Materials modeling by ab initio methods and machine learning interatomic potentials: a critical assessment
Publication or External Link
Atomistic modeling is a crucial research technique in materials science to simulate physical phenomena based on the interactions of atoms. Density functional theory (DFT) calculation has been the standard technique for evaluating atom interactions, but their applications are limited due to high computation cost. Classical interatomic potential, as another widely adopted technique, provides an alternative by enabling large-scale simulations that are computationally inexpensive. However, the employment of classical potentials for cheap atomistic simulations is achieved in exchange of accurately evaluating atom interactions and the transferability to different chemistries. Most recently, machine learning interatomic potential (MLP) emerges as a new computation technique to bridge the gap between first-principles computation and classical potentials.I first utilized the atomistic modeling based on DFT calculations to find novel Li superionic conductor, a key component of the emerging all-solid-state Li-ion battery technology. I performed a systematic study on the Li-ion conduction of lithium chloride materials system and predicted a dozen potential Li superionic conductors. I revealed that the Li-ion migration in the materials is greatly impacted by the Li content, the cation configuration, and the cation concentrations. I further demonstrated tuning these three factors in designing new chloride Li-ion conductors.
Then, I studied the atomistic dynamics predicted by MLPs in comparison to DFT calculations to answer the open question whether MLPs can accurately reproduce dynamical phenomena and related physical properties in molecular dynamics simulations. I examined the current state-of-the-art MLPs and uncovered a number of discrepancies related to atomistic dynamics, defects, and rare events compared to DFT methods. I found that testing averaged errors of MLPs are insufficient and developed evaluation metrics that better indicate the accurate prediction of related properties by MLPs in MD simulations. I further demonstrated that the MLPs optimized by the proposed evaluation metrics have improved prediction in multiple properties. I also study the performance of MLPs in assessing the elemental orderings in a large variety of phases across composition range in alloy system. Using the Li-Al alloy system as a case study, I trained MLPs using only a few phases and the trained MLP demonstrated good performances over a number of existing and other hypothetical materials in- and out- of the training data across the Li-Al binary alloy system. We developed several new evaluation metrics on energy rankings to evaluate the elemental ordering, which is critical for studying the phase stabilities of materials. I tested MLP transferability to other phases and the limits of MLP applications on commonly performed simulations. I also studied the effect of diverse training data on MLP performances.
With these efforts evaluating MLP performances for a number of properties, metrics, prediction errors, and dynamical phenomena, I constructed a dataset with a large number of MLPs and their performances and performed an empirical analysis to identify the challenging properties to be predicted by the MLPs. Further, I identified pairs of properties that are challenging to predict.
This series of works demonstrated that atomistic modeling is an effective computation technique for studying atomistic dynamic mechanisms, evaluating materials thermodynamics, and guiding materials discovery. As an emerging computational technique, MLPs show good performance on many materials and have great potential to enhance the materials research, but the results show that critical assessments are needed to examine their capabilities to accurately reproduce physical phenomena and understand their limits of performing reliable simulations. My thesis evaluates MLP performances on a number of critical issues related to materials simulations and provides guidance to improve MLPs for future studies.