Session N62: Emerging Trends in Molecular Dynamics Simulations and Machine Learning III

Exascale Electronic Structure and Quantum Transport Calculations

The creation of robust, adaptive software and algorithms that can fully exploit exascale capabilities and future computing architectures is critical to designing advanced materials and devices with targeted properties. We have developed an open-source code that discretizes the DFT equations on real-space grids distributed over the nodes of a massively parallel system via domain decomposition. Multigrid techniques are used to greatly accelerate convergence while only requiring nearest neighbor communications, while a novel adaptive finite differencing scheme dramatically improves accuracy. The real-space multigrid (RMG) code achieves the same average accuracy as plane wave codes on the well-known Delta test for 71 elements in the periodic table and can handle all Bravais lattice types. It scales from desktops and clusters to supercomputers, including the exascale Frontier and pre-exascale Summit, Perlmutter and Polaris systems, while utilizing all CPU cores and GPUs in each node. RMG is distributed via www.rmgdft.org, with over 4,000 downloads to date. Due to its computational efficiency, RMG is very suitable for large-scale survey approaches, including Materials-Genome and Machine-Learning projects. Advanced functionalities are provided through interfaces to other codes, including QMCPACK, BerkeleyGW, Phonopy, and ALAMODE. A localized-orbitals RMG module enables high-accuracy calculations with much-reduced memory footprint and scaling, saving 90% of wave function memory in 1,000-atom supercells. It forms the basis for a non-equilibrium Green’s function module able to study quantum transport properties for devices containing tens of thousands of atoms with full DFT accuracy. Several large-scale applications will be discussed.

*In collaboration with E. L. Briggs and W. Lu.   

Controlling Ferroelectric Domains in Moire MoS2

Moire supercells, formed by the stacking of two-dimensional (2D) van der Waals materials with small twists, have been shown to exhibit novel electronic and optical properties. The stacking order of these 2D materials plays a crucial role in determining their electronic and optical characteristics. These supercells have also been observed to give rise to a superlattice of out-of-plane ferroelectric domains. In this study, we employ molecular dynamics simulations to investigate the formation of polarized domain patterns in a bi-layer MoS₂ at finite temperatures. Specifically, we examine how the initial twist angle of the stacked 2D materials affects the formation of polarized domains and how these domains evolve with increasing temperature. Additionally, we explore how the initial twist angle can be utilized to control the size of the ferroelectric domains. Our findings could provide significant insights into the design and development of novel electronic and optical devices based on 2D materials.   

Hydrogen Diffusion through Polymer Using Deep Reinforcement Learning

Hydrogen energy has the potential to reshape the energy landscape by replacing significant amounts of fossil fuels in many fields. Versatile storage methods have been developed to store hydrogen safely and efficiently, including physical and chemical storage. To pursue a sustainable and low-carbon future, containers with polymer liners have emerged as a low-cost hydrogen storage solution due to their chemical inertness and low permeability. Here, the understanding of long-time diffusion mechanism is critical for a rational design of next-generation hydrogen storage. In response, we have developed a computational framework with deep reinforcement learning (DRL) combined with transition state theory to investigate molecular diffusion at experimentally relevant time scale. Based on the Deep Q-network architecture and distributed training framework, DRL agent is capable of learning energy-efficient pathways in a variety of polymer morphologies. In this talk, I will discuss atomistic mechanisms of long-time molecular diffusion for polymer liner applications.   

Neural Network Quantum Molecular Dynamics Simulation of Topological Defects in YMnO3 Manganite

Perovskite multi-ferroic manganite materials possess a unique Mexican hat-shaped potential energy surface, which makes them an ideal candidate for investigation of topological-defect formation under rapid quenching. Quantum-mechanical methods based on density functional theory (DFT) can accurately describe atomic interactions but fail, due to limitations in system size, in simulating large-scale topological defect formation. Recent developments in machine learning (ML) have made it possible to investigate the kinetics of complex phase transitions. In this work, we train an equivariant Allegro-Legato neural network quantum molecular dynamics (NNQMD) model using DFT training data for YMnO₃, and thereby investigate dependence of topological defect formation on the quenching rate. We have also performed diffuse neutron scattering experiments on the quenched YMnO₃ to quantify topological defects for validating our NNQMD simulations.   

Indirect Learning of an Interatomic Potential to Model the Phase-Change Material Ge2Sb2Te5

Chalcogenide phase change materials such as Ge₂Sb₂Te₅ (GST) have received attention due to their ability to transition rapidly between stable solid states, giving them binary data retention properties ideal for computer memory. Thus, there is wide interest in studying this material via atomistic modeling. Recently, a Gaussian Approximation Potential (GAP) was trained to reproduce Density Functional Theory (DFT) energies and forces at a fraction of the computational cost; however, GAP simulations are still far too slow to study the phase change properties that make GST so valuable. Here we present a machine-learned (ML) potential that performs three orders of magnitude faster than GAP. Rather than training the ML potential directly from DFT, we use the concept of indirect learning and train from configurations generated using the GAP potential, allowing us to generate a far more extensive training set than could be considered using DFT. The resulting potential - trained using the Atomic Cluster Expansion (ACE) model - reproduces the structure and thermodynamics of the GAP potential. The tremendous improvement in speed of the ACE potential allows us to study in depth the phase change behavior of GST, as well as explore the possible fragile-to-strong crossover in amorphous phases.   

Materials Discovery Using Simulations and Deep Learning

Despite the recent advances in physical simulations and machine learning, the exploration of novel inorganic crystals remains constrained by the expensive trial-and-error approaches. Recent developments in deep learning have shown that models can attain emergent predictive capabilities with increasing data and computation, in fields such as language, vision, and biology. In this talk, we will present our recent results on how graph networks trained at scale can reach unprecedented levels of generalization, improving the efficiency of materials discovery by an order of magnitude. We will show how the scale and diversity unlock surprising modeling capabilities for downstream applications, including predictions of crystalline stability, ionic conduction, and structural properties of amorphous materials.   

Investigating Water Adsorption in Amine-Appended Metal-Organic Frameworks using Density Functional Theory-Derived Neural Network Potentials

CO₂ adsorption in polyamine-appended metal-organic frameworks (MOFs) exhibits a cooperative step-shaped isotherm [1], allowing for complete adsorption/desorption within small pressure/temperature swings. As a consequence, these materials are promising candidates for reversible carbon capture, particularly in humid conditions as the amine-appended group protects open metal sites from water poisoning. Nonetheless, the role of water molecules on the adsorption properties and mechanism in such systems is not clearly understood. In this context, ab initio neural network potentials (NNPs), derived from density functional theory calculations, are a promising tool to simulate atomistic dynamics for such large and complex systems at finite temperature. In this work, we investigate water effects on CO₂ adsorption in the MOF framework Mg₂(olz) (olz4- = (E)-5,5′-(diazene-1,2-diyl) bis(2-oxidobenzoate)) with a sample of appended diamine groups. Exploiting the computational efficiency and near-first-principles accuracy of our NNPs, we are able to identify water binding sites, discover lower energy structures, and ultimately predict water and CO₂ diffusion coefficients in these systems. 

[1]  McDonald, Thomas M., et al, J. Am. Chem. Soc. 2012, 134, 16, 7056–7065   

Hole Doping of Hydrogenated Diamond by MoO3: Effect of Oxygen Vacancy

Surface transfer doping has been explored to be a potential strategy for doping diamond, a material which is known hard to dope traditionally, for applications in high power electronics. Although MoO₃ has been identified as effective surface electron acceptor for hydrogen-terminated diamond, the impact of oxygen vacancies, which are commonly present, still remain elusive. To address this, we conducted reactive molecular dynamics simulations to investigate the deposition of MoO_3-x on hydrogenated diamond (111) surface. As more vacancies are introduced into the material, the local atomic arrangement is observed to become more compact and the Mo-O bonds are found to get stronger. Further we employed first-principles calculations based on density functional theory to investigate the charge transfer and electronic structures. Bader charge calculations further confirm MoO_3-x as effective surface electron acceptors. The change in Bader charge and notable shift of the electronic band alignment upon doping reveal that oxygen vacancy limits the hole-doping capability of MoO_3-x. Spatial distribution of doped holes is characterized after deposition, which demonstrates a widespread distribution of hole density, so-called two-dimensional hole gas, at the interface that can support exceptional transport properties.    

Uncertainty Estimation and Robust Training of Materials Graph Neural Network Models

Graph Neural Network (GNN) models have become a promising new tool for materials simulations. One key advancement of GNN models is the invention of universal interatomic potentials (UIPs) that are transferable to the whole periodic table. Such GNN UPs have trainable parameters at the scale of millions, and they achieve reasonable training and test accuracies for millions of training and test structures spanning the space of known chemistry. Such UIPs open the opportunity to explore and discovery hypothetical materials at the unprecedented scales of tens to hundards of millions. However, even UIPs suffer from poor reliability for extrapolation, and their prediction accuracy for unseen structures are worse than that for training structures. The usefulness of UIPs would be questionable if their reliability cannot be confirmed. In our work, we tackle this challenge by introducing a method to estimate the prediction uncertainty of GNN models for materials simulations, with which we also come up with a training set specially aimed for ambient-temperatures simulations of all existing structures in the Materials Project.   

Oral: Denoising of Transmission Electron Microscopic Data

In-situ TEM is a powerful technique to understand dynamic changes in nanoparticles as they undergo phase transformation and structural deformation. As these transformations can occur spontaneously, conducting high-speed in-situ TEM studies to achieve atomic resolution information across a wide field of view while minimizing the impact of electron dose presents a significant challenge. The in-situ TEM video often contains high levels of noise, making it difficult to obtain structural and atomic information. 

 

We have applied denoising methods to a TEM dataset affected by a combination of Poisson and Gaussian noise. A distinctive challenge in this study lies in the absence of ground truth or previously denoised images for benchmarking purposes. We use a variety of denoising techniques including Total Variation, Non-Local Means, Noise2Void, Noise2Fast, and BM3D with Anscombe transform. The denoising approaches reveal new features hidden in the noisy data. The denoised TEM images show a stronger contrast, making it easier for us to distinguish the boundary between the nanoparticle and the background (carbon support). Comparing the FFTs of the noisy and denoised images, we reveal an additional spot corresponding to the  Fe₃O₄ (400) reflection.  The FFT of the denoised image shows an increase in the clarity and number of  “Thon” rings which suggests an increase in the information content of the image as a function of spatial frequency, or resolution.   

Uncertainty-aware machine learning surrogates of molecular dynamics simulations

We introduce an approach to use the statistical uncertainties associated with outputs of molecular dynamics simulations of soft materials to improve the training of deep neural networks when creating machine learning surrogates for predicting the relationship between input parameters and simulation outputs. The approach is illustrated by designing a surrogate model for molecular dynamics simulations of confined electrolytes, which aims to predict the intricate relationship between the features of the electrolyte solution (e.g., ion size, electrolyte concentration) and the resulting ionic structure. We propose a surrogate model to consider the direct regression and replace the point estimation with a probabilistic distribution in the output space. By incorporating probabilistic embeddings with Kullback–Leibler divergence in the loss function, we show that the model can significantly reduce prediction errors for samples in the unseen test dataset as well as yield higher generalizability and robustness among different datasets.