EMPLOYING MACHINE LEARNING POTENTIAL TO ELUCIDATE NON-STOICHIOMETRIC FES STRUCTURE

• Main Author: Rizky Rahman, Muhammad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/71731
• Institution: Institut Teknologi Bandung

Iron (II) sulfide (FeS) is of interest due to their potential usage in a variety of industries, including energy storage, catalysis, and environmental remediation. FeS is a Metal-Insulator Transition (MIT) material in which its electronic properties can be modified as initial temperature treatment. FeS has P62c space group at below 400 K and P63/mmc at above 590 K. Nevertheless, since a stoichiometric ratio is challenging to attain, multistep methods are needed to synthesize FeS. In this study, we employ artificial neural network (ANN) based on Behler-Parinello approach to investigate the phase changes of FeS at 300 – 600 K. Currently, we try to elucidate the best multilayer perceptron (MLP) ANN structure to construct adequate atomic interaction potential of FeS. The potentials are constructed from more than 4000 database structures, which include pristine and defected structures. We obtain that 2 hidden layers MLP with 20 nodes in each show fine training and testing set error. We do force minimization of particular FeS structure using ANN atomic potential. We obtain excellent results as no overlapping Fe-S bond length in the final structure. The resulting structure in the finite temperature simulation shows a changed structure but remains within reasonable limits.