Session B62: Emerging Trends in Molecular Dynamics Simulations and Machine Learning I

Towards Large-scale Quantum Accuracy Materials Simulations

Electronic structure calculations, especially those using density functional theory (DFT), have been very useful in understanding and predicting a wide range of materials properties. Despite the wide adoption of DFT, and the tremendous progress in theory and numerical methods over the decades, the following challenges remain. Firstly, many widely used implementations of DFT suffer from domain-size and geometry restrictions, limiting the complexity of materials systems that can be treated using DFT calculations. Secondly, there are many materials systems (such as strongly-correlated systems) where the widely used model exchange-correlation functionals in DFT, which account for the many-body quantum mechanical interactions between electrons, are not sufficiently accurate. 

 

This talk will discuss the recent advances towards addressing the aforementioned challenges. In particular, the development of computational methods and numerical algorithms for conducting fast and accurate large-scale DFT calculations using adaptive finite-element discretization will be presented, which form the basis for the recently released DFT-FE open-source code [1,2]. The computational efficiency, scalability and performance of DFT-FE will be presented, which demonstrates a significant outperformance of widely used plane-wave DFT codes. Some recent application studies that highlight the capabilities of DFT-FE will be presented. In addressing the second challenge, our recent progress in accurately solving the inverse DFT problem will be presented, which has enabled the computation of exact exchange-correlation potentials for polyatomic systems [3,4,5]. Ongoing efforts on using the exact exchange-correlation potentials to improve the exchange-correlation functional description in DFT will be discussed. 

 

[1] P. Motamarri et al., Comput. Phys. Commun. 246, 106853 (2020).

[2] S. Das et al., Comput. Phys. Commun. 280, 108473 (2022).

[3] B. Kanungo, P. Zimmerman, V. Gavini, Nature Communications 10, 4497 (2019).

[4] B. Kanungo, P. Zimmerman, V. Gavini, J. Phys. Chem. Lett. 12, 12012-12019 (2021).

[5] B. Kanungo, J. Hatch P. Zimmerman, V. Gavini, Exact and model exchange-correlation potentials for open shell systems, to appear in J. Phys. Chem. Lett. (2023); arXiv:2305.15620.

    

Quantum-accurate large-scale atomistic simulation of materials with LAMMPS and FitSNAP

Molecular dynamics (MD) is a powerful materials simulation approach whose accuracy is limited by the interatomic potential (IAP). The quest for improved accuracy has resulted in a decades-long growth in the complexity of IAPs, many of which are now implemented in the LAMMPS MD code[1]. Traditional physics-based IAPs are now being rapidly supplanted by machine-learning IAPs. The SNAP (Spectral Neighbor Analysis Potential) machine-learning approach is one example of this[2]. SNAP is formulated in terms of a set of general four-body descriptors that characterize the local neighborhood of each atom. The FitSNAP software[3], tightly integrated with LAMMPS, provides an automated methodology for generating accurate and robust application-specific IAPs. This approach has been used to develop potentials for diverse materials, including metal alloys, semiconductors, plasma-facing materials, and even magnetic materials such as iron. Each SNAP IAP is trained on quantum electronic structure calculations of energy, force, and stress for many small configurations of atoms. The resultant potentials enable high-fidelity large-scale MD simulations of these materials, yielding insight into their behavior on lengthscales and timescales unreachable by other methods. The relatively large computational cost of SNAP is offset by combining LAMMPS' spatial parallel algorithms with Kokkos-based hierarchical multithreading, enabling the efficient use of Exa-scale CPU and GPU platforms, allowing large-scale production simulations at 30 ns/day with millions to billions of atoms. Finally, I will discuss extensions of FitSNAP and LAMMPS to handle other ML styles, including neural network libraries and the more general Atomic Cluster Expansion descriptors.   

Benchmarking anharmonicity in machine learned interatomic potentials

Machine learning (ML) approaches have recently emerged as powerful tools to probe structure-property relationships in crystals and molecules. Specifically, ML interatomic potentials have been shown to predict ground state properties with near density functional theory (DFT) accuracy at a cost similar to conventional interatomic potential approaches. While ML potentials have been extensively tested across various classes of materials and molecules, there is no clear understanding of how well the anharmonicity of any given system is encoded. Here, we benchmark popular ML interatomic potentials using third and fourth order phonon interactions in fluorite crystals. An anharmonic hamiltonian was constructed from DFT using our highly accurate and efficient irreducible derivative methods, which were then used to train three classes of ML potentials: Gaussian Approximation Potentials, Behler-Parrinello Neural Networks, and Graph Neural Networks. We evaluate their accuracy in not only reproducing anharmonic interaction terms but also in observables such as phonon linewidths and lineshifts. We then present the results of the models trained on a DFT dataset, showing good and reasonable agreement with the DFT computed third and fourth order interactions, respectively. Finally, we discuss strategies to leverage anharmonic terms in the training procedure to improve the accuracy of ML interatomic potentials.   

Modeling Atomic Structure of Platinum Deposition on Graphene with Machine Learning Interatomic Potentials

Depositing catalytic metals, such as Platinum, in forms like monolayers, sub-monolayers, or nanoparticles onto two-dimensional substrates like graphene, enhances surface area and reduces material usage, resulting in significant cost savings. Computational modeling of these catalytic systems is crucial for predicitng their reactivity. However, many computational investigations, primarily based on First-Principles methods like Density Functional Theory (DFT), often rely on arbitrarily assumed crystal configurations of metal clusters on graphene for computational tractability. Given that DFT electronic and chemical properties heavily depend on the underlying atomic configurations, it is imperative to meticulously study the atomic structure of these metal depositions before making property predictions with DFT. Recent advancements in machine learning interatomic potentials have shown considerable promise in enabling large-scale simulations while achieving accuracy comparable to DFT at a fraction of the computational cost. In this study, we employ neural network potentials to conduct extensive Monte Carlo simulations to investigate Platinum growth on graphene. Our results provide a comprehensive analysis of the competition between lateral and vertical growth of Platinum under a wide range of loadings and demonstrate excellent agreement with experimental measurements conducted using transmission electron microscopy.   

Extracting Catalytic Reaction Mechanisms from Large-Scale Simulations Accelerated by Machine Learning Interatomic Potentials

Machine learning interatomic potentials (MLIPs) have become a prevalent approach to bridging the gap between slow-but-accurate ab initio calculations and fast-but-inaccurate empirical potentials for molecular dynamics. Among MLIPs, there is a pareto front of models with different tradeoffs between accuracy and speed. The FLARE interatomic potential aims to push the boundary of scalability and performance, while maintaining sufficient accuracy to study complex, reactive systems.

In our recent work [1], we demonstrated the ability of FLARE to simulate heterogeneous catalytic reactions in systems of up to 0.5 trillion atoms using 27336 GPUs, with an accurate Bayesian MLIP efficiently trained using active learning. With the exceptional speed of FLARE on modern HPC architectures, we can combine the atomistic resolution of molecular dynamics simulations with the large length and time scales required for realistic simulations and sufficient sampling of rare catalytic events, while providing the near-quantum level of accuracy provided by the MLIP. In this talk, we will showcase how this combination of scales and accuracy enables new insights into chemically complex processes such as heterogeneous catalysis through high-resolution measurements of reaction rates and residence times.

[1] arXiv:2204.12573   

The quest for spin crossover molecules for low-power nonvolatile memory applications

A spin crossover molecular complex consists of a transition metal atom core, usually with a 3d4 to 3d7 electronic configuration, that can be switched from its low-spin to high-spin state with an external perturbation. Of relevance to this talk is the voltage control of spin-state switching leading to nonvolatile conductance change, which shows promise for applications in high-density, low-power nonvolatile memory. Yet, there is little fundamental understanding of how to reliably design a spin crossover complex with low enough resistance while at the same time retaining a very high (~10⁴) ON/OFF ratio. In this talk, I will summarize recent understanding of the characteristics of spin crossover complexes, as single molecules, in organized layers, and as supported on functional low dimensional materials and organic ferroelectrics. The interactions of spin crossover complexes with substrates will be explored with the goal of unraveling the effect of ferroelectric polarization on the molecular spin states. Through examples of a handful of “good SCO complexes” such as the Fe(II) complex Fe[H₂B(Pz)₂]₂(bipy), some possible descriptors of “goodness” will be presented that may accelerate the discovery of novel spin crossover complexes. Could we design redox-active ligands for substantial charge delocalization and enhanced intermolecular charge transfer in the solid state that may promote high conductivity in the ON state of the device? Could AI help us discover new “good SCO complexes” that would make molecular electronics with spin crossover complexes feasible?  Answers to these and related questions will be discussed.   

Machine-Learning Potentials for All-Atom Simulation of CO2’s Chemisorption

The application of solid sorbent materials for carbon capture has been proposed as an alternative to amine-based liquid sorbents due to their lower desorption energy requirement. Among various types of porous solid sorbent, metal-organic frameworks (MOFs) are highly promising because they typically exhibit both high CO₂ uptake and CO₂/N₂ selectivity, which are required for carbon capture. The chemisorption of CO₂ in MOFs (such as on open metal sites) generally yields an extremely high CO₂/N₂ selectivity, manifesting itself more prominently at lower partial pressures (which is particularly relevant for direct air capture). The atomistic modelling of this process (e.g. bond forming) cannot be performed using efficient classical force fields and requires the first-principle based simulation. The latter is still computationally costly and is not suitable for screening a large amount of MOFs. Here, we report the quantum-informed machine-learning potentials for atomistic simulations, including both molecular dynamics (MD) and grand canonical monte carlo (GCMC), of CO₂ in MOFs. We demonstrate that the method has a much higher computational efficiency than the first-principle one while predicting accurate forces on atoms (in MD simulations) and energies (in GCMC simulations).  We further explored the transferability of machine-learning potentials among MOFs with similar atomic structures.   

Machine Learning-Based Predictions of Threshold Displacement Energy in Materials

The threshold displacement energy (E_d) of materials is a crucial parameter for most of the radiation damage models that predict the performance of materials in extreme radiation environments. Traditionally, E_d calculations have been based on complex experiments or computationally expensive methods, such as Molecular Dynamics or ab initio Molecular Dynamics. However, recent advancements in machine learning have opened new avenues for efficiently estimating E_d values for different materials. This work explores the application of machine learning techniques to predict E_d with accuracy and reduced computational cost compared to traditional methods. We discuss the methodology, data sources, and model performance, highlighting the potential of machine learning to revolutionize our understanding of radiation damage in materials.   

Investigating MXene properties via simulated surface acoustic waves

MXenes are a class of 2D materials, consisting of transition metal carbides and nitrides. They have tremendous technological promise with applications ranging from energy storage, sensors, photocatalysis, and even neural electrodes. They can be single or multi-layer, terminated with a variety of functional groups, and free standing or supported on almost any substrate. In this presentation, we study the mechanical properties of single and multi-layer Ti3C2 MXene sheets via atomistic simulations. Specifically, we are interested in surface acoustic waves (SAWs), mechanical waves focused at the surface of a material which are coupled strongly to the electronic structure. We will discuss SAWs characteristics and propagation mechanisms in this 2D material as a function of the wave amplitude, sheet characteristics, and termination to infer their unique mechanical properties.   

Merging Quantum Mechanics with Machine Learning for Navigating Chemical Space

 

Innovative materials are needed to tackle current major challenges in energy storage and generation. However, the design of new materials largely relies on experimental trial and error, limiting the number explored compounds relative to the entire space of possible compounds. 

In this presentation, I will discuss our approach to materials design, which integrates machine learning (ML) techniques with quantum mechanics-based computations. Our ML models are trained to predict quantum mechanical properties, like formation energy and band gap, using only the material's structure as input. I'll highlight a few applications in computational materials design where ML offers significant advantages, such as circumventing costly electronic structure calculations to provide statistical approximations of high-fidelity band structures.[1]This allows for an improved computational understanding of a material's electronic properties.

         One limitation of using structure-based ML models in material screening is due to the prerequisite of having a well-defined structure for optimal accuracy. To address these challenges, we have developed a generalizable ML-based interatomic potential (MLIPs) for structure prediction of complicated organic/inorganic materials.[2] Additionally, these MLIPs are highly computationally efficient which allows for modeling the temperature-dependent dynamics of local structural domains. Such insights can be valuable when assessing a material's suitability for energy-based applications. Overall, the application of ML and computation shows great promise in facilitating the exploration of the chemical space to allow for transformative advancements in energy storage and generation.