Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.22817

Multiple Instance Learning for Interpretable Plasmoid Detection in the FLARE Device

Charles Marsden·Adam Robbins·Hantao Ji
View original source ↗Read source document ↗Export metadata ↓

Abstract

Multiple instance learning (MIL) is a form of weakly supervised learning widely used in healthcare and biomedical imaging, where labels may be available for groups of observations ("bags") but not individual observations ("instances"). This setup is well-suited to plasmoid detection in magnetic reconnection experiments, since identifying dynamic magnetic structures is best accomplished taking in data from many time steps as a whole, as opposed to viewing frames in isolation. Instance-level consideration enables researchers to determine the range of times most affecting model output, thus serving as a tool for the temporal localization of plasmoids. Despite its success in the life sciences, to the best of our knowledge, MIL is yet to be reported in peer-reviewed plasma physics literature. We evaluate a MIL model against a convolutional recurrent neural network (CRNN), an architecture commonly found in plasma physics literature, on a plasmoid detection task, using a highly imbalanced dataset of 1152 magnetic reconnection shots from the FLARE device at PPPL. We demonstrate significant improvements in test performance (PR AUC: $Δ$=+0.0932, p=0.0039; F1: $Δ$=+0.0259, p=0.0488; Wilcoxon signed-rank test) against the CRNN model on full-channel input. We investigate the interpretability advantage provided by our MIL model for the temporal localization of plasmoids. Through its attention mechanism, we can discern relevant frames and even the number of plasmoids present in a shot. Overall, we show that MIL is a promising framework for similar event classification tasks in plasma physics.

Explore related subjects

physics.plasm-ph

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Charles Marsden, Adam Robbins, Hantao Ji. 2026-10-16. Multiple Instance Learning for Interpretable Plasmoid Detection in the FLARE Device. https://arxiv.org/abs/2610.22817

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Horizon-Aware Early Event Prediction for Tokamak Disruption Alarms

Disruption predictors must warn the plasma control system early enough for the intended response, and the required warning time differs between mitigation and avoidance. We view a warning system as a predictor that outputs a disruption risk and an alarm policy that acts on it, and ask which risk the predictor should learn. Binary classifiers do not specify when a disruption occurs, and survival analysis models the full distribution of the time to disruption, whereas an alarm depends only on the risk within a finite horizon. We introduce two horizon-focused targets into tokamak disruption prediction: a horizon-truncated survival objective and fixed-horizon early event prediction (EEP) with temporal label smoothing (TLS). We compare them with binary classification and survival analysis on the DIII-D, Alcator C-Mod, and EAST databases of DisruptionBench, evaluating the risk by its threshold-free ranking at a 40 ms deadline and the alarm policy by thresholds calibrated on validation data. On every device, TLS outperforms the deep survival models in both ranking and alarm performance; TLS performs best on DIII-D and EAST, and a random forest trained on a hard horizon target performs best on C-Mod. TLS also improves the calibrated alarm performance over hard targets placed at either end of its smoothing range on all devices, isolating the benefit of label smoothing. Horizon truncation alone did not consistently improve survival analysis. Alarm performance remains limited on C-Mod and DIII-D and varies strongly across training seeds, and ablations show that the target should extend to the required warning time.

physics.plasm-ph↗
arXiv2026

Properties and challenges of machine learning closures for ten-moment fluid models in the case of Landau damping

Two-fluid ten-moment models for plasma simulations evolve the velocity moments of the Velocity Distribution Function (VDF) up to the pressure tensor and require a closure for the divergence of the next velocity moment, the heat flux. We investigate the use of Machine Learning (ML) models for the discovery of heat flux closures, as an alternative to the typically employed Hammett-Perkins-like analytical closures. As a test case, we investigate Landau damping, which describes a resonant process of energy transfer between particles and waves. We find that fluid simulations with ML closures correctly reproduce both the damping frequency and the non-linear phase of Landau damping, and thereby outperform Hammett-Perkins-like closures, within the parameter ranges seen during training and when interpolating within those ranges. When extrapolating beyond the training data, the performance of the ML closure degrades with increasing distance from known parameters. We then investigate how well offline metrics (i.\,e.\; evaluating the trained model on fully-kinetic reference simulations, without inserting it into two-fluid simulations) can predict online performance, when the model is actually used in the two-fluid simulation. Our results show that offline metrics do not fully explain online performance, highlighting important implications for method evaluation in this field. We further investigate suitable metrics for online performance evaluation in the absence of a fully-kinetic reference simulation, which is often computationally expensive to run. We propose using the temporal evolution of entropy to evaluate the performance of reduced models with ML closures when no or few kinetic reference simulations are available.

physics.plasm-ph↗
arXiv2026

Role of Device Size on Island Divertor Performance of W7-X-like Configurations

To examine how the island divertor performs in future low-shear stellarator reactors, we simulate (using EMC3-Eirene [1]) 3 different scales of W7-X: its current size, increased by a factor 2, and increased by a factor 4, where 4 times W7-X corresponds roughly to a reactor-scale device. For a single magnetic configuration, the standard magnetic configuration in W7-X, we examine how increasing the size of W7-X alters the divertor performance of the island divertor. We analyze how device scale shifts the location of the ionization source and the downstream density and temperature trends in the island divertor for achieving detachment. Local field line analysis shows that factor 2 and factor 4 size scalings achieve target densities $>10^{21} \ \mathrm{m}^{-3}$ via plasma ``self-baffling" within the open island divertor geometry. Finally, the heat and particle flux deposition on the scaled-up island divertor configurations are calculated. We observe a separation in the target heat and particle flux distribution with increased machine size in agreement with [2]. The target heat fluxes are $<5 \ \mathrm{MW}/\mathrm{m}^2$ at high separatrix densities. This shows that the separation of the heat and particle flux channels remains to be largely unproblematic. This analysis helps guide how to improve the open island divertor geometries for future stellarator reactors.

physics.plasm-ph↗