PREDIKSI MULTI-PROPERTI MATERIAL PEROVSKITE ABX3 MENGGUNAKAN XGBOOST DENGAN SELEKSI FITUR METAHEURISTIK

Faizah, Wanda Nur (2026) PREDIKSI MULTI-PROPERTI MATERIAL PEROVSKITE ABX3 MENGGUNAKAN XGBOOST DENGAN SELEKSI FITUR METAHEURISTIK. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002221158-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (13MB) | Request a copy

Abstract

Material perovskite ABX3 memiliki potensi besar sebagai material sel surya karena efisiensinya yang tinggi dan biaya produksinya yang relatif rendah. Namun, penerapannya masih terkendala oleh rendahnya stabilitas serta toksisitas timbal (Pb) pada perovskite berbasis halida. Perovskite kalkogenida (S, Se, dan Te) menjadi alternatif yang menjanjikan, tetapi eksplorasinya melalui eksperimen maupun perhitungan Density Functional Theory (DFT) memerlukan waktu dan biaya yang relatif tinggi. Oleh karena itu, penelitian ini menerapkan pembelajaran mesin untuk memprediksi lima properti material perovskite ABX3, yaitu kestabilan termodinamika, jenis celah pita (direct/indirect band gap), konduktivitas, energi formasi, dan band gap, berdasarkan komposisi kimianya. Data diperoleh dari Materials Project Database, dengan 1.673 material yang digunakan setelah melalui tahap pra-pemrosesan dari total 3.735 material awal. Fitur diekstraksi menggunakan metode Meredig, Magpie, dan MEGNet Element Embedding, kemudian dimodelkan menggunakan XGBoost. Ketidakseimbangan kelas ditangani menggunakan SMOTE, optimasi hyperparameter dilakukan menggunakan RandomizedSearchCV, dan seleksi fitur dilakukan menggunakan dua metode metaheuristik, yaitu Particle Swarm Optimization (PSO) dan Genetic Algorithm (GA). Hasil penelitian menunjukkan bahwa performa PSO dan GA sebagai metode seleksi fitur bervariasi tergantung pada target prediksi. Dari kelima target yang diuji, GA memberikan performa terbaik pada dua target, yaitu konduktivitas dengan F1-score sebesar 0,8599 dan celah pita dengan R2 sebesar 0,8493. PSO memberikan performa terbaik pada dua target lainnya, yaitu kestabilan termodinamika dengan F1-score sebesar 0,8000 dan energi formasi dengan R2 sebesar 0,9640. Sementara itu, pada prediksi sifat celah pita, model baseline tanpa seleksi fitur tetap memberikan performa terbaik dibandingkan kedua metode seleksi fitur. Hasil penelitian ini menunjukkan bahwa pendekatan melalui komposisi kimia dengan seleksi fitur metaheuristik berpotensi mendukung proses identifikasi material perovskite baru tanpa memerlukan informasi struktur kristal maupun perhitungan DFT tambahan.
====================================================================================================================================
The ABX3 perovskite materials have great potential for solar cell applications due to their high efficiency and relatively low production costs. However, their application is still hampered by low stability and the toxicity of lead (Pb) in halide-based perovskites. Chalcogenide perovskites (S, Se, and Te) offer a promising alternative, but their exploration through experiments and Density Functional Theory (DFT) calculations requires a relatively high investment of time and resources. Therefore, this study applies machine learning to predict five properties of ABX3 perovskite materials namely, thermodynamic stability, band gap nature (direct/indirect), conductivity, formation energy, and band gap based on their chemical composition. Data were obtained from the Materials Project Database, yielding 1,673 materials for use after preprocessing an initial set of 3,735 materials. Features were extracted using the Meredig, Magpie, and MEGNet Element Embedding methods, and subsequently modeled using XGBoost. Class imbalance was addressed using SMOTE, hyperparameter optimization was performed using RandomizedSearchCV, and feature selection was carried out using two metaheuristic methods, Particle Swarm Optimization (PSO) and Genetic Algorithm (GA). The results showed that the performance of PSO and GA as feature selection methods varied depending on the prediction target. Among the five targets evaluated, GA achieved the best performance on two targets, conductivity with an F1-score of 0.8599, and band gap value with an R2 of 0.8493. PSO achieved the best performance on two other targets, thermodynamic stability with an F1-score of 0.8000, and formation energy with an R2 of 0.9640. Meanwhile, for band gap nature prediction, the baseline model without feature selection remained superior to both feature selection methods. The findings of this study indicate that a chemical composition-based approach combined with metaheuristic feature selection has the potential to support the identification of new perovskite materials without requiring crystal structure information or additional DFT calculations.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Material perovskite ABX3, XGBoost, SHAP, seleksi fitur meta-heuristik, kalkogenida, stabilitas, band gap, fotovoltaik. ABX3 perovskite materials, XGBoost, SHAP, metaheuristic feature selection, chalcogenides, stability, band gap, photovoltaics.
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines.
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Wanda Nur Faizah
Date Deposited:	30 Jul 2026 08:50
Last Modified:	30 Jul 2026 08:50
URI:	http://repository.its.ac.id/id/eprint/140080

Actions (login required)

View Item