- Home
- Institut Néel
- Research teams
- Technical Groups & Services
- Work at the Institut
- Partnerships
The defence will be in English.
Abstract: Recent advances in machine learning and artificial intelligence have brought upon a paradigm shift, which also lead to big implications for materials simulations. Traditional methods have often faced challenges with performing accurate atomistic simulations of complex materials, such as GeSbTe, due to the trade-off between simulation accuracy and computational cost. Machine learning interatomic potentials (MLIPs) have recently been presented as a solution to this problem, which offer accurate simulations with high computational efficiency.
In this work, we evaluate the viability of Behler and Parrinello’s high-dimensional neural network potential (HDNNP) and E(3)-equivariant neural network potential for large-scale molecular dynamics (MD) simulations. After an overview of the necessary background information on neural networks and their training, a library for training HDNNPs and performing MD simulations that was developed during this work is introduced. The ability of HDNNPs to compute much larger structures than the training structures is demonstrated. Finally, the performance gain and accuracy of the predictions are surveyed against density functional theory (DFT).
Issues of extrapolation on out-of-distribution (OOD) data are also addressed by examining three uncertainty estimation techniques. Active learning, a method for automatically building a robust training dataset using OOD data, is discussed and investigated for SiGe alloy systems using the committee approach.
Finally, neural network potentials are used to perform large MD simulations of silicon, germanium and germanium telluride (GeTe) in order to compute their lattice thermal conductivities using the Green-Kubo method. A comparison of the predictions to experimentally observed values is made.
