Aplikasi Machine Learning Dalam Meramalkan Properti Baterai Berbasis Ion Natrium

Abstract

Penelitian ini berfokus pada aplikasi machine learning dalam meramalkan properti baterai berbasis ion natrium. Menggunakan data dari situs Material Project sebanyak 309 data. Penelitian ini mengkaji prediksi terhadap properti average voltage, gravimetric capacity, volumetric capacity, gravimetric energy, dan volumetric energy. Fitur-fitur yang digunakan meliputi formula, spacegroup, Nsites, E above hull, formation energy, volume, density, dan bandgap. Untuk mendapatkan fitur struktur, unsur-unsur dalam fitur formula diurai menggunakan library Chemparse. Tiga kombinasi model Artificial Neural Network diterapkan: kombinasi pertama menggunakan fitur propert material, kombinasi kedua menggunakan fitur seluruh fitur properti dan unsur yang diurai, sementara kombinasi ketiga hanya menggunakan fitur unsur yang diurai saja. Evaluasi performa model dilakukan menggunakan metrik R2 score, Mean Squared Error, dan Root Mean Squared Error. Hasil penelitian menunjukkan bahwa kombinasi kedua yang menggunakan seluruh fitur mendapatkan performal yang paling baik. Selain itu, didapatkan temuan bahwa struktur material berupa unsur yang diurai mampu menguatkan performa model.

This research focuses on the application of machine learning to predict the properties of sodium-ion-based batteries. Using data from the Material Project, comprising 309 entries. This study examines the prediction of properties such as average voltage, gravimetric capacity, volumetric capacity, gravimetric energy, and volumetric energy. The features used include formula, spacegroup, Nsites, E above hull, formation energy, volume, density, and bandgap. To obtain structural features, the elements in the formula feature were parsed using the Chemparse library. Three combinations of Artificial Neural Network models were applied: the first combination used material property features, the second combination used all property features and parsed elements, while the third combination used only the parsed elements. The performance of the models was evaluated using R2 score, Mean Squared Error, and Root Mean Squared Error metrics. The results indicate that the second combination, which utilized all features, achieved the best performance. Additionally, it was found that the structural features in the form of parsed elements were able to enhance the model's performance.