| dc.description.abstract | Machine learning force fields (MLFF) have become gradually more popular within the field of material science as of late. Especially within molecular dynamics (MD) simulations have MLFFs seen prominent results, with both accuracy and efficiency comparable to traditional methods. In this study, a MLFF software called NeuralIL has been used to calculate the interatomic forces of a hybrid organic-inorganic perovskite material (DMMgF). Various NeuralIL architectures were explored, and several models showing promising results were selected for flexible cell MD simulations in the functional code Jax-MD. Finally, the accuracy of the NeuralIL-based MD was assessed by investigating whether the simulations could reproduce the phase transition of DMMgF in accordance with experimental data.
The interatomic force calculations provided by NeuralIL showcased the model's high capability to reproduce ab initio levels of accuracy. A mean absolute error of 0.020 eV/Å was the lowest seen in test sets that had configurations with ground truth forces calculated from density functional theory.
From a variety of NeuralIL architectures explored, a selection was chosen for Jax-MD simulations. Stability in the volume fluctuations was achieved for a single model, with the rest crashing or showing un-physical results. The results shed light on challenges related to training data and overfitting when using MLFF in MD simulations.
The phase transition of DMMgF was not shown in accordance with experimental data, although indications of structural changes concurrent with expectations were evident in some simulations. Possible weaknesses in the methodology are discussed as reasons, with special emphasis on the diversity of the training data. | |
