Using stacked ensemble learning, a UK research team achieved high-precision prediction of the asteroseismic indices of 251 Delta Scuti stars.
A research team from the University of Warwick in the UK has developed a stacked ensemble learning framework to directly predict the key asteroseismic parameters of δ Scuti stars from TESS light curves. This method has achieved remarkable results on a sample of 643 stars: the coefficient of determination R² for all target parameters is higher than 0.77, and it shows good generalization ability on 60 stars that were not involved in the training. The prediction results are highly consistent with traditional asteroseismic analysis.
Asteroseismology is one of the most penetrating research methods in modern stellar physics. It analyzes the natural oscillation signals of stars to invert their internal structure and evolutionary state. Among many research objects, δ Scuti stars (with a mass about 1.5–2.5 times that of the Sun) have become an important experimental field for asteroseismology due to their rich pulsation modes and highly dense oscillation spectra. The pulsation of these stars is mainly driven by the opacity (κ) mechanism in the helium ionization zone, and their internal convective cores further trigger complex processes such as convective overshoot, chemical mixing, and angular momentum redistribution. At the same time, the relatively fast rotation causes the oscillation modes to couple and the frequencies to split, greatly increasing the difficulty of mode identification and parameter extraction.
In asteroseismic analysis, parameters such as the frequency corresponding to the highest peak in the power spectrum, the frequency of the maximum oscillation power, and the large frequency spacing Δν are particularly important. Among them, Δν is extremely sensitive to the average density of the star and is a core indicator for characterizing its overall structure. However, for δ Scuti stars, the rapid rotation and multi - mode aliasing disrupt the originally regular frequency spacing, posing significant challenges to traditional methods in measuring Δν.
In recent years, the large - scale, high - precision light curve data obtained by the TESS satellite have greatly expanded the research samples of this type of stars. However, the data processing process is still computationally intensive and experience - dependent, and it is still not easy to achieve high - precision extraction of parameters. Against this background, machine learning provides a new technical path. Compared with traditional methods, ensemble learning can integrate the prediction results of multiple models to achieve higher accuracy and stability in complex data environments. Methods such as random forest, gradient boosting, and ridge regression have shown good potential in astronomical data analysis in recent years.
Based on this idea, the research team from the University of Warwick in the UK developed a stacked ensemble learning framework to directly predict the key asteroseismic parameters of δ Scuti stars from TESS light curves. This method has achieved remarkable results on a sample of 643 stars: the coefficient of determination R² for all target parameters is higher than 0.77, and it shows good generalization ability on 60 stars that were not involved in the training. The prediction results are highly consistent with traditional asteroseismic analysis.
The relevant research results, titled "Ensemble Machine Learning Approach to Estimate the Asteroseismic Indices for δ Scuti Stars Observed by TESS", have been published in The Astronomical Journal.
Research Highlights:
* A machine learning framework for directly estimating key asteroseismic parameters from light curves is proposed, breaking through the limitations of traditional methods and significantly improving the efficiency of parameter extraction.
* High - precision prediction is achieved through optimized feature selection and model architecture, and its reliability is verified on independent samples.
* The asteroseismic indices of 251 δ Scuti stars are determined, a new star catalog is constructed, and the parameter database of relevant stars is enriched, providing important data support for future large - sample statistical analysis and stellar evolution research.
Paper URL: https://beta.iopscience.iop.org/article/10.3847/1538-3881/ae4bd8
Dataset: TESS Light Curve Screening and Asteroseismic Sample Construction
The core dataset used in this study contains the TESS light curves of 643 δ Scuti stars, as well as three key asteroseismic indices: ν(Aₘₐₓ), νₘₐₓ, and Δν. The initial sample included 677 δ Scuti stars, and after multiple rounds of screening, 643 were retained as the core dataset. The screening criteria include: having TESS 2 - minute short - exposure light curves (from the MAST archive); having no less than 7,000 data points in each observation field; the light curves being processed by PDC - SAP correction; and the three asteroseismic parameters being completely available.
On this basis, the researchers additionally selected 251 δ Scuti stars as a supplementary sample. These stars also have high - quality light curves, but their corresponding asteroseismic parameters have not been published. The criteria for their selection are: covering at least 3 observation fields, and having no less than 7,000 data points in each field. This part of the sample is mainly used for the actual prediction and verification of the model.
Frequency histogram of 643 δ Scuti stars
Model: An Ensemble Regression Framework with Multiple Base Models Stacked
The goal of the model in this study is to estimate the asteroseismic parameters of stars based on the features of light curves. The overall process includes feature extraction, data pre - processing, ensemble modeling, and hyperparameter optimization.
In terms of feature construction, two types of features are used: one type is statistical features (such as mean, standard deviation, median, etc.) to describe the basic properties of the photometric distribution; the other type is frequency - domain features, including principal component analysis (PCA), autocorrelation function (ACF), fast Fourier transform (FFT), and discrete wavelet transform (DWT), to extract the periodic and multi - scale structure information in the oscillation signals.
In the data pre - processing stage, samples with missing values are first removed, and the features are normalized. In addition, to address the problem of uneven distribution of some features, a resampling method based on statistical distribution is introduced to generate synthetic data and reduce the bias, thereby improving the stability of model training.
In terms of the framework, the model uses a stacked ensemble regression framework, with random forest, gradient boosting regression, and ridge regression as base models: the first two improve the prediction performance from the perspectives of reducing variance and bias respectively, and ridge regression addresses the collinearity problem between features through regularization. The outputs of the base models are further used as inputs to train a meta - regressor for fusion, thereby improving the overall generalization ability and reducing the prediction error.
During the model training process, the researchers also use random search combined with cross - validation to optimize the key hyperparameters (such as the number of trees, maximum depth, and learning rate) to obtain a stable and high - performance model configuration.
Testing Generalization with 60 Independent Stars, R² > 0.77 for All Asteroseismic Indices
The experimental verification includes three parts: model training, generalization ability evaluation, and prediction of new samples.
During the training stage, the researchers randomly selected 583 stars from the 643 stars for model construction, and divided the training set and test set at a ratio of 8:2, repeating this 100 times to reduce the impact of randomness. The remaining 60 stars were used as an independent test set to evaluate the generalization ability of the model. In addition, 251 unlabeled samples were used for the final prediction.
Comparison of the measured and predicted values, relative errors, and error distributions of 583 stars
On the training and test samples, the R² values of the model's predictions for ν(Aₘₐₓ), νₘₐₓ, and Δν are 0.95, 0.93, and 0.87 respectively, and the relative errors of most samples are less than 0.2. Feature importance analysis shows that the autocorrelation function (ACF) contributes the most, followed by FFT and DWT, and some statistical features (such as skewness and kurtosis) also play a certain role. The learning curve shows that the model converges stably, and the hyperparameter optimization is effective.
Model learning curve
On the independent test set, the model still maintains good performance. The R² values of the three parameters are 0.91, 0.87, and 0.77 respectively, and the prediction results are highly consistent with the observed values. The results of multiple repeated experiments show little fluctuation, indicating that the model has good stability and robustness. Finally, the researchers applied the model to 251 unlabeled stars and obtained the predicted values of their asteroseismic parameters. The results generally fall within the reasonable range of δ Scuti stars.
Conclusion
Overall, this work is not a replacement for traditional asteroseismic methods but a targeted supplement: in the context of the rapid accumulation of large - scale observational data, efficient parameter estimation is achieved through data - driven methods, and then in - depth analysis is carried out combined with fine physical modeling. This approach is particularly relevant for targets like δ Scuti stars, which have complex oscillation modes and are difficult to standardize.
This article is from the WeChat official account "HyperAI Super Neural". Author: Tian Xiaoyao. Republished by 36Kr with permission.