Prediksi Multi-Properti Double Perovskite Menggunakan LightGBM dengan Seleksi Fitur Algoritma Genetika dan Particle Swarm Optimization

Artha, Selly Purnama (2026) Prediksi Multi-Properti Double Perovskite Menggunakan LightGBM dengan Seleksi Fitur Algoritma Genetika dan Particle Swarm Optimization. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 50002221118-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (12MB) | Request a copy

Abstract

Material double perovskite dengan formula A2BB′X6 dan AA′BB′X6 merupakan salah satu material yang dikaji untuk aplikasi fotovoltaik. Penelitian ini bertujuan menerapkan Light Gradient Boosting Machine (LightGBM) untuk memprediksi lima target, yaitu sifat konduktivitas, stabilitas termodinamika, tipe band gap, nilai band gap, dan formation energy, serta mengevaluasi pengaruh seleksi fitur menggunakan Genetic Algorithm (GA) dan Particle Swarm Optimization (PSO). Data material diperoleh dari Materials Project, sedangkan ekstraksi fitur menggunakan deskriptor Meredig, Magpie, dan MEGNet menghasilkan 347 fitur berbasis komposisi material. Optimasi hyperparameter dilakukan menggunakan RandomizedSearchCV dengan 5-fold cross validation, serta SMOTE diterapkan pada target klasifikasi dengan distribusi kelas tidak seimbang. Hasil seleksi fitur menunjukkan bahwa GA menghasilkan 138–180 fitur terpilih, sedangkan PSO menghasilkan 172–181 fitur terpilih. Pada data uji, model dengan fitur hasil seleksi GA memperoleh nilai accuracy sebesar 0,9081 untuk target is conductor dan thermodynamic stability, serta 0,8728 untuk target is direct. Pada target regresi, diperoleh nilai R2 sebesar 0,8673 untuk band gap dan 0,8860 untuk formation energy. Analisis SHAP menunjukkan bahwa fitur yang berkontribusi terhadap hasil prediksi berbeda pada setiap target. Secara keseluruhan, GA dan PSO mampu mengurangi jumlah fitur dengan tetap menghasilkan performa prediksi yang baik, sedangkan GA menunjukkan peningkatan performa yang lebih konsisten pada penelitian ini.
====================================================================================================================================
Double perovskite materials with the formulas A2BB′X6 and AA′BB′X6 are among the materials being studied for photovoltaic applications. This study aims to apply the Light Gradient Boosting Machine (LightGBM) to predict five targets, namely conductivity, thermodynamic stability, band gap type, band gap value, and formation energy, as well as to evaluate the effect of feature selection using the Genetic Algorithm (GA) and Particle Swarm Optimization (PSO). Material data were obtained from the Materials Project, while feature extraction using the Meredig, Magpie, and MEGNet descriptors yielded 347 features based on material composition. Hyperparameter optimization was performed using RandomizedSearchCV with 5-fold cross-validation, and SMOTE was applied to classification targets with an imbalanced class distribution. Feature selection results showed that GA yielded 138–180 selected features, while PSO yielded 172–181 selected features. On the test data, the model using features selected by GA achieved an accuracy of 0.9081 for the is conductor and thermodynamic stability targets, and 0.8728 for the is direct target. For the regression targets, an R2 value of 0.8673 was obtained for band gap and 0.8860 for formation energy. SHAP analysis showed that the features contributing to the prediction results differed for each target. Overall, both GA and PSO were able to reduce the number of features while still achieving good prediction performance, whereas GA demonstrated more consistent performance improvements in this study.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Double Perovskite, LightGBM, Seleksi Fitur, Feature Selection, Genetic Algorithm, ParticleSwarm Optimization, SHAP.
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines.
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics
Depositing User:	Selly Purnama Artha
Date Deposited:	30 Jul 2026 09:05
Last Modified:	30 Jul 2026 09:05
URI:	http://repository.its.ac.id/id/eprint/140081

Actions (login required)

View Item