Unified graph neural network force-field for the periodic table: solid state applications

Choudhary, Kamal; DeCost, Brian; Major, Lily; Butler, Keith T.; Thiyagalingam, Jeyarajan; Tavazza, Francesca

doi:10.1039/d2dd00096b

Cited by 42 publications

(29 citation statements)

References 73 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…However, other studies have found that the difference in scale between inter- and intramolecular interactions means that capturing both with a single model is often limiting and that training separate intermolecular and intramolecular models yields improved performance. , Partitioning the energy will also make applying the workflow over multiple landscapes more practical, which could allow for training transferable rather than system-specific models. The development of universal models for organic molecules ,− and inorganic materials, − has produced impressive results with good transferability; similar models for organic crystals could have an important impact in the field of CSP.…”

Section: Discussionmentioning

confidence: 99%

Machine-Learned Potentials by Active Learning from Organic Crystal Structure Prediction Landscapes

Butler,

Hafizi,

Day

2024

J. Phys. Chem. A

View full text Add to dashboard Cite

A primary challenge in organic molecular crystal structure prediction (CSP) is accurately ranking the energies of potential structures. While high-level solid-state density functional theory (DFT) methods allow for mostly reliable discrimination of the low-energy structures, their high computational cost is problematic because of the need to evaluate tens to hundreds of thousands of trial crystal structures to fully explore typical crystal energy landscapes. Consequently, lower-cost but less accurate empirical force fields are often used, sometimes as the first stage of a hierarchical scheme involving multiple stages of increasingly accurate energy calculations. Machine-learned interatomic potentials (MLIPs), trained to reproduce the results of ab initio methods with computational costs close to those of force fields, can improve the efficiency of the CSP by reducing or eliminating the need for costly DFT calculations. Here, we investigate active learning methods for training MLIPs with CSP datasets. The combination of active learning with the well-developed sampling methods from CSP yields potentials in a highly automated workflow that are relevant over a wide range of the crystal packing space. To demonstrate these potentials, we illustrate efficiently reranking large, diverse crystal structure landscapes to near-DFT accuracy from force field-based CSP, improving the reliability of the final energy ranking. Furthermore, we demonstrate how these potentials can be extended to more accurately model structures far from lattice energy minima through additional on-the-fly training within Monte Carlo simulations.

show abstract

Section: Discussionmentioning

confidence: 99%

Machine-Learned Potentials by Active Learning from Organic Crystal Structure Prediction Landscapes

Butler,

Hafizi,

Day

2024

J. Phys. Chem. A

View full text Add to dashboard Cite

show abstract

“…Models generated in this way are typically expected to be directly applicable to downstream tasks in which the explored configurations are effectively covered by the training data. Some examples include models such as M3GNet [15], CHGNet [16] and MACE-MP-0 [17], which are all trained on snapshots from DFT relaxations of the Material Project [18] structures, with M3GNet utilizing 88k configurations across 89 chemical species and both CHGNet and MACE-MP-0 being trained on 1.58M inorganic crystal frames from the concurrently introduced MPtraj dataset [16]; GNoME [19], trained on a dataset of inorganic crystals also starting from MP, but nearly two orders of magnitude larger than MPtrj; PreFerred Potential (PFP), trained on approximately 9M frames of 45 elements [20]; and ALIGNN, trained on 307K data frames of 89 elements [21]. However, several limitations exist: (1) Simultaneously training multiple datasets from different application fields is not feasible due to the variations in labeling with different DFT settings.…”

Section: Mainmentioning

confidence: 99%

DPA-2: Towards a universal large atomic model for molecular and materials simulation

Wang,

Zhang,

Liu

et al. 2024

Preprint

View full text Add to dashboard Cite

The rapid development of artificial intelligence (AI) is driving significant changes in the field of atomic modeling, simulation, and design. AI-based potential energy models have been successfully used to perform large-scale and long-time simulations with the accuracy of ab initio electronic structure methods. However, the model generation process still hinders applications at scale. We envision that the next stage would be a model-centric ecosystem, in which a large atomic model (LAM), pre-trained with as many atomic datasets as possible and can be efficiently fine-tuned and distilled to downstream tasks, would serve the new infrastructure of the field of molecular modeling. We show that DPA-2 can accurately represent a diverse range of chemical systems and materials, enabling high-quality simulations and predictions with significantly reduced efforts compared to traditional methods. Our approach paves the way for a universal large atomic model that can be widely applied in molecular and material simulation research, opening new opportunities for scientific discoveries and industrial applications.

show abstract

“…Machine learning potentials (MLPs) have recently emerged as highly promising tools in computational materials science due to their near-DFT accuracy, nearly linear scaling with system size, and exceptional transferability to diverse chemical environments. Prominent examples of MLPs include neural network potentials (NNPs), − Gaussian approximation potentials, , moment tensor potentials, spectral neighbor analysis potentials, atomic CE potentials, , and graph NNPs. − The high flexibility of MLPs allows for broad applicability across different types of matter, encompassing bulk − and 2D crystals, amorphous materials, , liquids, − interfaces, , and clusters. , In the domain of disordered systems, MLPs were employed to investigate binary alloys spanning a wide range of compositions, − high-entropy alloys, − and grain boundaries. − However, a conspicuous gap exists in the literature concerning the application of MLPs to nonstoichiometric systems characterized by varying elevated vacancy concentrations. Herein, we address this gap by examining the efficacy of NNPs for modeling nonstoichiometric chromium sulfides, a material that has not been explored in the existing literature using either machine learning or conventional potentials.…”

Section: Introductionmentioning

confidence: 99%

Modeling Chemical Exfoliation of Non-van der Waals Chromium Sulfides by Machine Learning Interatomic Potentials and Monte Carlo Simulations

Ibrahim,

Wines,

Ataca

2024

J. Phys. Chem. C

View full text Add to dashboard Cite

The chemical exfoliation of non-van der Waals (vdW) materials to ultrathin nanosheets remains a formidable challenge. This difficulty arises from the strong preference of these materials to engage in three-dimensional chemical bonding, resulting in uncontrolled atomic migration into the vdW gaps during cation deintercalation from the bulk structure, ultimately leading to unpredictable structural disorder. Computational models capable of comprehending the widespread nonstoichiometric local environments resulting from disordered atomic migrations during the exfoliation of non-vdW materials are crucial for understanding the underlying mechanisms. Here, we propose a generic framework using neural network potentials (NNPs) to accurately model nonstoichiometric systems over a broad range of vacancy concentrations. We apply our framework to investigate the crystal structures and phase transformations occurring during the exfoliation of non-vdW nonstoichiometric Cr (1−x) S systems, a compelling material category with substantial potential for twodimensional (2D) magnetic applications. The efficacy of the NNP outperforms conventional cluster expansion, exhibiting superior accuracy and transferability to unexplored crystal structures and compositions. By employing the NNP in simulated annealing optimizations, we predict low-energy Cr (1−x) S structures anticipated to result from experimental synthesis. A notable structural transition is discerned at the Cr 0.5 S composition, with half of the Cr atoms preferentially migrating to vdW gaps. This aligns with experimental observations in the chemical exfoliation of 2D CrS 2 and emphasizes the vital role of excess Cr atoms beyond the Cr/S = 1/2 composition ratio in stabilizing vdW gaps. Additionally, we utilize the NNP in a large-scale vacancy diffusion Monte Carlo simulation to illustrate the impact of lateral compressive strains in catalyzing the formation of vdW gaps within non-vdW CrS 2 slabs through Poisson's axial expansion. This provides a direct pathway for more facile exfoliation of ultrathin nanosheets from non-vdW materials through strain engineering. The implemented methodology, leveraging machine learning potentials, is imperative to bridge the quantum-level accuracy to large scales necessary for modeling the intricate mechanisms underlying the chemical exfoliation of non-vdW materials.

show abstract

Unified graph neural network force-field for the periodic table: solid state applications

Abstract: Classical force fields (FF) based on machine learning (ML) methods show great potential for large scale simulations of solids. MLFFs have hitherto largely been designed and fitted for specific systems...

Cited by 42 publications

References 73 publications

Machine-Learned Potentials by Active Learning from Organic Crystal Structure Prediction Landscapes

Machine-Learned Potentials by Active Learning from Organic Crystal Structure Prediction Landscapes

DPA-2: Towards a universal large atomic model for molecular and materials simulation

Modeling Chemical Exfoliation of Non-van der Waals Chromium Sulfides by Machine Learning Interatomic Potentials and Monte Carlo Simulations

Contact Info

Product

Resources

About