توصيفگر ها :
كتابخانه‌هاي طيفي ستاره‌اي , تابع پايه شعاعي , شبكه عصبي مصنوعي , جنگل تصادفي
چكيده فارسي :
در اين پايان‌نامه از كتابخانه‌هاي طيفي ستاره‌اي تجربي براي درون‌يابي طيف‌هاي ستاره‌اي بر اساس سه پارامتر جوي ستاره‌اي استفاده شده است. از آنجايي كه توزيع ستارگان در اين كتابخانه‌ها نامنظم است، درون‌يابي‌هاي سنتي نتايج دقيقي نخواهند داشت. بنابراين شبكه تابع پايه شعاعي و ساختار آن مرور شده است. همچنين مروري بر محدوديت اعمال شده بر روي تابع هسته گاوسي انجام شده است تا بهبودي براي شبكه تابع شعاعي باشد. در اين كار دو تكنيك يادگيري ماشين يعني شبكه عصبي مصنوعي و جنگل تصادفي براي درون يابي طيف‌هاي ستاره اعمال شده‌اند و نتايج آن ها با نتايج روش هاي درون‌يابي ديگر مقايسه شده است. نشان مي‌دهيم كه به طور كلي مدل‌هاي شبكه عصبي و جنگل تصادفي نسبت به روش هاي ديگر به درون‌يابي بهتري دست مي‌يابند.
چكيده انگليسي :
In this thesis, empirical stellar spectral libraries have been used for the interpolation of stellar spectra based on three stellar atmospheric parameters. Since the distribution of stars in these libraries is irreg- ular, traditional interpolations will not yield accurate results. First, the radial basis function (RBF) network and its structure have been reviewed. In stellar population synthesis models, the existence of empirical stellar spectral libraries is essential for the integrated spectra of stellar populations. The up- graded RBF network and its comparison with other kernel methods like SPH have been investigated. A constraint related to the kernel function in the RBF network has been added, which describes the relationship between the σ in the Gaussian kernel function of RBF network and the spatial density of samples in the parameter space. Additionally, an anisotropic kernel function has been considered by relating it to the inhomogeneous distribution of stars in the stellar atmospheric parameter space. Axial local dispersion has also been used to determine this anisotropic kernel function. Three control parameters, c0, c1, and c2, are obtained for spectral interpolation calculation. The beetle antennae search algorithm is used to find the best value for these parameters. The semi-empirical BaSeL-3.1 library is used for training, and the MILES library is used for testing. For testing, the comparison between the original and interpolated spectra for each star in the MILES library is used. The up- graded RBF interpolator generally performs well except at the edge of the low-temperature region. This undesirable performance at the edge of the low-temperature region can be due to incomplete spectral coverage in the stellar atmospheric parameter space, insufficient parameters, and inconsistent atmospheric parameters. The RBFup interpolator shows a significant improvement over the RBF18 except at the edge of the low-temperature region. Both interpolators do not perform well in this region. Then, two machine learning techniques, namely artificial neural networks (ANN) and random forests (RF), have been applied for stellar spectrum interpolation, and their results have been com- pared with other interpolation methods. A fully-connected neural network model and a random forest model have been trained to produce a stellar spectrum for a given set of three stellar atmospheric parameters. The empirical spectral library MILES is used for training, and the CFLIB library is used for testing. By comparing various interpolations RBF, TGM2, ANN, and RF, we demonstrated that the ANN and RF models perform significantly better than RBF and TGM2 interpolations. The performance of the ANN and RF models is comparable. Error statistics indicate overall high-quality reconstruction. By examining the most inconsistent cases with the largest residuals, it was found that only 2% of the spectra in the entire training set were not satisfactorily reconstructed. Additionally, to test the robustness of the model, Gaussian noise from 1% to 10% is simultaneously introduced to the stellar atmospheric parameters. The analysis showed that the presented ML models are fully robust to small errors in the input atmospheric parameters, and their behavior does not change significantly with the introduction of noise.
