22 May 2023 in The Journal of Chemical Physics
A reliable uncertainty estimator is a key ingredient in the successful use of machine-learning force fields for predictive calculations. Important considerations are correlation with error, overhead during training and inference, and efficient workflows to systematically improve the force field. However, in the case of neural-network force fields, simple committees are often the only option considered due to their easy implementation. Here, we present a generalization of the deep-ensemble design based on multiheaded neural networks and a heteroscedastic loss. It can efficiently deal with uncertainties in both energy and forces and take sources of aleatoric uncertainty affecting the training data into account. We compare uncertainty metrics based on deep ensembles, committees, and bootstrap-aggregation ensembles using data for an ionic liquid and a perovskite surface. We demonstrate an adversarial approach to active learning to efficiently and progressively refine the force fields. That active learning workflow is realistically possible thanks to exceptionally fast training enabled by residual learning and a nonlinear learned optimizer.
23 December 2021 in Journal of Chemical Information and Modeling
We present NeuralIL, a model for the potential energy of an ionic liquid that accurately reproduces first-principles results with orders-of-magnitude savings in computational cost. Built on the basis of a multilayer perceptron and spherical Bessel descriptors of the atomic environments, NeuralIL is implemented in such a way as to be fully automatically differentiable. It can thus be trained on ab initio forces instead of just energies, to make the most out of the available data, and can efficiently predict arbitrary derivatives of the potential energy. Using ethylammonium nitrate as the test system, we obtain out-of-sample accuracies better than 2 meV atom–1 (<0.05 kcal mol–1) in the energies and 70 meV Å–1 in the forces. We show that encoding the element-specific density in the spherical Bessel descriptors is key to achieving this. Harnessing the information provided by the forces drastically reduces the amount of atomic configurations required to train a neural network force field based on atom-centered descriptors. We choose the Swish-1 activation function and discuss the role of this choice in keeping the neural network differentiable. Furthermore, the possibility of training on small data sets allows for an ensemble-learning approach to the detection of extrapolation. Finally, we find that a separate treatment of long-range interactions is not required to achieve a high-quality representation of the potential energy surface of these dense ionic systems.
14 September 2016 in Journal of Physics: Condensed Matter
This work describes the behaviour of water molecules in 1-butyl-3-methylimidazolium tetrafluoroborate ionic liquid under nanoconfinement, between graphene sheets. By means of molecular dynamics simulations, the adsorption of water molecules at the graphene surface is studied. A depletion of water molecules in the vicinity of the neutral and negatively charged graphene surfaces, and their adsorption at the positively charged surface are observed in line with the preferential hydration of the ionic liquid anions. The findings are appropriately described using a two-level statistical model. The confinement effect on the structure and dynamics of the mixtures is thoroughly analyzed using the density and the potential of mean force profiles, as well as by the vibrational densities of the states of water molecules near the graphene surface. The orientation of water molecules and the water-induced structural transitions in the layer closest to the graphene surface are also discussed.
28 November 2016 in The Journal of Chemical Physics
TMTrinidad Méndez-Morales
+1
We report a molecular dynamics study on the effect of electronic polarization on the structure and single-particle dynamics of mixtures of the aprotic ionic liquid 1-ethyl-3-methylimidazolium bis-(trifluoromethylsulfonyl)-imide ([EMIM][TFSI]) doped with a lithium salt with the same anion at 298 K and 1 bar. In particular, we analyze the effect of electron density fluctuations on radial distribution functions, velocity autocorrelation functions, cage correlation functions, mean-squared displacements, and vibrational densities of states, comparing the predictions of the quantum-chemistry-based Atomistic Polarizable Potential for Liquids, Electrolytes, & Polymers (APPLE&P) with those of its nonpolarizable version and those of the standard non-polarizable Optimized Potentials for Liquid Simulations-All Atom (OPLS-AA). We found that the structure of the mixture is scarcely modified by the fluctuations in electron charge of their constituents, but their transport properties are indeed quite drastically changed, with larger mobilities being predicted for the different species in the bulk mixtures with the polarizable force field. Specifically, the mean-squared displacements are larger for the polarizable potentials at identical time intervals and the intermediate subdiffusive plateaus are greatly reduced, so the transition to the diffusive regime takes place much earlier than in the non-polarizable media. Moreover, the correlations of the added cations inside their cages are weakened out earlier and their vibrational densities of states are slightly red-shifted, reflecting the weakening effect of the electronic polarization on the Coulomb coupling in these dense ionic media. The comparison of OPLS-AA with non-polarizable APPLE&P indicates that adding polarization to OPLS-AA is not sufficient to achieve results close to experiments.
1 May 2020 in Journal of Molecular Liquids
28 March 2017 in The Journal of Chemical Physics
TMTrinidad Méndez-Morales
+2
The structural and dynamical properties of bulk mixtures of long-chained primary and secondary alcohols (propanol, butanol, and 2-pentanol) with protic ionic liquids (ethylammonium and butylammonium nitrate) were studied by means of molecular dynamics simulations and small angle X-ray scattering (SAXS). Changes in the structure with the alcohol concentration and with the alkyl chain length of the alcohol moieties were found, showing variations in the radial distribution function and in the number of hydrogen bonds in the bulk liquids. Moreover, the structural behaviour of the studied mixtures is further clarified with the spatial distribution functions. The global picture in the local scale is in good agreement with the nanostructured solvation paradigm [T. Méndez-Morales et al. Phys. Chem. B 118, 761 (2014)], according to which alcohols are accommodated into the hydrogen bonds’ network of the ionic liquid instead of forming clusters in the bulk. Indeed, our study reveals that the alcohol molecules are placed with their polar heads at the interfaces between polar and nonpolar nanodomains in the ionic liquid, with their alkyl chains inside the nonpolar organic nanodomains. The influence of alcohol chain length in the single-particle dynamics of the mixtures is also reported calculating the velocity autocorrelation function and vibrational densities of states of the different species in the ionic liquid-alcohol mixtures, and a weak caging effect for the ethylammonium cations independent of the chain size of the alcohols was found. However, the SAXS data collected for the studied mixtures show an excess of the scattering intensities which indicates that there are also some structural heterogeneities at the nanoscale.