimize impurity measure to determine splitting value at each node:

leaves

decision node

1 Plasma Current Error Fraction ip_error_frac

2 Internal Inductance li

3 Greenwald Fraction n/nG

4 q95 (Safety Factor at r = 0.95a) q95

5 Poloidal Beta Ratio betap

6 Loop Voltage Vloop

7 Radiated Power Fraction prad_frac

Multi-Class Classification:• ‘non-disrupted’ = sample from

shot with no disruption• ‘far from disr’ = sample from

disrupted shot > 40 ms from disruption

• ‘close to disr’ = sample from disrupted shot < 40 ms from disruption

Classification Accuracy:• Non-Disrupted: 97.4 % • Far from Disr: 37.3 %• Close to Disr: 53.3 %

Overall Accuracy: 90.1 %

• Large overlap of internal inductance distributions compared to DIII-D for time slices grouped via the multi-class classification case;

Supervised Learning

Learn 𝑌 = 𝑓(𝑥)

Unsupervised Learning

Search 𝑋 = Ԧ𝑥 for structure & patterns

ClusteringDiscover groupings in parameter space

Machine Learning Algorithms

AssociationDiscover rules that

relate data

Classification𝑌 = discrete (class)

Regression𝑌 = continuous

(likelihood or time)

Linear regression, neural networks, random forest, etc.

Random forest, logistic regression, support vector machines, etc.

K-means clustering, self-organizing maps, Gaussian mixture models, etc.

Apriori algorithm, equivalence class transformation, etc.

40

par

amet

ers

Ip (MA)

ne (m-3)

q95

[1] C. Rea et al. APS (Oct. 2017)

Shot # 1150806029

C-Mod li distribution mean

Power Spike Before C-Mod Disruption