-
Machine-learning interatomic potentials from a users perspective: A comparison of accuracy, speed and data efficiency
Authors:
Niklas Leimeroth,
Linus C. Erhard,
Karsten Albe,
Jochen Rohrer
Abstract:
Machine learning interatomic potentials (MLIPs) have massively changed the field of atomistic modeling. They enable the accuracy of density functional theory in large-scale simulations while being nearly as fast as classical interatomic potentials. Over the last few years, a wide range of different types of MLIPs have been developed, but it is often difficult to judge which approach is the best fo…
▽ More
Machine learning interatomic potentials (MLIPs) have massively changed the field of atomistic modeling. They enable the accuracy of density functional theory in large-scale simulations while being nearly as fast as classical interatomic potentials. Over the last few years, a wide range of different types of MLIPs have been developed, but it is often difficult to judge which approach is the best for a given problem setting. For the case of structurally and chemically complex solids, namely Al-Cu-Zr and Si-O, we benchmark a range of machine learning interatomic potential approaches, in particular, the Gaussian approximation potential (GAP), high-dimensional neural network potentials (HDNNP), moment tensor potentials (MTP), the atomic cluster expansion (ACE) in its linear and nonlinear version, neural equivariant interatomic potentials (NequIP), Allegro, and MACE. We find that nonlinear ACE and the equivariant, message-passing graph neural networks NequIP and MACE form the Pareto front in the accuracy vs. computational cost trade-off. In case of the Al-Cu-Zr system we find that MACE and Allegro offer the highest accuracy, while NequIP outperforms them for Si-O. Furthermore, GPUs can massively accelerate the MLIPs, bringing them on par with and even ahead of non-accelerated classical interatomic potentials (IPs) with regards to accessible timescales. Finally, we explore the extrapolation behavior of the corresponding potentials, probe the smoothness of the potential energy surfaces, and finally estimate the user friendliness of the corresponding fitting codes and molecular dynamics interfaces.
△ Less
Submitted 5 May, 2025;
originally announced May 2025.
-
Understanding phase transitions of $α$-quartz under dynamic compression conditions by machine-learning driven atomistic simulations
Authors:
Linus C. Erhard,
Christoph Otzen,
Jochen Rohrer,
Clemens Prescher,
Karsten Albe
Abstract:
Characteristic shock effects in silica serve as a key indicator of historical impacts at geological sites. Despite this geological significance, atomistic details of structural transformations under high pressure and shock compression remain poorly understood. This ambiguity is evidenced by conflicting experimental observations of both amorphization and crystallization transitions. Utilizing a new…
▽ More
Characteristic shock effects in silica serve as a key indicator of historical impacts at geological sites. Despite this geological significance, atomistic details of structural transformations under high pressure and shock compression remain poorly understood. This ambiguity is evidenced by conflicting experimental observations of both amorphization and crystallization transitions. Utilizing a newly developed machine-learning interatomic potential, we examine the response of $α$-quartz to shock compression with a peak pressure of 60 GPa over nano-second timescales. We initially observe amorphization before recrystallization into a d-NiAs-structured silica with disorder on the silicon sublattice, accompanied by the formation of domains with partial order of silicon. Investigating a variety of strain conditions enables us to identify the non-hydrostatic stress and strain states that allow the direct diffusionless formation of rosiaite-structured silica.
△ Less
Submitted 25 June, 2024;
originally announced June 2024.
-
Structure-property relations of silicon oxycarbides studied using a machine learning interatomic potential
Authors:
Niklas Leimeroth,
Jochen Rohrer,
Karsten Albe
Abstract:
Silicon oxycarbides show outstanding versatility due to their highly tunable composition and microstructure. Consequently, a key challenge is a thorough knowledge of structure-property relations in the system. In this work, we fit an atomic cluster expansion potential to a set of actively learned DFT training data spanning a wide configurational space. We demonstrate the ability of the potential t…
▽ More
Silicon oxycarbides show outstanding versatility due to their highly tunable composition and microstructure. Consequently, a key challenge is a thorough knowledge of structure-property relations in the system. In this work, we fit an atomic cluster expansion potential to a set of actively learned DFT training data spanning a wide configurational space. We demonstrate the ability of the potential to produce realistic amorphous structures and rationalize the formation of different morphologies of the turbostratic free carbon phase. Finally, we relate the materials stiffness to its composition and microstructure, finding a delicate dependence on Si-C bonds that contradicts commonly assumed relations to the free carbon phase.
△ Less
Submitted 15 March, 2024;
originally announced March 2024.
-
From electrons to phase diagrams with classical and machine learning potentials: automated workflows for materials science with pyiron
Authors:
Sarath Menon,
Yury Lysogorskiy,
Alexander L. M. Knoll,
Niklas Leimeroth,
Marvin Poul,
Minaam Qamar,
Jan Janssen,
Matous Mrovec,
Jochen Rohrer,
Karsten Albe,
Jörg Behler,
Ralf Drautz,
Jörg Neugebauer
Abstract:
We present a comprehensive and user-friendly framework built upon the pyiron integrated development environment (IDE), enabling researchers to perform the entire Machine Learning Potential (MLP) development cycle consisting of (i) creating systematic DFT databases, (ii) fitting the Density Functional Theory (DFT) data to empirical potentials or MLPs, and (iii) validating the potentials in a largel…
▽ More
We present a comprehensive and user-friendly framework built upon the pyiron integrated development environment (IDE), enabling researchers to perform the entire Machine Learning Potential (MLP) development cycle consisting of (i) creating systematic DFT databases, (ii) fitting the Density Functional Theory (DFT) data to empirical potentials or MLPs, and (iii) validating the potentials in a largely automatic approach. The power and performance of this framework are demonstrated for three conceptually very different classes of interatomic potentials: an empirical potential (embedded atom method - EAM), neural networks (high-dimensional neural network potentials - HDNNP) and expansions in basis sets (atomic cluster expansion - ACE). As an advanced example for validation and application, we show the computation of a binary composition-temperature phase diagram for Al-Li, a technologically important lightweight alloy system with applications in the aerospace industry.
△ Less
Submitted 8 March, 2024;
originally announced March 2024.
-
Modelling atomic and nanoscale structure in the silicon-oxygen system through active machine learning
Authors:
Linus C. Erhard,
Jochen Rohrer,
Karsten Albe,
Volker L. Deringer
Abstract:
Silicon-oxygen compounds are among the most important ones in the natural sciences, occurring as building blocks in minerals and being used in semiconductors and catalysis. Beyond the well known silicon dioxide, there are phases with different stoichiometric composition and nanostructured composites. One of the key challenges in understanding the Si-O system is therefore to accurately account for…
▽ More
Silicon-oxygen compounds are among the most important ones in the natural sciences, occurring as building blocks in minerals and being used in semiconductors and catalysis. Beyond the well known silicon dioxide, there are phases with different stoichiometric composition and nanostructured composites. One of the key challenges in understanding the Si-O system is therefore to accurately account for its nanoscale heterogeneity beyond the length scale of individual atoms. Here we show that a unified computational description of the full Si-O system is indeed possible, based on atomistic machine learning coupled to an active-learning workflow. We showcase applications to very-high-pressure silica, to surfaces and aerogels, and to the structure of amorphous silicon monoxide. In a wider context, our work illustrates how structural complexity in functional materials beyond the atomic and few-nanometre length scales can be captured with active machine learning.
△ Less
Submitted 7 September, 2023;
originally announced September 2023.
-
Interface-controlled creep in metallic glass composites
Authors:
Constanze Kalcher,
Tobias Brink,
Jochen Rohrer,
Alexander Stukowski,
Karsten Albe
Abstract:
In this work we present molecular dynamics simulations on the creep behavior of $\rm Cu_{64}Zr_{36}$ metallic glass composites. Surprisingly, all composites exhibit much higher creep rates than the homogeneous glass. The glass-crystal interface can be viewed as a weak interphase, where the activation barrier of shear transformation zones is lower than in the surrounding glass. We observe that the…
▽ More
In this work we present molecular dynamics simulations on the creep behavior of $\rm Cu_{64}Zr_{36}$ metallic glass composites. Surprisingly, all composites exhibit much higher creep rates than the homogeneous glass. The glass-crystal interface can be viewed as a weak interphase, where the activation barrier of shear transformation zones is lower than in the surrounding glass. We observe that the creep behavior of the composites does not only depend on the interface area but also on the orientation of the interface with respect to the loading axis. We propose an explanation in terms of different mean Schmid factors of the interfaces, with the amorphous interface regions acting as preferential slip sites.
△ Less
Submitted 25 September, 2018; v1 submitted 11 August, 2017;
originally announced August 2017.
-
Influence of elastic strain on the thermodynamics and kinetics of lithium vacancy in bulk LiCoO2
Authors:
Ashkan Moradabadi,
Payam Kaghazchi,
Jochen Rohrer,
Karsten Albe
Abstract:
The influence of elastic strain on the lithium vacancy formation and migration in bulk LiCoO2 is evaluated by means of first-principles calculations within density functional theory (DFT). Strain dependent energies are determined directly from defective cells and also within linear elasticity theory from the elastic dipole tensor (Gij) for ground state and saddle point configurations. We analyze f…
▽ More
The influence of elastic strain on the lithium vacancy formation and migration in bulk LiCoO2 is evaluated by means of first-principles calculations within density functional theory (DFT). Strain dependent energies are determined directly from defective cells and also within linear elasticity theory from the elastic dipole tensor (Gij) for ground state and saddle point configurations. We analyze finite size-effects in the calculation of Gij, compare the predictions of the linear elastic model with those obtained from direct calculations of defective cells under strain and discuss the differences. Based on our data, we calculate the variations in vacancy concentration and mobility due to the presence of external strain in bulk LiCoO2 cathodes. Our results reveal that elastic in-plane and out-of-plane strains can significantly change the ionic conductivity of bulk LiCoO2 by an order of magnitude and thus strongly affect the performance of Li-secondary batteries.
△ Less
Submitted 6 June, 2017;
originally announced June 2017.
-
Role of sonication pre-treatment and cation valence in nano-cellulose suspensions sol-gel transition
Authors:
Cecilia Ada Maestri,
Michela Abrami,
Sharon Hazan,
Elena Chistè,
Yuval Golan,
Julia Rohrer,
Andreas Bernkop-Schnürch,
Mario Grassi,
Marina Scarpa,
Paolo Bettotti
Abstract:
Sol-gel transition of carboxylated cellulose nanocrystals is investigated using rheology, SAXS, NMR and optical spectroscopies to unveil the distinctive roles of ultrasounds treatment and ions addition. Besides cellulose fibers fragmentation, sonication treatment induces fast gelling of the solution. Gelation is induced independently on the addition of cations, while the final rheological properti…
▽ More
Sol-gel transition of carboxylated cellulose nanocrystals is investigated using rheology, SAXS, NMR and optical spectroscopies to unveil the distinctive roles of ultrasounds treatment and ions addition. Besides cellulose fibers fragmentation, sonication treatment induces fast gelling of the solution. Gelation is induced independently on the addition of cations, while the final rheological properties are highly influenced by the type, the concentration as well as on the sequence of the operations since salts must be added before sonication to produce stiff gels. Cations with various charge and dimension have been associated to ultrasounds to induce gelation and the gel elastic modulus increase proportionally with the charge over the ion size ratio. SAXS analysis of the Na+ hydrogel and Ca2+ hydrogel to which the ion was added after sonication shows the presence of structurally ordered domains where water is confined as indicated by 1H-NMR investigation of the dynamic of water exchange in the hydrogels. Conversely, separated phases containing essentially free water, characterize the hydrogels obtained by sonication after Ca2+ addition, confirming that this ion induces irreversible fiber aggregation. The rheological properties of the hydrogels depend on the duration of the ultrasound treatment and it enables the design of materials programmed with tailored energy dissipation response.
△ Less
Submitted 19 May, 2017;
originally announced May 2017.
-
Relative stability of $6H$-SiC$\{0001\}$ surface terminations and formation of graphene overlayers by Si evaporation
Authors:
Jochen Rohrer,
Eleni Ziambaras,
Per Hyldgaard
Abstract:
We present density functional theory (DFT) calculations for 6H-SiC$\{0001\}$ surfaces with different surface stackings and terminations. We compare the relative stability of different $(0001)$ and $(000\bar1)$ surfaces in terms of their surface free energies. Removing surface and subsurface Si atoms, we simulate the formation of graphene and graphene-like overlayers by Si evaporation. We find that…
▽ More
We present density functional theory (DFT) calculations for 6H-SiC$\{0001\}$ surfaces with different surface stackings and terminations. We compare the relative stability of different $(0001)$ and $(000\bar1)$ surfaces in terms of their surface free energies. Removing surface and subsurface Si atoms, we simulate the formation of graphene and graphene-like overlayers by Si evaporation. We find that overlayers with a different nature of bonding are preferred at the two non-equivalent surface orientations. At $(0001)$, a chemically bonded, highly strained and buckled film is predicted. At $(000\bar1)$, a van der Waals (vdW) bonded overlayer is preferred. We quantify the vdW binding and show that it can have a doping effect on electron behavior in the overlayer.
△ Less
Submitted 10 February, 2011;
originally announced February 2011.
-
Computational scheme for ab-initio predictions of chemical compositions interfaces realized by deposition growth
Authors:
Jochen Rohrer,
Per Hyldgaard
Abstract:
We present a novel computational scheme to predict chemical compositions at interfaces as they emerge in a growth process. The scheme uses the Gibbs free energy of reaction associated with the formation of interfaces with a specific composition as predictor for their prevalence. It explicitly accounts for the growth conditions by rate-equation modeling of the deposition environment. The Bell-Evans…
▽ More
We present a novel computational scheme to predict chemical compositions at interfaces as they emerge in a growth process. The scheme uses the Gibbs free energy of reaction associated with the formation of interfaces with a specific composition as predictor for their prevalence. It explicitly accounts for the growth conditions by rate-equation modeling of the deposition environment. The Bell-Evans-Polanyi principle motivates our emphasis on an effective nonequilibrium thermodynamic description inspired by chemical reaction theory. We illustrate the scheme by characterizing the interface between TiC and alumina. Equilibrium thermodynamics favors a nonbinding interface, being in conflict with the wear-resistant nature of TiC/alumina multilayer coatings. Our novel scheme predicts that deposition of a strongly adhering interface is favored under realistic conditions.
△ Less
Submitted 5 January, 2011;
originally announced January 2011.
-
Stacking and band structure of van der Waals bonded graphane multilayers
Authors:
Jochen Rohrer,
Per Hyldgaard
Abstract:
We use density functional theory and the van der Waals density functional (vdW-DF) method to determine the binding separation in bilayer and bulk graphane and study the changes in electronic band structure that arise with the multilayer formation. The calculated binding separation (distance between center-of-mass planes) and binding energy are 4.5-5.0 Å (4.5-4.8 Å) and 75-102 meV/cell (93-127 meV/…
▽ More
We use density functional theory and the van der Waals density functional (vdW-DF) method to determine the binding separation in bilayer and bulk graphane and study the changes in electronic band structure that arise with the multilayer formation. The calculated binding separation (distance between center-of-mass planes) and binding energy are 4.5-5.0 Å (4.5-4.8 Å) and 75-102 meV/cell (93-127 meV/cell) in the bilayer (bulk), depending on the choice of vdW-DF version. We obtain the corresponding band diagrams using calculations in the ordinary generalized gradient approximation for the geometries specified by our vdW-DF results, so probing the indirect effect of vdW forces on electron behavior. We find significant band-gap modifications by up to -1.2 eV (+4.0 eV) in various regions of the Brillouin zone, produced by the bilayer (bulk) formation.
△ Less
Submitted 10 February, 2011; v1 submitted 14 October, 2010;
originally announced October 2010.
-
Understanding adhesion at as-deposited interfaces from ab initio thermodynamics of deposition growth: thin-film alumina on titanium carbide
Authors:
Jochen Rohrer,
Per Hyldgaard
Abstract:
We investigate the chemical composition and adhesion of chemical vapour deposited thin-film alumina on TiC using and extending a recently proposed nonequilibrium method of ab initio thermodynamics of deposition growth (AIT-DG) [Rohrer J and Hyldgaard P 2010 Phys. Rev. B 82 045415]. A previous study of this system [Rohrer J, Ruberto C and Hyldgaard P 2010 J. Phys.: Condens. Matter 22 015004] found…
▽ More
We investigate the chemical composition and adhesion of chemical vapour deposited thin-film alumina on TiC using and extending a recently proposed nonequilibrium method of ab initio thermodynamics of deposition growth (AIT-DG) [Rohrer J and Hyldgaard P 2010 Phys. Rev. B 82 045415]. A previous study of this system [Rohrer J, Ruberto C and Hyldgaard P 2010 J. Phys.: Condens. Matter 22 015004] found that use of equilibrium thermodynamics leads to predictions of a non-binding TiC/alumina interface, despite the industrial use as a wear-resistant coating. This discrepancy between equilibrium theory and experiment is resolved by the AIT-DG method which predicts interfaces with strong adhesion. The AIT-DG method combines density functional theory calculations, rate-equation modelling of the pressure evolution of the deposition environment and thermochemical data. The AIT-DG method was previously used to predict prevalent terminations of growing or as-deposited surfaces of binary materials. Here we extent the method to predict surface and interface compositions of growing or as-deposited thin films on a substrate and find that inclusion of the nonequilibrium deposition environment has important implications for the nature of buried interfaces.
△ Less
Submitted 6 September, 2010;
originally announced September 2010.
-
Ab-initio thermodynamics of deposition growth: surface terminations of CVD titanium carbide and nitride
Authors:
Jochen Rohrer,
Per Hyldgaard
Abstract:
We present a calculational method to predict terminations of growing or as-deposited surfaces as a function of the deposition conditions. Such characterizations are valuable for understanding catalysis and growth phenomena. The method combines ab-initio density functional theory (DFT) calculations and experimental thermodynamical data with a rate-equations description of partial pressures in the r…
▽ More
We present a calculational method to predict terminations of growing or as-deposited surfaces as a function of the deposition conditions. Such characterizations are valuable for understanding catalysis and growth phenomena. The method combines ab-initio density functional theory (DFT) calculations and experimental thermodynamical data with a rate-equations description of partial pressures in the reaction chamber. The use of rate equations enables a complete description of a complex gas environment in terms of a few, (experimentally accessible) parameters. The predictions are based on comparisons between free energies of reaction associated with the formation of surfaces with different terminations. The method has an intrinsic non-equilibrium character. In the limit of dynamic equilibrium (with equal chemical potential in the surface and the gas phase) we find that the predictions of the method coincide with those of standard ab-initio based equilibrium thermodynamics. We illustrate the method for chemical vapor deposition (CVD) of TiC(111) and TiN(111), and find that the emerging termination can be controlled both by the environment and the growth rate.
△ Less
Submitted 28 June, 2010; v1 submitted 12 April, 2010;
originally announced April 2010.
-
Ab initio structure modeling of complex thin-film oxides: thermodynamical stability of TiC/thin-film alumina
Authors:
Jochen Rohrer,
Carlo Ruberto,
Per Hyldgaard
Abstract:
We present an efficient and general method to identify promising candidate configurations for thin-film oxides and to determine structural characteristics of (metastable) thin-film structures using ab initio calculations. At the heart of this method is the complexity of the oxide bulk structure, from which a large number of thin films with structural building blocks, that is motifs, from metasta…
▽ More
We present an efficient and general method to identify promising candidate configurations for thin-film oxides and to determine structural characteristics of (metastable) thin-film structures using ab initio calculations. At the heart of this method is the complexity of the oxide bulk structure, from which a large number of thin films with structural building blocks, that is motifs, from metastable bulk oxide systems can be extracted. These span a coarse but well-defined network of initial configurations for which density functional theory (DFT) calculations predict and implement dramatic atomic relaxations in the corresponding, resulting thin-film candidates. The network of thin-film candidates (for various film thicknesses and stoichiometries) can be ordered according to their variation in ab initio total energy or in ab initio equilibrium Gibbs free energy. Analysis of the relaxed atomic structures for the most favored structures gives insight into the nature of stable and metastable thin-film oxides. We investigate ultrathin alumina nucleated on TiC as a model system to illustrate this method.
△ Less
Submitted 6 March, 2009;
originally announced March 2009.
