arXiv subjects

Gábor Csányi

Publications and source records attributed to Gábor Csányi.

At least 55 records · Page 3Linked to original sources

arXiv2023

Gaussian Approximation Potentials: theory, software implementation and application examples

Gaussian Approximation Potentials are a class of Machine Learned Interatomic Potentials routinely used to model materials and molecular systems on the atomic scale. The software implementation provides the means for both fitting models using ab initio data and using the resulting potentials in atomic simulations. Details of the GAP theory, algorithms and software are presented, together with detailed usage examples to help new and existing users. We review some recent developments to the GAP framework, including MPI parallelisation of the fitting code enabling its use on thousands of CPU cores and compression of descriptors to eliminate the poor scaling with the number of different chemical elements.

cond-mat.mtrl-sci↗

arXiv2023

ACEpotentials.jl: A Julia Implementation of the Atomic Cluster Expansion

We introduce ACEpotentials.jl, a Julia-language software package that constructs interatomic potentials from quantum mechanical reference data using the Atomic Cluster Expansion (Drautz, 2019). As the latter provides a complete description of atomic environments, including invariance to overall translation and rotation as well as permutation of like atoms, the resulting potentials are systematically improvable and data efficient. Furthermore, the descriptor's expressiveness enables use of a linear model, facilitating rapid evaluation and straightforward application of Bayesian techniques for active learning. We summarize the capabilities of ACEpotentials.jl and demonstrate its strengths (simplicity, interpretability, robustness, performance) on a selection of prototypical atomistic modelling workflows.

physics.comp-ph↗

arXiv2023

Machine learning of microscopic structure-dynamics relationships in complex molecular systems

In many complex molecular systems, the macroscopic ensemble's properties are controlled by microscopic dynamic events (or fluctuations) that are often difficult to detect via pattern-recognition approaches. Discovering the relationships between local structural environments and the dynamical events originating from them would allow unveiling microscopic level structure-dynamics relationships fundamental to understand the macroscopic behavior of complex systems. Here we show that, by coupling advanced structural (e.g., Smooth Overlap of Atomic Positions, SOAP) with local dynamical descriptors (e.g., Local Environment and Neighbor Shuffling, LENS) in a unique dataset, it is possible to improve both individual SOAP- and LENS-based analyses, obtaining a more complete characterization of the system under study. As representative examples, we use various molecular systems with diverse internal structural dynamics. On the one hand, we demonstrate how the combination of structural and dynamical descriptors facilitates decoupling relevant dynamical fluctuations from noise, overcoming the intrinsic limits of the individual analyses. Furthermore, machine learning approaches also allow extracting from such combined structural/dynamical dataset useful microscopic-level relationships, relating key local dynamical events (e.g., LENS fluctuations) occurring in the systems to the local structural (SOAP) environments they originate from. Given its abstract nature, we believe that such an approach will be useful in revealing hidden microscopic structure-dynamics relationships fundamental to rationalize the behavior of a variety of complex systems, not necessarily limited to the atomistic and molecular scales.

physics.chem-ph↗

arXiv2023

Accurate Energy Barriers for Catalytic Reaction Pathways: An Automatic Training Protocol for Machine Learning Force Fields

In this study, we introduce a training protocol for developing machine learning force fields (MLFFs), capable of accurately determining energy barriers in catalytic reaction pathways. The protocol is validated on the extensively explored hydrogenation of carbon dioxide to methanol over indium oxide. With the help of active learning, the final force field obtains energy barriers within 0.05 eV of Density Functional Theory. Thanks to the computational speedup, not only do we reduce the cost of routine in-silico catalytic tasks, but also find a 40\% reduction in the previously established rate-limiting step. Furthermore, we illustrate the importance of finite-temperature effects and compute free energy barriers. The transferability of the protocol is demonstrated on the experimentally relevant, yet unexplored, top-layer reduced indium oxide surface. The ability of MLFFs to enhance our understanding of extensively studied catalysts underscores the need for fast and accurate alternatives to direct ab-intio simulations.

physics.chem-ph↗

arXiv2023

wfl Python Toolkit for Creating Machine Learning Interatomic Potentials and Related Atomistic Simulation Workflows

Predictive atomistic simulations are increasingly employed for data intensive high throughput studies that take advantage of constantly growing computational resources. To handle the sheer number of individual calculations that are needed in such studies, workflow management packages for atomistic simulations have been developed for a rapidly growing user base. These packages are predominantly designed to handle computationally heavy ab initio calculations, usually with a focus on data provenance and reproducibility. However, in related simulation communities, e.g. the developers of machine learning interatomic potentials (MLIPs), the computational requirements are somewhat different: the types, sizes, and numbers of computational tasks are more diverse, and therefore require additional ways of parallelization and local or remote execution for optimal efficiency. In this work, we present the atomistic simulation and MLIP fitting workflow management package wfl and Python remote execution package ExPyRe to meet these requirements. With wfl and ExPyRe, versatile Atomic Simulation Environment based workflows that perform diverse procedures can be written. This capability is based on a low-level developer-oriented framework, which can be utilized to construct high level functionality for user-friendly programs. Such high level capabilities to automate machine learning interatomic potential fitting procedures are already incorporated in wfl, which we use to showcase its capabilities in this work. We believe that wfl fills an important niche in several growing simulation communities and will aid the development of efficient custom computational tasks.

physics.comp-ph↗

arXiv2023

Structural Dynamics Descriptors for Metal Halide Perovskites

Metal halide perovskites have shown extraordinary performance in solar energy conversion technologies. They have been classified as "soft semiconductors" due to their flexible corner-sharing octahedral networks and polymorphous nature. Understanding the local and average structures continues to be challenging for both modelling and experiments. Here, we report the quantitative analysis of structural dynamics in time and space from molecular dynamics simulations of perovskite crystals. The compact descriptors provided cover a wide variety of structural properties, including octahedral tilting and distortion, local lattice parameters, molecular orientations, as well as their spatial correlation. To validate our methods, we have trained a machine learning force field (MLFF) for methylammonium lead bromide (CH$_3$NH$_3$PbBr$_3$) using an on-the-fly training approach with Gaussian process regression. The known stable phases are reproduced and we find an additional symmetry-breaking effect in the cubic and tetragonal phases close to the phase transition temperature. To test the implementation for large trajectories, we also apply it to 69,120 atom simulations for CsPbI$_3$ based on an MLFF developed using the atomic cluster expansion formalism. The structural dynamics descriptors and Python toolkit are general to perovskites and readily transferable to more complex compositions.

cond-mat.mtrl-sci↗

arXiv2023

Dynamic Local Structure in Caesium Lead Iodide: Spatial Correlation and Transient Domains

Metal halide perovskites are multifunctional semiconductors with tunable structures and properties. They are highly dynamic crystals with complex octahedral tilting patterns and strongly anharmonic atomic behaviour. In the higher temperature, higher symmetry phases of these materials, several complex structural features have been observed. The local structure can differ greatly from the average structure and there is evidence that dynamic two-dimensional structures of correlated octahedral motion form. An understanding of the underlying complex atomistic dynamics is, however, still lacking. In this work, the local structure of the inorganic perovskite CsPbI$_3$ is investigated using a new machine learning force field based on the atomic cluster expansion framework. Through analysis of the temporal and spatial correlation observed during large-scale simulations, we reveal that the low frequency motion of octahedral tilts implies a double-well effective potential landscape, even well into the cubic phase. Moreover, dynamic local regions of lower symmetry are present within both higher symmetry phases. These regions are planar and we report the length and timescales of the motion. Finally, we investigate and visualise the spatial arrangement of these features and their interactions, providing a comprehensive picture of local structure in the higher symmetry phases.

cond-mat.mtrl-sci↗

arXiv2023

MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields

Creating fast and accurate force fields is a long-standing challenge in computational chemistry and materials science. Recently, several equivariant message passing neural networks (MPNNs) have been shown to outperform models built using other approaches in terms of accuracy. However, most MPNNs suffer from high computational cost and poor scalability. We propose that these limitations arise because MPNNs only pass two-body messages leading to a direct relationship between the number of layers and the expressivity of the network. In this work, we introduce MACE, a new equivariant MPNN model that uses higher body order messages. In particular, we show that using four-body messages reduces the required number of message passing iterations to just two, resulting in a fast and highly parallelizable model, reaching or exceeding state-of-the-art accuracy on the rMD17, 3BPA, and AcAc benchmark tasks. We also demonstrate that using higher order messages leads to an improved steepness of the learning curves.

stat.ML↗

arXiv2022

The Design Space of E(3)-Equivariant Atom-Centered Interatomic Potentials

The rapid progress of machine learning interatomic potentials over the past couple of years produced a number of new architectures. Particularly notable among these are the Atomic Cluster Expansion (ACE), which unified many of the earlier ideas around atom density-based descriptors, and Neural Equivariant Interatomic Potentials (NequIP), a message passing neural network with equivariant features that showed state of the art accuracy. In this work, we construct a mathematical framework that unifies these models: ACE is generalised so that it can be recast as one layer of a multi-layer architecture. From another point of view, the linearised version of NequIP is understood as a particular sparsification of a much larger polynomial model. Our framework also provides a practical tool for systematically probing different choices in the unified design space. We demonstrate this by an ablation study of NequIP via a set of experiments looking at in- and out-of-domain accuracy and smooth extrapolation very far from the training data, and shed some light on which design choices are critical for achieving high accuracy. Finally, we present BOTNet (Body-Ordered-Tensor-Network), a much-simplified version of NequIP, which has an interpretable architecture and maintains accuracy on benchmark datasets.

stat.ML↗

arXiv2022

Hyperactive Learning (HAL) for Data-Driven Interatomic Potentials

Data-driven interatomic potentials have emerged as a powerful class of surrogate models for {\it ab initio} potential energy surfaces that are able to reliably predict macroscopic properties with experimental accuracy. In generating accurate and transferable potentials the most time-consuming and arguably most important task is generating the training set, which still requires significant expert user input. To accelerate this process, this work presents \text{\it hyperactive learning} (HAL), a framework for formulating an accelerated sampling algorithm specifically for the task of training database generation. The key idea is to start from a physically motivated sampler (e.g., molecular dynamics) and add a biasing term that drives the system towards high uncertainty and thus to unseen training configurations. Building on this framework, general protocols for building training databases for alloys and polymers leveraging the HAL framework will be presented. For alloys, ACE potentials for AlSi10 are created by fitting to a minimal HAL-generated database containing 88 configurations (32 atoms each) with fast evaluation times of <100 microsecond/atom/cpu-core. These potentials are demonstrated to predict the melting temperature with excellent accuracy. For polymers, a HAL database is built using ACE, able to determine the density of a long polyethylene glycol (PEG) polymer formed of 200 monomer units with experimental accuracy by only fitting to small isolated PEG polymers with sizes ranging from 2 to 32.

physics.comp-ph↗

arXiv2022

Atomistic fracture in bcc iron revealed by active learning of Gaussian approximation potential

The prediction of atomistic fracture mechanisms in body-centred cubic (bcc) iron is essential for understanding its semi-brittle nature. Existing atomistic simulations of the crack-tip deformation mechanisms under mode-I loading based on classical interatomic potentials yield contradicting predictions. To enable fracture prediction with quantum accuracy, we develop a Gaussian approximation potential (GAP) using an active learning strategy by extending a density functional theory (DFT) database of ferromagnetic bcc iron. We apply the active learning algorithm and obtain a Fe GAP model with a maximum predicted error of 8 meV/atom over a broad range of stress intensity factors (SIFs) and for four crack systems. The learning efficiency of the approach is analysed, and the predicted critical SIFs are compared with Griffith and Rice theories. The simulations reveal that cleavage along the original crack plane is the crack tip mechanism for {100} and {110} crack planes at T=0K, thus settling a long-standing dispute. Our work also highlights the need for a multiscale approach to predicting fracture and intrinsic ductility, whereby finite temperature, finite loading rate effects and pre-existing defects (e.g. nanovoids, dislocations) should be taken explicitly into account.

cond-mat.mtrl-sci↗

arXiv2022

Nested sampling for physical scientists

We review Skilling's nested sampling (NS) algorithm for Bayesian inference and more broadly multi-dimensional integration. After recapitulating the principles of NS, we survey developments in implementing efficient NS algorithms in practice in high-dimensions, including methods for sampling from the so-called constrained prior. We outline the ways in which NS may be applied and describe the application of NS in three scientific fields in which the algorithm has proved to be useful: cosmology, gravitational-wave astronomy, and materials science. We close by making recommendations for best practice when using NS and by summarizing potential limitations and optimizations of NS.

stat.CO↗

arXiv2022

Multilayer atomic cluster expansion for semi-local interactions

Traditionally, interatomic potentials assume local bond formation supplemented by long-range electrostatic interactions when necessary. This ignores intermediate range multi-atom interactions that arise from the relaxation of the electronic structure. Here, we present the multilayer atomic cluster expansion (ml-ACE) that includes collective, semi-local multi-atom interactions naturally within its remit. We demonstrate that ml-ACE significantly improves fit accuracy compared to a local expansion on selected examples and provide physical intuition to understand this improvement.

cond-mat.mtrl-sci↗

arXiv2022

Comment on "Manifolds of quasi-constant SOAP and ACSF fingerprints and the resulting failure to machine learn four body interactions"

The "quasi-constant" SOAP and ACSF fingerprint manifolds recently discovered by Parsaeifard and Goedecker are a direct consequence of the presence of degenerate pairs of configurations, a known shortcoming of all low-body-order atom-density correlation representations of molecular structures. Contrary to the configurations that are rigorously singular -- that we demonstrate can only occur in finite, discrete sets -- the continuous "quasi-constant" manifolds exhibit low, but non-zero, sensitivity to atomic displacements. Thus, it is possible to build interpolative machine-learning models of high-order interactions along the manifold, even though the numerical instabilities associated with proximity to the exact singularities affect the accuracy and transferability of such models, to an extent that depends on numerical details of the implementation.

physics.chem-ph↗

arXiv2022

A Gaussian Approximation Potential for Amorphous Si:H

Hydrogenation of amorphous silicon (a-Si:H) is critical for reducing defect densities, passivating mid-gap states and surfaces, and improving photoconductivity in silicon-based electro-optical devices. Modelling the atomic scale structure of this material is critical to understanding these processes, which in turn is needed to describe c-Si/a-Si:H heterjunctions that are at the heart of the modern solar cells with world record efficiency. Density functional theory (DFT) studies achieve the required high accuracy but are limited to moderate system sizes a hundred atoms or so by their high computational cost. Simulations of amorphous materials in particular have been hindered by this high cost because large structural models are required to capture the medium range order that is characteristic of such materials. Empirical potential models are much faster, but their accuracy is not sufficient to correctly describe the frustrated local structure. Data driven, "machine learned" interatomic potentials have broken this impasse, and have been highly successful in describing a variety of amorphous materials in their elemental phase. Here we extend the Gaussian approximation potential (GAP) for silicon by incorporating the interaction with hydrogen, thereby significantly improving the degree of realism with which amorphous silicon can be modelled. We show that our Si:H GAP enables the simulation of hydrogenated silicon with an accuracy very close to DFT, but with computational expense and run times reduced by several orders of magnitude for large structures. We demonstrate the capabilities of the Si:H GAP by creating models of hydrogenated liquid and amorphous silicon, and showing that their energies, forces and stresses are in excellent agreement with DFT results, and their structure as captured by bond and angle distributions, with both DFT and experiments.

cond-mat.mtrl-sci↗

arXiv2021

Compressing local atomic neighbourhood descriptors

Many atomic descriptors are currently limited by their unfavourable scaling with the number of chemical elements $S$ e.g. the length of body-ordered descriptors, such as the Smooth Overlap of Atomic Positions (SOAP) power spectrum (3-body) and the Atomic Cluster Expansion (ACE) (multiple body-orders), scales as $(NS)^ν$ where $ν+1$ is the body-order and $N$ is the number of radial basis functions used in the density expansion. We introduce two distinct approaches which can be used to overcome this scaling for the SOAP power spectrum. Firstly, we show that the power spectrum is amenable to lossless compression with respect to both $S$ and $N$, so that the descriptor length can be reduced from $\mathcal{O}(N^2S^2)$ to $\mathcal{O}\left(NS\right)$. Secondly, we introduce a generalized SOAP kernel, where compression is achieved through the use of the total, element agnostic density, in combination with radial projection. The ideas used in the generalized kernel are equally applicably to any other body-ordered descriptors and we demonstrate this for the Atom Centered Symmetry Functions (ACSF). Finally, both compression approaches are shown to offer comparable performance to the original descriptor across a variety of numerical tests.

cond-mat.mtrl-sci↗

arXiv2021

Local invertibility and sensitivity of atomic structure-feature mappings

The increasingly common applications of machine-learning schemes to atomic-scale simulations have triggered efforts to better understand the mathematical properties of the mapping between the Cartesian coordinates of the atoms and the variety of representations that can be used to convert them into a finite set of symmetric descriptors or features. Here, we analyze the sensitivity of the mapping to atomic displacements, showing that the combination of symmetry and smoothness leads to mappings that have singular points at which the Jacobian has one or more null singular values (besides those corresponding to infinitesimal translations and rotations). This is in fact desirable, because it enforces physical symmetry constraints on the values predicted by regression models constructed using such representations. However, besides these symmetry-induced singularities, there are also spurious singular points, that we find to be linked to the incompleteness of the mapping, i.e. the fact that, for certain classes of representations, structurally distinct configurations are not guaranteed to be mapped onto different feature vectors. Additional singularities can be introduced by a too aggressive truncation of the infinite basis set that is used to discretize the representations.

physics.chem-ph↗

arXiv2021

Machine learning force fields based on local parametrization of dispersion interactions: Application to the phase diagram of C$_{60}$

We present a comprehensive methodology to enable addition of van der Waals (vdW) corrections to machine learning (ML) atomistic force fields. Using a Gaussian approximation potential (GAP) [Bartók et al., Phys. Rev. Lett. 104, 136403 (2010)] as baseline, we accurately machine learn a local model of atomic polarizabilities based on Hirshfeld volume partitioning of the charge density [Tkatchenko and Scheffler, Phys. Rev. Lett. 102, 073005 (2009)]. These environment-dependent polarizabilities are then used to parametrize a screened London-dispersion approximation to the vdW interactions. Our ML vdW model only needs to learn the charge density partitioning implicitly, by learning the reference Hirshfeld volumes from density functional theory (DFT). In practice, we can predict accurate Hirshfeld volumes from the knowledge of the local atomic environment (atomic positions) alone, making the model highly computationally efficient. For additional efficiency, our ML model of atomic polarizabilities reuses the same many-body atomic descriptors used for the underlying GAP learning of bonded interatomic interactions. We also show how the method enables straightforward computation of gradients of the observables, even when these remain challenging for the reference method (e.g., calculating gradients of the Hirshfeld volumes in DFT). Finally, we demonstrate the approach by studying the phase diagram of C$_{60}$, where vdW effects are important. The need for a highly accurate vdW-inclusive reactive force field is highlighted by modeling the decomposition of the C$_{60}$ molecules taking place at high pressures and temperatures.

cond-mat.mtrl-sci↗