Preamble

Astronomical Techniques and Instruments, Vol. 2, September 2025, 310–318 Review Open Access Stellar flare detection methods in TESS data: application and per- formance study Min Li , Liang Wang 4,5,6 , Zhiqiang Zou 1,2,3 , Ali Luo 3,6,7 , Bo Qiu , Peng Jia , Ying Shan

1 College of Computer

Nanjing University of Posts and Telecommunications Nanjing 210023, China 2 Jiangsu Key Laboratory of Big Data Security and Intelligent Processing Nanjing 210023, China

3 University of Chinese Academy of Sciences , Nanjing 211135, China

4 Nanjing Institute of Astronomical Optics Technology Chinese Academy of Sciences Nanjing 210042, China 5 CAS Key Laboratory of Astronomical Optics Technology Nanjing Institute of Astronomical Optics Technology Nanjing 210042, China

INTRODUCTION

Stellar flares are explosive phenomena at the surface and in the atmosphere of stars . They are caused by a release of energy induced by the star’s magnetic field and are a very common manifestation of stellar activity. Flares are most common in late-type M dwarfs , but they also occur frequently in main-sequence stars and, more rarely, evolved giant stars . The study of stellar flares is essen- tial to better understand stellar evolution and planetary hab- itability . For example, intense stellar flare activity can strongly affect the surrounding planets , notably by heat- ing their atmosphere or causing ionospheric distur- bances Research on stellar flares has traditionally relied on observational data collected by ground-based and space- borne telescopes . Recently, the development of larger- scale space missions, such as the TESS , has consider- ably transformed flare studies. To date, TESS has pro- vided an extensive, high-quality dataset that has already markedly accelerated progress in stellar flare research and improved our understanding of flare properties, energy release mechanisms, and role in stellar activity cycles.

Flare detection methods can be categorized into tradi- tional and ML approaches. Traditional methods typically involve detrending and statistical analysis of stellar observa- tions. For example, the WARPFINDER algorithm devel- oped by Pietras et al. employs a three-step approach (detrending, differencing, and curve fitting) to detect flare events automatically. Although effective, such methods fre- quently require manual parameter adjustment and cannot process large datasets efficiently.

ML methods have been increasingly applied to the anal- ysis of astronomical datasets, including TESS data.

6 University of Chinese Academy of Sciences , Beijing 100049, China

7 CAS Key Laboratory of Optical Astronomy National Astronomical Observatories Beijing 100101, China 8 School of Intelligent Science and Technology University of Science and Technology Beijing Beijing 100083, China 9 College of Electronic Information and Optical Engineering Taiyuan University of Technology Taiyuan 030024, China *Correspondences: © 2025 Editorial Office of Astronomical Techniques and Instruments, Yunnan Observatories, Chinese Academy of Sciences. This is an open access article under the CC BY 4.0 license ( Citation: Li, M., Wang, L., Zou, Z. Q., et al. 2025. Stellar flare detection methods in TESS data: application and performance study.

Astronomical Techniques and Instruments (5): 310−318.

Abstract

The detection of stellar flares is crucial to understanding dynamic processes at the stellar surface and their potential impact on surrounding exoplanetary systems. Extensive time series data acquired by the Transiting Exoplanet Survey Satellite (TESS) offer valuable opportunities for large-scale flare studies. A variety of methods is currently employed for flare detection, with machine learning (ML) approaches demonstrating strong potential for automated classification tasks, particularly for the analysis of astronomical time series. This review provides an overview of the methods used to detect stellar flares in TESS data and evaluates their performance and effectiveness. It includes our assessment of both traditional detection techniques and more recent methods, such as ML algorithms, highlighting their strengths and limitations. By addressing current challenges and identifying promising approaches, this manuscript aims to support further studies and promote the development of stellar flare research.

Keywords

Stellar flare detection; TESS light curve; ML; Automatic classification

Recent advances include the application of convolutional neural networks (CNNs) such as Stella, developed by Fein- stein et al. , to study the relationship between stellar prop- erties and flare activity; of recurrent neural networks (RNNs) by Vida et al. for flare detection; and of CNNs with ensemble learning by Tu et al. to identify super- flares (stellar burst phenomena with longer duration and larger energy than typical stellar flares) in TESS data These models, trained on large volumes of TESS data, have enhanced detection efficiency and minimized the need for manual intervention. ML methods provide robust and systematic solutions to process large observational datasets, enabling automatic identification and classifica- tion of flare events.

In this review, we provide a comprehensive overview of stellar flare detection methods from TESS data that cov- ers traditional techniques and ML approaches. We evalu- ate the advantages and limitations of each method, focus- ing on accuracy and computational efficiency. Addition- ally, we discuss the challenges and opportunities of apply- ing ML to stellar flare detection. The article structure is as follows: in Section 2, we introduce TESS data; in Sec- tion 3, we describe traditional and ML flare detection meth- ods; in Section 4, we present experimental results and eval- uate method performance; in Section 5, we discuss the chal- lenges of applying ML to astronomical research; and in Sec- tion 6, we summarize flare detection methods and sug- gest future research directions.

TESS DATA OVERVIEW TESS Mission and Scientific Objectives The primary objective of the TESS mission is the search for exoplanets, especially small telluric planets located in the habitable zone of their star. These planets are identified with the transit method, which consists of observing a slight decrease in stellar brightness caused by a planet passing in front of its star. Additionally, the high precision of TESS data is well-suited to study stellar physics. By analyzing light curves (LCs), scientists can study stellar properties such as rotation, pulsation, and mag- netic activity that provide a basis to characterize stellar structure and evolution.

During its primary mission, TESS conducted a two- year survey of the solar neighborhood. On July 4, 2020, its primary mission ended; TESS is now in its extended mission. The planets discovered by TESS range from small telluric planets to gas giants, clearly demonstrating the diversity of planets in our galaxy. These exoplanets and, additionally, active stars identified by TESS during its primary mission have been defined as priority observa- tion targets for subsequent missions, such as the James Webb Space Telescope , to improve our understanding of planetary atmospheres and stellar activity. During its pri- mary mission (Section 2.2), TESS imaged approximately 75% of the sky, discovered 623 confirmed exoplanets, and listed nearly 7 643 additional candidates, which are cur- rently awaiting confirmation by independent instruments or methods TESS Data Acquisition and Processing TESS is equipped with four wide-field cameras, which are fixed to the satellite platform. To achieve point- ing toward different regions of the sky, the entire space- craft is rotated. During its primary mission phase, TESS executed 26 distinct pointings, with each pointing main- tained for approximately two orbital cycles. By the end of the primary mission, the full southern and northern celes- tial hemispheres had been surveyed.

To maximize sky coverage, these pointings were evenly distributed along the ecliptic longitude. Each individ- ual camera provides a field of view of 24° × 24°, and the four cameras together form a combined synthetic field of view of 24° × 96°, covering approximately 2 300 square degrees. One of the cameras is directed toward the eclip- tic pole, while the remaining three are spaced at equal inter- vals between the ecliptic pole and an ecliptic latitude of 6°, thereby optimizing overall sky coverage.

During operation, each camera nominally acquires an image every 2 s. To optimize storage and processing time, single images are accumulated onboard by sets of 60 to generate composite images with equivalent expo- sure times of 2 min that are stored in the onboard solid- state buffer (SSB). The 2-min equivalent images, or “postage stamps”, focus on areas surrounding the target source; their usual size is 10 × 10 pixels. Their main use is to characterize light variability near a target for subse- quent analysis. Then, each image is analyzed by aperture photometry to generate a flux array (the LC), ultimately stored in data files labeled “LC files” . An example of generated photometric data is shown in . In addi- tion, aggregated FFIs are collected every 30 min and also stored in the SSB for larger-scale light variation analysis and for stellar research. Although it covers a consider- ably larger field of view than the postage stamps, each FFI remains highly sensitive to small signal variations, pro- viding high-resolution, high-sensitivity, large-scale data for exoplanet and stellar research.

Every 13.7 days, from 8 h before to 8 h after perigee, data stored in the SSB, including all postage stamps and FFIs, are transmitted to the ground segment through the National Aeronautics and Space Administra- tion Deep Space Network Characteristics of TESS LCs Two types of LC data are acquired by TESS for its tar- get stars (440 000 targets as of February 2023) at inter- vals of 2 min (standard LCs) and 20 s (fast-LCs) Because the number of target stars with available fast- LCs is small, we restrict our review to the TESS 2-min LC data. Their characteristics are described in this section.

Standard LCs of some target stars exhibit gradual brightness increases or decreases that can be attributed to

Every 2 min Every 30 min Full frame images slow alterations of the internal structure or physical state of these stars. In other cases, small amplitude fluctua- tions can be discerned in the LCs. These fluctuations may be associated with small-scale activity at the star’s sur- face or with star vibrations. Conversely, sporadic events, such as stellar flares or planetary transits, induce sudden brightness changes materialized by peaks or troughs in the LC. Such events provide essential data on stellar atmo- spheric activity and magnetic field strength.

In addition, periodic brightness variations are fre- quently observed in the LCs of TESS targets; they usu- ally reflect the star’s rotation, surface activity, or possi- ble planetary transits. The star’s rotation period can be inferred from these periodic LC patterns. The amplitude of the brightness variations reflects the intensity of stellar activity, especially if they are caused by magnetic activ- ity or stellar rotation. Stellar spots on the star’s surface can also affect the amplitude of the fluctuations.

In TESS data, flare events can be identified by the shape of the LC features: they are usually short and charac- terized by a rapid and steep brightness increase followed by a slow decrease, forming a typical “spike-shaped” peak, as illustrated in , with a maximum brightness largely exceeding background brightness fluctuations. In some active stars, multiple flare “bursts” can be observed, appearing as independent peaks in the LC. These multi- ple bursts are important to study the periodicity of stellar magnetic activity.

FLARE DETECTION METHODS

In this section, we present an overview of stellar flare detection methods, which are broadly categorized into tradi- tional threshold-based approaches and ML techniques.

Traditional Threshold Detection Methods In traditional threshold detection methods, candidate flare events are identified by LC brightness variations exceeding a preset statistical threshold (e.g., 2.5 standard deviations or higher) . Here, the standard deviation Time/day Light curve files

1.0 Flare

Time/day ) represents the statistical dispersion of stellar bright- ness measurements in an LC when the observed star is in its quiescent, non-flaring state. It quantifies the variabil- ity in the star brightness induced by instrumental noise, stel- lar activity, or other non-flare-related fluctuations. Mathe- matically, if represents the brightness measurements and is the mean quiescent flux, is given by

1 N

i = 1

The flare start and end times are determined from the set of consecutive data points exceeding the preset thresh- . Standard steps of threshold detection methods include LC detrending, noise removal, threshold filtering, and manual validation.

Detrending methods The objective of detrending is to eliminate long-term variations in the LC, such as stellar spot modulation and instrument drift, while preserving short-term flare signals . To achieve this, common techniques such as polynomial fitting, spline smoothing, and iterative denois- ing are employed . Polynomial fitting uses low-order poly- nomials (e.g., cubic polynomials) to model the long-term LC trend. This trend is then subtracted from the original Target pixel files

data. This effectively removes low-frequency variations from the LC . Spline smoothing and Gaussian filtering are two common techniques used to smooth LCs, aiming to reduce low-frequency noise while preserving short-term flare signals . Finally, iterative denoising removes out- liers or large flares through multiple iterations and gradu- ally converges toward a stable, detrended LC In summary, detrending methods are effective in remov- ing undesired, long-term brightness variations from observed LCs, thereby enhancing the observation sensitiv- ity for flare signal detection. However, such methods also present limitations: for example, overfitting might occur, potentially resulting in the misidentification of large flares as long-term trends; moreover, their overall effectiveness st- rongly depends on the selection of appropriate parameters.

Difference methods Different methods use a threshold detection approach relying on brightness differences to identify flare events . Standard difference calculation methods include “running differences” and “resting model” differ- ences. The running difference method calculates flux differ- ences between temporally adjacent LC points and com- pares them with a predefined threshold. Differences that exceed threshold marked flare candidates . The resting model difference method glob- ally subtracts a “resting model” (e.g., the detrended LC) from the original LC. This process results in a different LC that can effectively capture rapid brightness changes unrelated to long-term trends . Different methods are par- ticularly effective in identifying sudden brightness increases during flares, hence to detect flares on short timescales. Additionally, they are computationally simple and can be easily parallelized. However, such methods are sensitive to noise, which might cause misidentifica- tion of cosmic rays or short-term instrumental errors as flares. Therefore, combining different methods with detrend- ing techniques is essential to reduce the likelihood of erro- neously interpreting cosmic rays or transient instrumental effects as real flaring events.

ML Methods The core idea in applying ML methods to stellar flare detection is to apply data-driven models automatically to “learn” stellar flare features, thereby enhancing the model’s detection accuracy and generalization ability. A cru- cial first step in this process is data preprocessing, which often includes LC normalization —a technique that rescales the data to a common baseline to reduce observa- tional noise. Additionally, the Synthetic Minority Over-sam- pling Technique (SMOTE) is commonly applied to address class imbalance, a situation where flare (rare events) images are considerably outnumbered by non-flare images, by synthetically generating more examples of the minority image class (flares). In terms of feature extrac- tion, some methods use manual feature definition, e.g., flare duration, peak amplitude, or equivalent duration while others rely on automatic extraction of time or fre- quency-domain features, for example, using CNNs pixel-level data . For model selection, commonly used supervised learning methods include random forests (RFs) and extreme gradient boosting (XGBoost) while learning approaches include , and deep neural networks (DNNs) Finally, model ensemble techniques improve detection stability and accuracy through multi-algorithm voting (i.e., several models are applied to the same data and detec- tion is confirmed if at least two models concurrently iden- tify a flare event) or enhance the robustness of the detec- tion system through stacking or weighted fusion combination of ML techniques can thus provide a more reli- able and comprehensive solution than traditional methods for flare detection.

EXPERIMENTAL

RESULTS

PERFORMANCE EVALUATION Performance Evaluation Metrics In ML methods, the stellar flare detection task can be formulated as a binary classification problem, because its primary objective is to differentiate between two distinct states: “flare” and “non-flare.” The confusion matrix is an essential tool to evaluate model performance for such binary classification tasks; it provides detailed characteriza- tion of the model’s predictive abilities using complemen- tary metrics . As illustrated in , the confusion matrix is a 2 × 2 table that includes the four possible out- comes of binary classification: true positives (TP), false pos- itives (FP), true negatives (TN), and false negatives (FN).

Predicted true Predicted false Actual true Actual false In this study, we use the confusion matrix to quan- tify the models’ predictive accuracy for flare detection by calculating key performance metrics: accuracy, recall, and precision. (1) Accuracy Accuracy is the proportion of correctly predicted flare events among the total sample set

Accuray = TP + TN TP + TN + FP + FN . (2)

For stellar flare detection, high accuracy indicates that the model correctly identifies flare events, with fewer FPs. (2) Recall Recall describes the model’s ability to identify posi- tive samples, i.e., the proportion of detected flare events rel- atively to the total number of real flare events,

Recall = TP TP + FN . (3)

High recall indicates that the model can effectively iden- tify a large proportion of real flare events and avoids FNs (3) Precision Precision is the proportion of correctly predicted posi- tive samples among all positive predictions,

Precision = TP TP + FP . (4)

It measures the “positive prediction accuracy” of the model Comparison Between Traditional

Methods

We designed an experiment to evaluate and compare the stellar flare detection performance of traditional thresh- old detection methods and ML methods. The dataset used for this experiment consisted of the TESS 2-min data for M dwarf stars. To ensure data comparability under identi- cal evaluation conditions, we preprocessed LC data by applying normalization, binary segmentation into flare and non-flare samples, and annotation of the corresponding labels.

Specifically, for standard threshold methods, we used the approach proposed by Yang et al. as a detrending method and the brightness difference detection technique introduced by Shibayama et al. as a different method.

For ML, we used a CNN developed by Feinstein et al.

Although the implementation of the traditional thresh- old detection and ML methods differs substantially, their outputs are consistent: both types of methods predict a binary class label (“flare” or “non-flare”) for each sample.

This consistency allows for direct comparison between the results of both approaches and the true labels and for calculation of the accuracy, precision, and recall. We also compared the relative computation time required by all methods to assess their practical efficiency.

The experimental results ( ) demonstrated that, in terms of accuracy, precision, and recall, ML methods Accuracy Precision Recall Relative runtime Metric value Difference clearly outperformed traditional methods, illustrating the strong ability of CNNs to handle complex LC data.

Although a single CNN architecture was tested in our analy- sis, its performance suggests that all deep learning mod- els, particularly those with convolutional structures, are well suited to capture local and hierarchical features in LC data. This result is consistent with prior studies that reported similar advantages of CNNs for time-series and astrophysical signal classification tasks, indicating the poten- tial generalizability of our results. In our experiment, tradi- tional methods—particularly the difference method—per- formed better in terms of detection speed, mostly because of the computational efficiency of simple brightness differ- ence calculations. In contrast, the detrending and ML approaches required additional processing steps such as polynomial fitting and model training, resulting in higher computational costs. The traditional methods evaluated here represent commonly used techniques reported in the literature; although they do not include all existing tradi- tional approaches, they provide a practical comparison framework. Despite their processing speed advantage, these methods exhibited lower accuracy. The consider- able improvement in detection performance achieved by our CNN-based model highlights the higher suitability of ML techniques—especially deep learning models—for reli- able stellar flare identification in complex LC data.

With this experiment, we validated the effectiveness of ML methods for stellar flare detection while highlight- ing the computational advantages of traditional methods.

Future directions of research can include investigating com- bined approaches to develop more efficient and accurate flare detection algorithms.

ML Performance For stellar flare detection tasks, the performance vari- ability of ML algorithms evaluated using accuracy, recall, and precision reflects their strengths and limitations for data feature extraction, temporal information capture, and overall detection accuracy. summarizes the perfor- mance of well-established ML algorithms (RF, XGBoost) Detrend Performance comparison of flare detection methods

and deep learning algorithms (CNN, RNN, DNN) in terms of accuracy, recall and precision.

Method

Application datasets Accuracy/(%) Recall/(%) Precision/(%) Feinstein et al. (CNN) TESS LCs Tu et al. (CNN) TESS pixel-level images Vida et al. (RNN) Kepler and LCs combined with simulated data Tu et al. (ensemble deep learning, Stacking) TESS pixel-level images Tu et al. (ensemble deep learning, Voting) TESS pixel-level images Lin et al. (DNN) TESS LCs Lin et al. (RF) TESS LCs Lin et al. (XGBoost) TESS LCs Feinstein et al. and Vida et al. used the LC flux as primary input feature. Feinstein et al. applied CNNs to LC data from TESS, achieving accuracy, recall, and pre- cision values of 97%, 97%, and 92%, respectively; their results indicated that, although the input data consisted only of raw LC flux data, their CNNs effectively cap- tured flare patterns with high predictive ability. Similarly, Vida et al. applied RNNs to TESS LC data and achieved accuracy, recall, and precision values of 80%, 80%, and 70%, respectively. Compared with the results of Feinstein et al. , the RNN performance was notably lower, possibly because of the model complexity and limita- tions in handling time-series data, particularly for flare detection over long timescales.

In contrast, Tu et al. used pixel-level TESS image data as input features for automatic feature extraction.

Their CNN achieved accuracy, recall, and precision val- ues of 96%, 96%, and 91%, respectively, demonstrating the advantage of CNNs for image data processing, with high efficiency for direct feature extraction and flare identi- fication. Moreover, Tu et al. further improved model per- formance by applying ensemble learning methods. Using the stacking method, they achieved accuracy, recall, and precision values of 97%, 96%, and 96%, respectively.

This ensemble learning method enhances overall perfor- mance by combining the strengths of multiple models.

With the ensemble learning voting method, they achieved their best results, with equally high accuracy, recall, and precision values of 99%. This indicates that combining pre- dictions from multiple models strongly improves flare iden- tification accuracy.

Lin et al. used LC flare features manually extracted from TESS data, including total flare duration, shock phase duration, decay phase duration, equivalent dura- tion, and peak amplitude, as input features for a DNN and achieved accuracy, recall, and precision values of 95%, 96%, and 94%, respectively; this indicates that com- bining manually computed features with deep learning mod- els can also yield high identification performance. Addition- ally, they evaluated RF and XGBoost algorithms and achieved accuracy, recall, and precision values of 97%, 97%, and 96%, respectively, for RF and 98%, 97%, and 98%, respectively, for XGBoost. The XGBoost algorithm exhibited outstanding (and the best overall) performance in handling data with structured features, particularly for such classification tasks.

In summary, the ensemble learning voting method of Tu et al. , relying on image data and combining predic- tions from multiple models, achieved highest perfor- mance, with strongly enhanced flare detection accuracy.

In contrast, deep learning models using only LC flux fea- tures, such as those of Feinstein et al. and Vida et , achieved comparatively poor performance, likely explained by know limitations of RNNs in capturing com- plex patterns. Finally, Lin et al. demonstrated that a com- bination of manually extracted features and established ML methods could also yield excellent performance.

DISCUSSION

Traditional methods for stellar flare detection primar- ily rely on threshold techniques and LC detrending.

Albeit simple and efficient, such methods are susceptible to noise interference. Astrophysical phenomena (such as star spots) or instrumental noise may produce FPs Low-amplitude flares may not reach the detection thresh- old and produce FNs With recent advances in ML and deep learning, meth- ods relying on feature engineering and data balancing have markedly improved flare detection accuracy. For exam- ple, Lin et al. , combining RF and XGBoost algorithms with SMOTE oversampling, achieved a flare recovery rate of 92% and additionally detected 2 000 small flare events . Deep learning methods, such as CNNs or “long short-term memory” neural networks, have further expanded model capabilities and overcome the limita- tions of traditional methods . Finally, ensemble learn- ing methods have provided, to date, the largest accuracy improvement for superflare detection, enhancing pixel- level classification accuracy to 99% . Despite the consid- erable potential of ML methods for stellar flare detection, there remain several challenges and limitations, described hereafter.

Data scarcity and imbalance: Stellar flare data repre- sent a very small proportion of the large observational datasets currently available, resulting in sample scarcity in

the ML model training data, particularly for low-energy flares. This sample imbalance can affect the model’s gener- alization performance, with good detection results for high-energy flares but difficult detection of low-energy events High rate of FP: LCs are affected by noise and interfer- ing astronomical phenomena (such as brightness peaks from pulsating stars) that can be misclassified as flares.

Even after data processing, such interferences can still affect deep learning models during flare detection, increas- ing the rate of FPs. This problem particularly affects large observational datasets with a substantial number of Long processing time and high computational cost:

The computational cost of deep learning models, espe- cially long short-term memory neural networks and CNNs, is high when processing large-scale LC data because of the time-consuming, but necessary, training and debugging, especially for parameter adjustment and hyperparameter optimization Noise interference and instrumental errors: The CCD sensors and data processing chains of spaceborne instru- ments introduce noise in the observation data, occasion- ally perceived erroneously as flare signals. Although ML models can be trained to filter this noise, the filtering is fre- quently incomplete and can cause misclassification of noise features as flares Challenges in cross-task data transfer: Observational data properties differ between space telescopes (such as Kepler and TESS), e.g., in terms of sampling rates and wavelength ranges. Therefore, a model trained on a spe- cific dataset, if applied to another dataset, frequently exhibits limited performance and requires adjustment or retraining to adapt to the new data Subjectivity in data labeling: Stellar flare event label- ing typically relies on visual inspection. However, this man- ual approach can be inconsistent, particularly when LCs exhibit low signal-to-noise ratios, which hampers human judgment. In such cases, flare labels may be assigned to ambiguous features, effectively introducing incompletely characterized labels—referred to as “noisy labels”—into the training dataset. These noisy labels can negatively affect the learning process and limit the model’s generaliza- tion ability These challenges reflect both the potential and the limi- tations of ML methods for stellar flare detection; they also indicate new research directions to improve models and data processing methods.

CONCLUSION

Stellar flare detection is an important astronomical research topic, investigated with either traditional or ML methods whose suitability depends on the application.

Before stellar flare catalogs became available or suffi- ciently extensive, early flare detection methods relied pri- marily on visual identification. Such traditional methods

use mathematical, unlabeled approaches; they are time-effi- cient and not labor-intensive. However, despite their sim- plicity and efficiency, their performance is considerably influenced by the accuracy of their initial settings (detec- tion threshold and detrending). ML can be used to opti- mize parameter selection (e.g., dynamic threshold adjust- ment) and improve detection sensitivity.

Because of their pattern recognition properties, ML methods can efficiently identify complex flare shapes and partly alleviate misclassification caused by human judg- ment errors. For example, they have demonstrably outper- formed traditional methods for the processing of large datasets, such as those from TESS or Kepler. However, their practical application still presents substantial chal- lenges. Model training relies on a large number of high- quality labeled samples; consequently, labeling of stellar flare events requires large volumes of observational data and expert cross-validation, which increases data acquisi- tion costs.

Therefore, to further generalize the application of ML techniques and enhance the accuracy of stellar flare detec- tion, we propose the following research directions:

(1) Intelligent transformation of detection methods Empirical threshold-based detection criteria are progres- sively replaced by ML algorithms, such as CNNs or XGBoost models, that improve identification accuracy and enhance generalization capabilities to weaker flares through autonomous learning of complex LC features, con- currently allowing for the automated processing and system- atic analysis of extensive datasets.

(2) Multisource data fusion and collaborative detection By combining high-quality data from multiple sources, e.g., from TESS for its high-precision photome- try data, Kepler for its long-term monitoring, and XMM- Newton for its multiwavelength X-ray data [ 61 ] , researchers can improve the signal-to-noise ratio of flare signals through cross-band feature correlation. Furthermore, multi- task collaborative analysis frameworks allow for simultane- ous investigation of flare activity, stellar magnetic fields, and stellar spot evolution, thereby maximizing the scien- tific output.

(3) High temporal resolution for dynamic process

Analysis

Current telescopes achieve very short exposure times, e.g., image acquisition by TESS at a 20-s cadence ; con- sequently, substructures of the stellar flare energy release can be resolved by analyzing these short-timescale data.

Combined with rapid-response spectroscopic and multiwave- length observations, this technological advance has been essential to characterize the temporal evolution of micro- physical processes such as magnetic reconnection and parti- cle acceleration, with essential theoretical breakthroughs on stellar eruption mechanisms.

Pursuing research on these combined topics will poten- tially induce considerable progress in stellar flare research and strongly expand our understanding of energy interac- tions between stars and their planetary systems.

ACKNOWLEDGEMENTS This work was supported by the National Natural Sci- ence Foundation of China (12473104 and U2031144).

AI DISCLOSURE STATEMENT AI-assisted tools such as StarWhisper and ChatGPT (OpenAI) were employed to translate the original Chi- nese text into English and refine the linguistic quality of this section. These tools were used solely for language improvement and did not contribute to data analysis, inter- pretation, or the formulation of scientific conclusions. The authors carefully reviewed, edited, and revised the AI-gener- ated texts to their own preferences, assuming ultimate responsibility for the content of the publication.

AUTHOR CONTRIBUTIONS Min Li conceived and designed the study, performed the experiments, analyzed the data, and wrote the paper.

Liang Wang provided key ideas, background knowledge, and contributed to data interpretation. Ying Shan guided and supported the experimental work. Zhiqiang Zou, Ali Luo, Bo Qiu, and Peng Jia reviewed the content and pro- vided valuable feedback on manuscript writing and aca- demic standards. All authors read and approved the final manuscript.

DECLARATION OF INTERESTS The authors declare no competing interests.

REFERENCES

Benz, A. O., Güdel, M. 2010. Physical processes in magnetically driven flares on the sun, stars, and young stellar objects.

Annual Review Astronomy Astrophysics (1): 241−287.

Walkowicz, L. M., Basri, G., Batalha, N., et al. 2011. White- light flares on cool stars in the Kepler quarter 1 data.

Astronomical Journal (2): 50.

Vida, K., Kriskovics, L., Oláh, K., et al. 2016. Investigating magnetic activity in very stable stellar magnetic fields-Long- term photometric and spectroscopic study of the fully convective M4 dwarf V374 Pegasi.

Astronomy & Astrophysics , 590 : A11.

Oláh, K., Seli, B., Kővári, Z., et al. 2022. Characteristics of flares on giant stars.

Astronomy & Astrophysics : A101. West, A. A., Hawley, S. L., Bochanski, J. J., et al. 2008.

Constraining the age-activity relation for cool stars: the sloan digital sky survey data release 5 low-mass star spectroscopic sample.

The Astronomical Journal (3): 785. Khodachenko, M. L., Ribas, I., Lammer, H., et al. 2007.

Coronal mass ejection (CME) activity of low mass M stars as an important factor for the habitability of terrestrial exoplanets. I. CME impact on expected magnetospheres of Earth-like exoplanets close-in habitable zones.

Astrobiology (1): 167−184. Yelle, R., Lammer, H., Ip, W. H. 2008. Aeronomy of extra- solar giant planets.

Space Science Reviews : 437−451.

Vida, K., Kővári, Z., Pál, A., et al. 2017. Frequent flaring in the TRAPPIST-1 system–unsuited for life?

The Astrophysical Journal , 841 (2): 124.

Roettenbacher, R. M., Kane, S. R. 2017. The stellar activity of TRAPPIST-1 and consequences for the planetary atmospheres.

The Astrophysical Journal (2): 77. Scalo, J., Kaltenegger, L., Segura, A., et al. 2007. M stars as targets for terrestrial exoplanet searches and biosignature detection.

Astrobiology (1): 85−166. Yelle, R. V. 2004. Aeronomy of extra-solar giant planets at small orbital distances.

Icarus (1): 167−179. Ricker, G. R., Vanderspek, R., Winn, J., et al. 2016. The transiting exoplanet survey satellite. In Proceedings of SPIE, 9904: 767−784.

Pietras, M., Falewicz, R., Siarkowski, M., et al. 2022.

Statistical analysis of stellar flares from the first three years of TESS observations.

The Astrophysical Journal (2): 143. Feinstein, A. D., Montet, B. T., Ansdell, M., et al. 2020.

Flare statistics for young stars from a convolutional neural network analysis of TESS data.

The Astronomical Journal (5): 219.

Vida, K., Bódi, A., Szklenár, T., et al. 2021. Finding flares in Kepler and TESS data with recurrent deep neural networks.

Astronomy & Astrophysics , 652 : A107.

Tu, Z. L., Wu, Q., Wang, W., et al. 2022. Convolutional neural networks for searching superflares from Pixel-level data of the transiting exoplanet survey satellite.

Astrophysical Journal (2): 90. Kalirai, J. 2018. Scientific discovery with the James Webb space telescope.

Contemporary Physics (3): 251−290. Guerrero, N., Seager, S., Huang, C. X., et al. 2021. The TESS objects of interest catalog from the TESS prime mission.

The Astrophysical Journal Supplement Series (2): 39.

Huang, C. X., Vanderburg, A., Pal, A., et al. 2020.

Photometry of 10 million stars from the first two years of TESS full frame images: part I.

Research Notes of the AAS (11): 204. Mullally, S. 2018. TESS Archive Manual, in TESS Archive Documentation Center, Space Telescope Science Institute. [Accessed 2024−10−25].

Oelkers, R., Stassun, K. 2018. Precision light curves from TESS full-frame images: A different imaging approach.

Astronomical Journal (3): 132. Hatt, E., Nielsen, M. B., Chaplin, W. J., et al. 2023.

Catalogue of solar-like oscillators observed by TESS in 120- s and 20-s cadence.

Astronomy & Astrophysics : A67.

Jackman, J. A., Shkolnik, E., Loyd, R. P. 2021. Stellar flares from blended and neighbouring stars in Kepler short cadence observations.

Monthly Notices of the Royal Astronomical Society , 502 (2): 2033−2042.

Meng, G., Zhang, L. Y., Su, T., et al. 2023. Properties of flare events on M stars from LAMOST spectral survey based on Kepler and TESS light curves.

Research in Astronomy and Astrophysics (5): 055001.

Davenport, J. R., Hawley, S. L., Hebb, L., et al. 2014.

Kepler flares. II. The temporal morphology of white-light flares on GJ 1243.

The Astrophysical Journal (2): 122. Gao, Q., Xin, Y., Liu, J. F., et al. 2016. White-light flares on close binaries observed with Kepler.

The Astrophysical

Journal Supplement Series (2): 37. Van Doorsselaere, T., Shariati, H., Debosscher, J. 2017.

Stellar flares observed in long-cadence data from the Kepler mission.

The Astrophysical Journal Supplement Series (2): 26.

Lu, H. P., Zhang, L. Y., Shi, J., et al. 2019. VizieR Online Data Catalog: M-type star magnetic activities from LAMOST & Kepler.

VizieR Online Data Catalog .

Huang, L. C., Ip, W. H., Lin, C. L., et al. 2020. M-dwarf Eclipsing Binaries with Flare Activity.

The Astrophysical Journal (1): 58. Günther, M. N., Zhan, Z. C., Seager, S., et al. 2020. Stellar flares from the first TESS data release: exploring a new sample of M dwarfs.

The Astronomical Journal (2): 60. Ilin, E. 2021. AltaiPony-Flare science in Kepler, K2 and TESS light curves.

Journal of Open Source Software (62): Yang, Z. L., Zhang, L. Y., Meng, G., et al. 2023. Properties of flare events based on light curves from the TESS survey.

Astronomy & Astrophysics : A15. Davenport, J. R. 2016. The Kepler catalog of stellar flares.

The Astrophysical Journal (1): 23. Yang, H. Q., Liu, J. F. 2019. The flare catalog and the flare activity in the Kepler mission.

The Astrophysical Journal Supplement Series (2): 29.

Brasseur, C., Osten, R. A., Tristan, I. I., et al. 2023.

Constraints on stellar flare energy ratios in the NUV and optical from a multiwavelength study of GALEX and Kepler flare stars.

The Astrophysical Journal (1): 5. Lurie, J. C., Davenport, J. R., Hawley, S. L., et al. 2015.

Kepler flares III: stellar activity on GJ 1245A and B.

Astrophysical Journal (2): 95. Shibayama, T., Maehara, H., Notsu, S., et al. 2013.

Superflares on solar-type stars observed with Kepler. I.

Statistical properties of superflares. The Astrophysical Journal Supplement Series (1): 5.

Okamoto, S., Notsu, Y., Maehara, H., et al. 2021. Statistical properties of superflares on solar-type stars: results using all of the Kepler primary mission data.

The Astrophysical Journal (2): 72. Wu, C. J., Ip, W. H., Huang, L. C. 2014. A study of variability in the frequency distributions of the superflares of G-type stars observed by the Kepler mission.

Astrophysical Journal (2): 92. Howard, W. S., MacGregor, M. A. 2022. No such thing as a simple flare: substructure and quasi-periodic pulsations observed in a statistical sample of 20 s cadence TESS flares.

The Astrophysical Journal (2): 204.

Chawla, N. V., Bowyer, K. W., Hall, L. O., et al. 2002.

SMOTE: synthetic minority over-sampling technique.

Journal of Artificial Intelligence Research , 16 : 321−357.

Lin, C. L., Apai, D., Giampapa, M. S., et al. 2024. Scalable, Advanced Machine Learning Based Approaches for Stellar Flare Identification: Application to TESS Short-cadence Data and Analysis of a New Flare Catalog.

Astronomical Journal (6): 234. Yamashita, R., Nishio, M., Do, R. K. G., et al. 2018.

Convolutional neural networks: an overview and application in radiology.

Insights into Imaging : 611−629.

Breiman, L. 2001. Random forests. Machine Learning , 45 : 5−32. [ 44 ]

Pedregosa, F., Varoquaux, G., Gramfort, A., et al. 2011.

Scikit-learn: Machine learning in Python. The Journal of Machine Learning Research , 12 : 2825−2830.

Chen, T. Q., Guestrin, C. 2016. XGBoost: A scalable tree boosting system. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.

Sun, Z. Y., Bobra, M. G., Wang, X. T., et al. 2022.

Predicting solar flares using CNN and LSTM on two solar cycles of active region data.

The Astrophysical Journal (2): 163. Liu, H., Liu, C., Wang, J. T., et al. 2019. Predicting solar flares using a long short-term memory network.

Astrophysical Journal (2): 121.

Kag, A., Saligrama, V. 2021. Training recurrent neural networks via forward propagation through time. In Proceedings of the 38th International Conference on Machine Learning.

Platts, J., Reale, M., Marsh, J., et al. 2022. Solar flare prediction with recurrent neural networks.

The Journal of the Astronautical Sciences (5): 1421−1440.

LeCun, Y., Bengio, Y., Hinton, G. 2015. Deep learning.

Nature , 521 (7553): 436−444. [ 51 ]

Baron, D. 2019. Machine learning in astronomy: A practical overview. arXiv: 1904.07248 . [ 52 ]

Shrestha, A., Mahmood, A. 2019. Review of deep learning algorithms and architectures.

IEEE Access , 7 : 53040−53065. [ 53 ]

Mahajan, P., Uddin, S., Hajati, F., et al. 2023. Ensemble learning for disease prediction: A review.

Healthcare (Basel) , 11 (12): 1808.

Mohammadifar, A., Gholami, H., Golzari, S. 2023.

Stacking-and voting-based ensemble deep learning models (SEDL and VEDL) and active learning (AL) for mapping land subsidence.

Environmental Science and Pollution Research , 30 : 26580−26595.

Luque, A., Carrasco, A., Martín, A., et al. 2019. The impact of class imbalance in classification performance metrics based on the binary confusion matrix.

Pattern Recognition , 91 : 216−231.

Zhou, Z. H. 2021. Machine Learning. Singapore: Springer Nature. [ 57 ]

Yang, H. Q., Liu, J. F., Gao, Q., et al. 2017. The flaring activity of M dwarfs in the Kepler field.

The Astrophysical Journal (1): 36. Althukair, A., Tsiklauri, D. 2023. Main sequence star super- flare frequency based on entire Kepler data.

Research in Astronomy and Astrophysics (8): 085017.

Feinstein, A. D., Seligman, D. Z., France, K., et al. 2024.

Evolution of flare activity in GKM stars younger than 300 Myr over five years of TESS observations. arXiv: 2405.

Stelzer, B., Caramazza, M., Raetz, S., et al. 2022. The Great Flare of 2021 November 19 on AD Leonis-Simultaneous XMM-Newton and TESS observations.

Astronomy & Astrophysics , 667 : L9.