[Paper Review] Towards Universal Sparse Gaussian Process Potentials: Application to Lithium Diffusivity in Superionic Conducting Solid Electrolytes

Amir Hajibabaei, Chang Woo Myung+1 more|arXiv (Cornell University)|Sep 28, 2020

Machine Learning in Materials Science4 citations

TL;DR

This paper introduces a data-efficient, adaptive sampling algorithm within sparse Gaussian process regression to develop universal interatomic potentials for solid electrolytes. It successfully reproduces experimental melting and glass-crystallization temperatures in Li7P3S11 and identifies a previously uncharacterized phase with significantly reduced lithium diffusivity, enabling scalable modeling via hierarchical expert model combinations.

ABSTRACT

For machine learning of interatomic potentials the sparse Gaussian process regression formalism is introduced with a data-efficient adaptive sampling algorithm. This is applied for dozens of solid electrolytes. As a showcase, experimental melting and glass-crystallization temperatures are reproduced for Li7P3S11 and an unchartered infelicitous phase is revealed with much lower Li diffusivity which should be circumvented. By hierarchical combinations of the expert models universal potentials are generated, which pave the way for modeling large-scale complexity by a combinatorial approach.

Motivation & Objective

• To create data-efficient, universal interatomic potentials for complex solid electrolytes using machine learning.
• To reproduce experimental thermal behavior (melting and glass-crystallization) in superionic conductors like Li7P3S11.
• To identify previously uncharacterized phases with poor ionic conductivity that could hinder performance.
• To enable large-scale modeling of complex materials through hierarchical combination of expert models.

Proposed method

• Adopting sparse Gaussian process regression to model interatomic interactions with high data efficiency.
• Implementing an adaptive sampling algorithm that prioritizes data acquisition in regions of high uncertainty.
• Training individual expert models on diverse solid electrolyte configurations to capture local chemical and structural complexity.
• Combining expert models hierarchically to construct universal potentials applicable across diverse materials.
• Using atomic environment descriptors as input features to the Gaussian process to encode local structure.
• Leveraging uncertainty estimates from the GP to guide active learning and improve sample efficiency.

Experimental results

Research questions

• RQ1 Can sparse Gaussian process regression with adaptive sampling accurately reproduce experimental thermal transitions in superionic conductors?
• RQ2 What are the structural and dynamical properties of previously uncharacterized phases in solid electrolytes?
• RQ3 How can expert models trained on specific materials be hierarchically combined to form universal interatomic potentials?
• RQ4 To what extent does the method reduce data requirements for reliable interatomic potential learning?

Key findings

• The method successfully reproduces experimental melting and glass-crystallization temperatures in Li7P3S11.
• A previously uncharted phase with significantly reduced lithium diffusivity was discovered, indicating a potential kinetic trap in the material landscape.
• The hierarchical combination of expert models enables the construction of universal potentials for complex materials systems.
• The adaptive sampling strategy achieves high data efficiency by focusing on regions of high uncertainty in the potential energy surface.