Machine-Learning Accelerated First-Principles Accurate Modeling of the Solid–Liquid Phase Transition in MgO under Mantle Conditions

Abstract

While accurate measurements of MgO under extreme high-pressure conditions are needed to understand and model planetary behavior, these studies are challenging from both experimental and computational modeling perspectives. Herein, we accelerate density functional theory (DFT) accurate calculations using deep neural network potentials (DNPs) trained over multiple phases and study the melting behavior of MgO via the two-phase coexistence (TPC) approach at 0–300 GPa and ≤9600 K. The resulting DNP–TPC melting curve is in excellent agreement with existing experimental studies. We show that the mitigation of finite-size effects that typically skew the predicted melting temperatures in DFT–TPC simulations in excess of several hundred kelvin requires models with ∼16 000 atoms and >100 ps molecular dynamics trajectories. In addition, the DNP can successfully describe MgO metallization well at increased pressures that are captured by DFT but missed by classical interatomic potentials.

Predicting melting temperatures across the periodic table with machine learning atomistic potentials

“…Interestingly, we could not nd any DFT (PBE) calculations in the literature for comparison, and thus, this work may be a rst glimpse of what a pressure-dependent T m curve would show for DFT. These hightemperature and high-pressure calculations and experiments are difficult to conduct, 104 exemplied by the few literature reports on T m and pressure dependence for each method. As the pressure increases, the deviation between DNP/DFT and experiment decreases; this observation has been noted in previous DFT melting simulations for a variety of materials 3,30,71,74 and is consistent with lower accuracy of rst-principles modeling using PBE functionals at pressures close to zero.…”

Editorial: Computational modelling of cardiovascular hemodynamics and machine learning

“…The neural network predictions are best described by conditional Bayesian statistics, which depend not only on the statistical distributions (variance, mean, outliers, missing values) of the data sets (e.g., feature space), but also on the modelling complexity (e.g., weight distribution) ( 6 – 8 ). There is a well-known trade-off between variance and bias which is determined by the modelling complexity and size of the feature space ( 6 – 8 ). Increasing complexity of the model at constant input, reduces the variance and improves training predictions but tend to increase the bias thereby reducing the generalizability of the model to new data sets.…”