Session P22: Emerging Trends in MD Simulations and Machine Learning V

Live

The Self-Learning Kinetic Monte Carlo (SLKMC) method [1] with its usage of a pattern recognition enabled the collection of a large database of diffusion pathways and their energetics for two-dimensional adatom islands containing 2-50 atoms on fcc(111) metal surfaces. A variety of diffusion mechanisms involving single and multiple island atoms were uncovered in long time (comparable to those in experiments) KMC simulations. In this talk, I will present results for the diffusion kinetics of two dimensional adatoms islands in two types of systems: homoepitaxial and heteroepitaxial. With examples of the diffusion of Ag and Pd adatom islands Ag(111) and Pd(111), and that of Cu and Ni adatoms islands on Ni(111) and Cu(111) [3], I will draw attention to the relative role of lateral interactions and binding energy in the size dependence of the island diffusion characteristics. Turning to the application of data driven techniques for extraction of descriptors and training of predictive models, I will present a summary of our recent success in extracting activation energy barriers that are accurate and obtained with little computational cost. These results are very promising for the development of tools suitable for multiscale modeling of the morphological evolution of nanostructured systems. Efforts in developing neural network derived interatomic potential informed from high throughput density functional theory calculations will also be presented.   

Live

We present a new machine learning method called multiscale reweighted stochastic embedding (MRSE) [1] for automatically constructing collective variables (CVs) to represent and drive the sampling of free energy landscapes in enhanced sampling simulations. The technique automatically finds CVs by learning a low-dimensional embedding of the high-dimensional feature space to the latent space via a deep neural network. Our work builds upon the popular t-distributed stochastic neighbor embedding approach [2]. We introduce several new aspects to stochastic neighbor embedding algorithms that make MRSE especially suitable for enhanced sampling simulations: (1) a well-tempered landmark selection scheme; (2) a multiscale representation of the high-dimensional feature space; and (3) a reweighting procedure to account for biased training data. We show the performance of MRSE by applying it to several model systems.

[1] J. Rydzewski, and O. Valsson, arXiv:2007.06377. https://arxiv.org/abs/2007.06377
[2] L. Maaten, and G. Hinton. J. Mach. Learn. Res. (2008) 9, 2579-2605.   

Live

Two temperature molecular dynamics based on finite electron temperature density functional theory (FT-DFT) has been successful in modeling optical excitation in materials as repopulation of the Fermi-Dirac distribution due to strong electron-field interactions. We employ FT-DFT based molecular dynamics to study the ferroelectric perovskite PbTiO₃ and find photo-induced charge transfer activation of optical phonons leads to a sub picosecond tetragonal/ferroelectric to cubic/paraelectric structural phase transition at 300K in the bulk structure. To study large polar domains common in PbTiO₃ based nanostructures we developed a neural-network force-field model based on ground state DFT and FT-DFT training data. Applications of the model to study polar vortex formation, domain wall structure, and their interactions in PbTiO₃ based nanostructures will be discussed.   

Live

Machine learned (ML) interatomic potentials have emerged as a powerful tool for performing large-scale molecular dynamics simulations at near-DFT accuracy, but training many-body ML potentials that are interpretable, efficient, and uncertainty-aware remains an important open challenge. In this talk, we present Bayesian force fields that unite three frameworks—the Atomic Cluster Expansion (ACE), Gaussian Approximation Potentials (GAP), and Spectral Neighbor Analysis Potentials (SNAP)—opening the door to scalable, uncertainty-aware molecular dynamics simulations of complex materials. We use a multi-species generalization of the N-body ACE descriptor to define a tunable many-body kernel for sparse Gaussian process (GP) regression, and show that mean predictions of the GP can be mapped onto equivalent and much faster models resembling SNAP and qSNAP models. The Bayesian error bars provided by the sparse GP make it possible to train force fields on the fly during molecular dynamics, and we apply this automated training protocol to model phase transitions in the shape memory alloy nickel titanium.   

Live

Free energy perturbation [1] is a bedrock technique for estimation of free energy differences. Fast, reliable convergence of the free energy difference, however, demands that the respective distributions share a large overlap in configuration space. One strategy to address this requirement is Targeted Free Energy Perturbation (TFEP) [2], whereby a bijective mapping on configuration space is used to increase an effective overlap. Despite its appeal, TFEP has seen little use in practice since it relies on handcrafting effective mappings. Here we turn TFEP into a machine learning problem whereby the mapping is represented by a deep neural network whose parameters are optimized so as to maximize overlap. We test the approach on a prototypical solvation system, employing a novel normalizing flow architecture that respects periodic boundary conditions and permutational symmetry of identical particles. Our technique leads to significant error reduction in free energy estimates compared to baselines, without requiring additional data.

[1] R. W. Zwanzig, J. Chem. Phys. 22, 1420 (1954)
[2] C. Jarzynski, Phys. Rev. E 65, 046122 (2002)

*PW and AJB contributed equally to this work.

This research has been first published in J. Chem. Phys. 153, 144112 (2020), with the permission of AIP Publishing.   

Live

Accurate computational models of the dynamics that govern restructuring of bimetallic alloy surfaces could accelerate experimental insights, guide materials synthesis efforts, and facilitate materials discovery for applications such as single-atom or single-cluster catalysis. Trends which bridge alloy composition and restructuring behavior would be valuable to future experiments. Machine learning force field models enable the study of long time scale molecular dynamics for large systems, but generating and selecting training data used to fit these models is a tedious and challenging task. Here, we demonstrate how integrating the FLARE codebase with workflow automation software enables data generation, model training, and iterative generation of further training data in a closed-loop cycle with minimal supervision. We apply the force fields generated from this process to a wide range of transition metals and bimetallic alloys to uncover trends in the relationship between chemical composition and restructuring behavior.   

Live

We present a study of Si bulk in diamond structure by fixed-node (FN) QMC since systems with Si tend to exhibit some of the smallest FN errors. This is due to the prevalence of single, spatially separated bonds as well as favorable spatial character and energetic ordering of the atomic levels. These features make it very convenient to study finite-size effects and scaling of both ground and excited states as the cell size is increased. This is carried out for HF and DFT sets of initial orbitals used in the construction of correlated wave functions in VMC and DMC. The cohesive energy is obtained very accurately, the FN error being ∼1% of the correlation energy that provides E_coh=4.629±0.08±0.002 eV where 1^st uncertainty is an estimated systematic error while 2^nd one is random error. The excitations Γ→Γ, and Γ→Χ are also studied using large cells of 64 and 216 atoms. We found that the thermodynamic limit (TDL) gaps can be reliably estimated using only these 2 data points with residual errors ≈0.2 eV. The bandgaps are found to be influenced by the orbital sets, and by the size and type of cells used. We emphasize the importance of reaching the TDL consistency as well as proper analysis in obtaining reliable estimates.   

Live

Coarse-grained (CG) models of polymers are developed to reduce computational effort yet capture the relaxation behavior imparted by the hierarchal structure. Bottom-up CG methods are parameterized using atomistic reference data, and in the case of Iterative Boltzmann inversion (IBI), target the recovery of the chemically specific structure and thus thermodynamics, but at the cost of dramatically sped up dynamics. Here, we aim to develop a chemically specific, thermodynamically consistent, and dynamically correct model by combining a conservative potential and a dissipative potential. The conservative potential is parametrized via IBI from short atomistic simulations; the CG force-field is tuned to recover the pair distribution function of the atomistic representation, and thus, thermodynamics. The dissipative potential is introduced to correct the dynamics of the CG force-field. Here, we apply a Langevin thermostat and characterize the friction factor needed to recover atomistic dynamics. We discuss the consistency of the parameterizable friction factor for melt state oligomers as characterized by multiple modes of dynamics and property calculations.   

Live

Descriptors of the environment of an atom in a material give a similarity metric between different structures, to be used in interatomic potentials or for analyzing patterns in large sets of atomistic configurations. Since different species have different typical bond lengths, the length scales that optimally describe them also differ. We present heuristics for a universal set of multiscale Smooth Overlap of Atomic Positions (SOAP) descriptors. They describe short range interactions precisely, include longer range interactions more approximately, handle the diversity of hetero-atomic length scales, and are stable when new species are added. Using them we have extended our automatic interatomic potential generation method [1] to multicomponent systems, implemented in a simple python workflow for structure search, configuration selection, and reference energy calculations required for the fit. We present results for generated single and multi-species potentials, and use the same descriptors to analyze the diversity in the generated databases.

[1] Bernstein et al., npj Comput. Mater. v. 5, 99 (2019)   

Live

Machine learning potentials (MLPs) play a vital role in atomistic simulations due to their nearly ab-initio accuracy and their high computational efficiency similar to empirical force fields. However, most MLPs depend on local atomic environments only without taking global charge distributions into account thus neglecting non-local effects. Recently, some non-local MLPs have been proposed [1-3] for tackling these problems. They are not only based on local environments but also on non-local electronic information, i.e. atomic charges depending on long-range charge transfer and the global charge. Here we use fourth-generation high-dimensional neural network potentials [3] to illustrate the role of non-local effects and suggest possible improvements for current state-of-the-art MLPs.
References:
1: Xie, X.; Persson, K. A.; Small, D. W. J. Chem. Theory Comput. 2020, 16, 4256–4270
2: Zubatyuk, R.; Smith, J. S.; Leszczynski, J.; Isayev, O. Sci. Adv. 2019, 5, eaav6490
3: Ko, T. W.; Finkler, J. A.; Goedecker, S.; Behler, J. arXiv:2009.064842020   

Live

Diffusion Monte Carlo (DMC) is a method for computing the Fermionic ground state that represents the wave function as a density of points called walkers, which converge to the ground state by moving in a stochastic way. The main limitation of DMC is that it requires a-priori knowledge of the sign of the wave function. The standard practice is to set this sign equal to that of a trial function, which is typically an approximate wave function produced by some other method. This fixed node approximation limits the accuracy of DMC to the accuracy of the method that produces the trial function. The work presented here explores a novel method that optimizes the location nodal surface from the walker distribution of DMC. The parameters that determine the location of the nodal surface are continually improved via gradient descent, to minimize the DMC energy.   

On Demand

The accuracy and efficiency of quantum Monte Carlo (QMC) methods can be improved directly by adapting the many body wavefunction employed in the importance sampling. Full wavefunction optimization introduced a decade ago enabled solving scientific challenges beyond chemical accuracy even with a very compact wavefunction ansatz. Recently algorithmic development paved the way for simulation with large electron counts although demonstrations were mostly limited to small molecular systems. When applying QMC in solids, using fixed orbitals calculated by density functional theory (DFT) together with variationally optimized Jastrow factors is still the most practical way. Recent study on bandgaps hints the necessity of improving single particle orbitals beyond Hartree-Fock or DFT. Thus, we enable orbital optimization schemes like rotating an extended set of orbitals and further directly optimizing orbital shapes with much more challenging amount of paramters. The improvement are benchmarked on a few selected solid state systems.