Finite Temperature Structure Prediction

Stanimir Bonev | 22-FS-014

Project Overview

The crystal structure of a solid is a basic characteristic and the starting point in determining material properties. Recent experimental measurements have discovered numerous temperature-driven phase transitions at high pressure, underlying the importance of including lattice dynamics in theoretical predictions of phase stability. However, direct structure prediction from first principles at finite temperature has remained a challenge. This project explores the feasibility of predicting solid phases stable at high pressure and finite temperature by performing large scale molecular dynamics simulations initiated in the liquid state. The approach exploits the similarities in short-range structural properties of compressed solids and liquids to train efficiently machine learning potentials using data generated from density functional theory molecular dynamics simulations of liquids. These potentials accurately describe structural properties in the vicinity of a target pressure and enable large scale molecular dynamics simulations where a liquid is quenched into a crystalline solid at finite temperature.

The feasibility of this approach is tested on dense lithium (Li), which has a rich phase diagram with multiple solid and liquid phase transitions as a function of pressure. We trained machine learning potentials on liquid Li data from 36 to 345 GPa pressure and established their accuracy in describing structural properties. The potentials were then used to quench liquids at four target pressures where Li crystal structures with the face centered cubic, cI16, oC40, and Oc88 space group symmetries are known to exist. We demonstrated that the liquids freeze into the correct crystalline structure at each of the four target pressures, which confirms the validity of the approach. Furthermore, we investigated the optimal pressure range for training the potentials, establishing that using data from a narrow pressure range improves the predictability of the method at that target pressure. Finally, we studied the sensitivity of the results to the rate of quenching, establishing a reliable range for this parameter. Our results demonstrate the feasibility of this novel approach for predicting high pressure solid phases.

Mission Impact

The project establishes the feasibility for developing a new capability for predictive simulations in the regime of high energy density, which will have an immediate impact on the design and interpretation of high-pressure experiments in support of the Stockpile Stewardship national security mission. It enhances Lawrence Livermore National Laboratory's Core Competencies in High Energy Density Science, Simulation and Data Science. The results of this project further key Lab investments in predictive first principles simulations, equation of state modeling and machine learning. The project has led to continued funding through follow-on LDRD-ER.

Publications, Presentations, and Patents

J. Chapman and S.A. Bonev, "Finite-temperature Crystal Structure Prediction of Lithium Using Machine Learning Potentials" (Presentation, 2022 MRS Spring Meeting and Exhibit, Honolulu, HI and Virtual Meeting, 2022).

A. Mukund, J. Chapman, and S.A. Bonev, "Crystal structure prediction starting from a liquid using machine-learning potentials" (Presentation, 2023 APS March Meeting, Las Vegas, NV, March 2023).