Machine learning potentials for complex aqueous systems made simple.

Simulation techniques based on accurate and efficient representations of potential energy surfaces are urgently needed for the understanding of complex systems such as solid-liquid interfaces. Here we present a machine learning framework that enables the efficient development and validation of models for complex aqueous systems. Instead of trying to deliver a globally optimal machine learning potential, we propose to develop models applicable to specific thermodynamic state points in a simple and user-friendly process. After an initial ab initio simulation, a machine learning potential is constructed with minimum human effort through a data-driven active learning protocol. Such models can afterward be applied in exhaustive simulations to provide reliable answers for the scientific question at hand or to systematically explore the thermal performance of ab initio methods. We showcase this methodology on a diverse set of aqueous systems comprising bulk water with different ions in solution, water on a titanium dioxide surface, and water confined in nanotubes and between molybdenum disulfide sheets. Highlighting the accuracy of our approach with respect to the underlying ab initio reference, the resulting models are evaluated in detail with an automated validation protocol that includes structural and dynamical properties and the precision of the force prediction of the models. Finally, we demonstrate the capabilities of our approach for the description of water on the rutile titanium dioxide (110) surface to analyze the structure and mobility of water on this surface. Such machine learning models provide a straightforward and uncomplicated but accurate extension of simulation time and length scales for complex systems.

Developing machine-learned potentials to simultaneously capture the dynamics of excess protons and hydroxide ions in classical and path integral simulations.

The transport of excess protons and hydroxide ions in water underlies numerous important chemical and biological processes. Accurately simulating the associated transport mechanisms ideally requires utilizing ab initio molecular dynamics simulations to model the bond breaking and formation involved in proton transfer and path-integral simulations to model the nuclear quantum effects relevant to light hydrogen atoms. These requirements result in a prohibitive computational cost, especially at the time and length scales needed to converge proton transport properties. Here, we present machine-learned potentials (MLPs) that can model both excess protons and hydroxide ions at the generalized gradient approximation and hybrid density functional theory levels of accuracy and use them to perform multiple nanoseconds of both classical and path-integral proton defect simulations at a fraction of the cost of the corresponding ab initio simulations. We show that the MLPs are able to reproduce ab initio trends and converge properties such as the diffusion coefficients of both excess protons and hydroxide ions. We use our multi-nanosecond simulations, which allow us to monitor large numbers of proton transfer events, to analyze the role of hypercoordination in the transport mechanism of the hydroxide ion and provide further evidence for the asymmetry in diffusion between excess protons and hydroxide ions.

Electronic structure of the solvated benzene radical anion.

The benzene radical anion is a molecular ion pertinent to several organic reactions, including the Birch reduction of benzene in liquid ammonia. The species exhibits a dynamic Jahn-Teller effect due to its open-shell nature and undergoes pseudorotation of its geometry. Here, we characterize the complex electronic structure of this condensed-phase system based on ab initio molecular dynamics simulations and GW calculations of the benzene radical anion solvated in liquid ammonia. Using detailed analysis of the molecular and electronic structure, we find that the spatial character of the excess electron of the solvated radical anion follows the underlying Jahn-Teller distortions of the molecular geometry. We decompose the electronic density of states to isolate the contribution of the solute and to examine the response of the solvent to its presence. Our findings show the correspondence between instantaneous molecular structure and spin density; provide important insights into the electronic stability of the species, revealing that it is, indeed, a bound state in the condensed phase; and offer electronic densities of states that aid in the interpretation of experimental photoelectron spectra.

Benzene Radical Anion Microsolvated in Ammonia Clusters: Modeling the Transition from an Unbound Resonance to a Bound Species.

The benzene radical anion, well-known in organic chemistry as the first intermediate in the Birch reduction of benzene in liquid ammonia, exhibits intriguing properties from the point of view of quantum chemistry. Notably, it has the character of a metastable shape resonance in the gas phase, while measurements in solution find it to be experimentally detectable and stable. In this light, our previous calculations performed in bulk liquid ammonia explicitly reveal that solvation leads to stabilization. Here, we focus on the transition of the benzene radical anion from an unstable gas-phase ion to a fully solvated bound species by explicit ionization calculations of the radical anion solvated in molecular clusters of increasing size. The computational cost of the largest systems is mitigated by combining density functional theory with auxiliary methods including effective fragment potentials or approximating the bulk by polarizable continuum models. Using this methodology, we obtain the cluster size dependence of the vertical binding energy of the benzene radical anion converging to the value of -2.3 eV at a modest computational cost.

Committee neural network potentials control generalization errors and enable active learning.

It is well known in the field of machine learning that committee models improve accuracy, provide generalization error estimates, and enable active learning strategies. In this work, we adapt these concepts to interatomic potentials based on artificial neural networks. Instead of a single model, multiple models that share the same atomic environment descriptors yield an average that outperforms its individual members as well as a measure of the generalization error in the form of the committee disagreement. We not only use this disagreement to identify the most relevant configurations to build up the model's training set in an active learning procedure but also monitor and bias it during simulations to control the generalization error. This facilitates the adaptive development of committee neural network potentials and their training sets while keeping the number of ab initio calculations to a minimum. To illustrate the benefits of this methodology, we apply it to the development of a committee model for water in the condensed phase. Starting from a single reference ab initio simulation, we use active learning to expand into new state points and to describe the quantum nature of the nuclei. The final model, trained on 814 reference calculations, yields excellent results under a range of conditions, from liquid water at ambient and elevated temperatures and pressures to different phases of ice, and the air-water interface-all including nuclear quantum effects. This approach to committee models will enable the systematic development of robust machine learning models for a broad range of systems.

Valence and Core-Level X-ray Photoelectron Spectroscopy of a Liquid Ammonia Microjet.

Photoelectron spectroscopy of microjets expanded into vacuum allows access to orbital energies for solute or solvent molecules in the liquid phase. Microjets of water, acetonitrile and alcohols have previously been studied; however, it has been unclear whether jets of low temperature molecular solvents could be realized. Here we demonstrate a stable 20 µm jet of liquid ammonia (-60 °C) in a vacuum, which we use to record both valence and core-level band photoelectron spectra using soft X-ray synchrotron radiation. Significant shifts from isolated ammonia in the gas-phase are observed, as is the liquid-phase photoelectron angular anisotropy. Comparisons with spectra of ammonia in clusters and the solid phase, as well as spectra for water in various phases potentially reveal how hydrogen bonding is reflected in the condensed phase electronic structure.

Decoding the spectroscopic features and time scales of aqueous proton defects.

Acid solutions exhibit a variety of complex structural and dynamical features arising from the presence of multiple interacting reactive proton defects and counterions. However, disentangling the transient structural motifs of proton defects in the water hydrogen bond network and the mechanisms for their interconversion remains a formidable challenge. Here, we use simulations treating the quantum nature of both the electrons and nuclei to show how the experimentally observed spectroscopic features and relaxation time scales can be elucidated using a physically transparent coordinate that encodes the overall asymmetry of the solvation environment of the proton defect. We demonstrate that this coordinate can be used both to discriminate the extremities of the features observed in the linear vibrational spectrum and to explain the molecular motions that give rise to the interconversion time scales observed in recent nonlinear experiments. This analysis provides a unified condensed-phase picture of the proton structure and dynamics that, at its extrema, encompasses proton sharing and spectroscopic features resembling the limiting Eigen [H3O(H2O)3]+ and Zundel [H(H2O)2]+ gas-phase structures, while also describing the rich variety of interconverting environments in the liquid phase.

Competing quantum effects in the free energy profiles and diffusion rates of hydrogen and deuterium molecules through clathrate hydrates.

Clathrate hydrates hold considerable promise as safe and economical materials for hydrogen storage. Here we present a quantum mechanical study of H2 and D2 diffusion through a hexagonal face shared by two large cages of clathrate hydrates over a wide range of temperatures. Path integral molecular dynamics simulations are used to compute the free-energy profiles for the diffusion of H2 and D2 as a function of temperature. Ring polymer molecular dynamics rate theory, incorporating both exact quantum statistics and approximate quantum dynamical effects, is utilized in the calculations of the H2 and D2 diffusion rates in a broad temperature interval. We find that the shape of the quantum free-energy profiles and their height relative to the classical free energy barriers at a given temperature, as well as the rate of diffusion, are strongly affected by competing quantum effects: above 25 K, zero-point energy (ZPE) perpendicular to the reaction path for diffusion between cavities decreases the quantum rate compared to the classical rate, whereas at lower temperatures tunneling outcompetes the ZPE and as a result the quantum rate is greater than the classical rate.

Ab initio molecular dynamics with nuclear quantum effects at classical cost: Ring polymer contraction for density functional theory.

Path integral molecular dynamics simulations, combined with an ab initio evaluation of interactions using electronic structure theory, incorporate the quantum mechanical nature of both the electrons and nuclei, which are essential to accurately describe systems containing light nuclei. However, path integral simulations have traditionally required a computational cost around two orders of magnitude greater than treating the nuclei classically, making them prohibitively costly for most applications. Here we show that the cost of path integral simulations can be dramatically reduced by extending our ring polymer contraction approach to ab initio molecular dynamics simulations. By using density functional tight binding as a reference system, we show that our ring polymer contraction scheme gives rapid and systematic convergence to the full path integral density functional theory result. We demonstrate the efficiency of this approach in ab initio simulations of liquid water and the reactive protonated and deprotonated water dimer systems. We find that the vast majority of the nuclear quantum effects are accurately captured using contraction to just the ring polymer centroid, which requires the same number of density functional theory calculations as a classical simulation. Combined with a multiple time step scheme using the same reference system, which allows the time step to be increased, this approach is as fast as a typical classical ab initio molecular dynamics simulation and 35× faster than a full path integral calculation, while still exactly including the quantum sampling of nuclei. This development thus offers a route to routinely include nuclear quantum effects in ab initio molecular dynamics simulations at negligible computational cost.

Structure, dynamics, and reactivity of hydrated electrons by ab initio molecular dynamics.

Understanding the properties of hydrated electrons, which were first observed using pulse radiolysis of water in 1962, is crucial because they are key species in many radiation chemistry processes. Although time-resolved spectroscopic studies and molecular simulations have shown that an electron in water (prepared, for example, by water photoionization) relaxes quickly to a localized, cavity-like structure ∼2.5 Šin radius, this picture has recently been questioned. In another experimental approach, negatively charged water clusters of increasing size were studied with photoelectron and IR spectroscopies. Although small water clusters can bind an excess electron, their character is very different from bulk hydrated species. As data on electron binding in liquid water have become directly accessible experimentally, the cluster-to-bulk extrapolations have become a topic of lively debate. Quantum electronic structure calculations addressing experimental measurables have, until recently, been largely limited to small clusters; extended systems were approached mainly with pseudopotential calculations combining a classical description of water with a quantum mechanical treatment of the excess electron. In this Account, we discuss our investigations of electrons solvated in water by means of ab initio molecular dynamics simulations. This approach, applied to a model system of a negatively charged cluster of 32 water molecules, allows us to characterize structural, dynamical, and reactive aspects of the hydrated electron using all of the system's valence electrons. We show that under ambient conditions, the electron localizes into a cavity close to the surface of the liquid cluster. This cavity is, however, more flexible and accessible to water molecules than an analogous area around negatively charged ions. The dynamical process of electron attachment to a neutral water cluster is strongly temperature dependent. Under ambient conditions, the electron relaxes in the liquid cluster and becomes indistinguishable from an equilibrated, solvated electron on a picosecond time scale. In contrast, for solid, cryogenic systems, the electron only partially localizes outside of the cluster, being trapped in a metastable, weakly bound "cushion-like" state. Strongly bound states under cryogenic conditions could only be prepared by cooling equilibrated, liquid, negatively charged clusters. These calculations allow us to rationalize how different isomers of electrons in cryogenic clusters can be observed experimentally. Our results also bring into question the direct extrapolation of properties of cryogenic, negatively charged water clusters to those of electrons in the bulk liquid. Ab initio molecular dynamics represents a unique computational tool for investigating the reactivity of the solvated electron in water. As a prototype, the electron-proton reaction was followed in the 32-water cluster. In accord with experiment, the molecular mechanism is a proton transfer process that is not diffusion limited, but rather controlled by a proton-induced deformation of the excess electron's solvent shell. We demonstrate the necessary ingredients of a successful density functional methodology for the hydrated electron that avoids potential pitfalls, such as self-interaction error, insufficient basis set, or lack of dispersion interactions. We also benchmark the density functional theory methods and outline the path to faithful ab initio simulations of dynamics and reactivity of electrons solvated in extended aqueous systems.

Reducing the Cost of Neural Network Potential Generation for Reactive Molecular Systems.

Although machine learning potentials have recently had a substantial impact on molecular simulations, the construction of a robust training set can still become a limiting factor, especially due to the requirement of a reference ab initio simulation that covers all the relevant geometries of the system. Recognizing that this can be prohibitive for certain systems, we develop the method of transition tube sampling that mitigates the computational cost of training set and model generation. In this approach, we generate classical or quantum thermal geometries around a transition path describing a conformational change or a chemical reaction using only a sparse set of local normal mode expansions along this path and select from these geometries by an active learning protocol. This yields a training set with geometries that characterize the whole transition without the need for a costly reference trajectory. The performance of the method is evaluated on different molecular systems with the complexity of the potential energy landscape increasing from a single minimum to a double proton-transfer reaction with high barriers. Our results show that the method leads to training sets that give rise to models applicable in classical and path integral simulations alike that are on par with those based directly on ab initio calculations while providing the computational speedup we have come to expect from machine learning potentials.

Solvation and ion-pairing properties of the aqueous sulfate anion: explicit versus effective electronic polarization.

We assessed the relative merits of two approaches for including polarization effects in classical force fields for the sulfate anion. One of the approaches is the explicit shell model for atomic polarization and the other is an implicit dielectric continuum representation of the electronic polarization, wherein the polarizability density is spatially uniform. Both the solvation and ion association properties of sulfate were considered. We carried out an ab initio molecular dynamics simulation for a single sulfate anion in aqueous solution to obtain a benchmark for the solvation structure. For the ion-pairing properties, the models were compared to experimental thermodynamic data through Kirkwood-Buff theory, which relates the integrals of the pair correlation functions to measurable properties. While deficiencies were found for both of the approaches, the continuum polarization model was not systematically worse than the shell model. The shell model was found to give a more structured solution than the continuum polarization model, both with respect to solvation and ion pairing.

Peptide salt bridge stability: from gas phase via microhydration to bulk water simulations.

The salt bridge formation and stability in the terminated lysine-glutamate dipeptide is investigated in water clusters of increasing size up to the limit of bulk water. Proton transfer dynamics between the acidic and basic side chains is described by DFT-based Born-Oppenheimer molecular dynamics simulations. While the desolvated peptide prefers to be in its neutral state, already the addition of a single water molecule can trigger proton transfer from the glutamate side chain to the lysine side chain, leading to a zwitterionic salt bridge state. Upon adding more water molecules we find that stabilization of the zwitterionic state critically depends on the number of hydrogen bonds between side chain termini, the water molecules, and the peptidic backbone. Employing classical molecular dynamics simulations for larger clusters, we observed that the salt bridge is weakened upon additional hydration. Consequently, long-lived solvent shared ion pairs are observed for about 30 water molecules while solvent separated ion pairs are found when at least 40 or more water molecules hydrate the dipeptide. These results have implications for the formation and stability of salt bridges at partially dehydrated surfaces of aqueous proteins.

From a localized H3O radical to a delocalized H3O+···e- solvent-separated pair by sequential hydration.

The impact of microhydration on the electronic structure and reactivity of the H(3)O moiety is investigated by ab initio calculations. In the gas phase, H(3)O is a radical with spin density localized on its hydrogen end, which is only kinetically stable and readily decomposes into a water molecule and a hydrogen atom. When solvated by a single water molecule, H(3)O preserves to a large extent its radical character, however, two water molecules are already capable to shift most of the spin density to the solvent. With three solvating water molecules this shift is practically completed and the system is best described as a solvent-separated pair of a hydronium cation and a hydrated electron. The electronic structure of this system and its proton transfer reactivity leading to formation of a hydrogen atom already resemble those of a proton-electron pair in bulk water.

Arrhenius analysis of anisotropic surface self-diffusion on the prismatic facet of ice.

We present an Arrhenius analysis of self-diffusion on the prismatic surface of ice calculated from molecular dynamics simulations. The six-site water model of Nada and van der Eerden was used in combination with a structure-based criterion for determining the number of liquid-like molecules in the quasi-liquid layer. Simulated temperatures range from 230 K-287 K, the latter being just below the melting temperature of the model, 289 K. Calculated surface diffusion coefficients agree with available experimental data to within quoted precision. Our results indicate a positive Arrhenius curvature, implying a change in the mechanism of self-diffusion from low to high temperature, with a concomitant increase in energy of activation from 29.1 kJ mol(-1) at low temperature to 53.8 kJ mol(-1) close to the melting point. In addition, we find that the surface self-diffusion is anisotropic at lower temperatures, transitioning to isotropic in the temperature range of 240-250 K. We also present a framework for self-diffusion in the quasi-liquid layer on ice that aims to explain these observations.

Chasing charge localization and chemical reactivity following photoionization in liquid water.

The ultrafast dynamics of the cationic hole formed in bulk liquid water following ionization is investigated by ab initio molecular dynamics simulations and an experimentally accessible signature is suggested that might be tracked by femtosecond pump-probe spectroscopy. This is one of the fastest fundamental processes occurring in radiation-induced chemistry in aqueous systems and biological tissue. However, unlike the excess electron formed in the same process, the nature and time evolution of the cationic hole has been hitherto little studied. Simulations show that an initially partially delocalized cationic hole localizes within ~30 fs after which proton transfer to a neighboring water molecule proceeds practically immediately, leading to the formation of the OH radical and the hydronium cation in a reaction which can be formally written as H(2)O(+) + H(2)O → OH + H(3)O(+). The exact amount of initial spin delocalization is, however, somewhat method dependent, being realistically described by approximate density functional theory methods corrected for the self-interaction error. Localization, and then the evolving separation of spin and charge, changes the electronic structure of the radical center. This is manifested in the spectrum of electronic excitations which is calculated for the ensemble of ab initio molecular dynamics trajectories using a quantum mechanics/molecular mechanics (QM∕MM) formalism applying the equation of motion coupled-clusters method to the radical core. A clear spectroscopic signature is predicted by the theoretical model: as the hole transforms into a hydroxyl radical, a transient electronic absorption in the visible shifts to the blue, growing toward the near ultraviolet. Experimental evidence for this primary radiation-induced process is sought using femtosecond photoionization of liquid water excited with two photons at 11 eV. Transient absorption measurements carried out with ~40 fs time resolution and broadband spectral probing across the near-UV and visible are presented and direct comparisons with the theoretical simulations are made. Within the sensitivity and time resolution of the current measurement, a matching spectral signature is not detected. This result is used to place an upper limit on the absorption strength and/or lifetime of the localized H(2)O(+) ((aq)) species.

Persistent Homology Metrics Reveal Quantum Fluctuations and Reactive Atoms in Path Integral Dynamics.

Nuclear quantum effects (NQEs) are known to impact a number of features associated with chemical reactivity and physicochemical properties, particularly for light atoms and at low temperatures. In the imaginary time path integral formalism, each atom is mapped onto a "ring polymer" whose spread is related to the quantum mechanical uncertainty in the particle's position, i.e., its thermal wavelength. A number of metrics have previously been used to investigate and characterize this spread and explain effects arising from quantum delocalization, zero-point energy, and tunneling. Many of these shape metrics consider just the instantaneous structure of the ring polymers. However, given the significant interest in methods such as centroid molecular dynamics and ring polymer molecular dynamics that link the molecular dynamics of these ring polymers to real time properties, there exists significant opportunity to exploit metrics that also allow for the study of the fluctuations of the atom delocalization in time. Here we consider the ring polymer delocalization from the perspective of computational topology, specifically persistent homology, which describes the 3-dimensional arrangement of point cloud data, (i.e. atomic positions). We employ the Betti sequence probability distribution to define the ensemble of shapes adopted by the ring polymer. The Wasserstein distances of Betti sequences adjacent in time are used to characterize fluctuations in shape, where the Fourier transform and associated principal components provides added information differentiating atoms with different NQEs based on their dynamic properties. We demonstrate this methodology on two representative systems, a glassy system consisting of two atom types with dramatically different de Broglie thermal wavelengths, and ab initio molecular dynamics simulation of an aqueous 4 M HCl solution where the H-atoms are differentiated based on their participation in proton transfer reactions.

Significance Of Nuclear Quantum Effects In Hydrogen Bonded Molecular Chains.

In hydrogen-bonded systems, nuclear quantum effects such as zero-point motion and tunneling can significantly affect their material properties through underlying physical and chemical processes. Presently, direct observation of the influence of nuclear quantum effects on the strength of hydrogen bonds with resulting structural and electronic implications remains elusive, leaving opportunities for deeper understanding to harness their fascinating properties. We studied hydrogen-bonded one-dimensional quinonediimine molecular networks which may adopt two isomeric electronic configurations via proton transfer. Herein, we demonstrate that concerted proton transfer promotes a delocalization of π-electrons along the molecular chain, which enhances the cohesive energy between molecular units, increasing the mechanical stability of the chain and giving rise to distinctive electronic in-gap states localized at the ends. These findings demonstrate the identification of a class of isomeric hydrogen-bonded molecular systems where nuclear quantum effects play a dominant role in establishing their chemical and physical properties. This identification is a step toward the control of mechanical and electronic properties of low-dimensional molecular materials via concerted proton tunneling.

Dynamics of electron localization in warm versus cold water clusters.

The process of electron localization on a cluster of 32 water molecules at 20, 50, and 300 K is unraveled using ab initio molecular dynamics simulations. In warm, liquid clusters, the excess electron relaxes from an initial diffuse and weakly bound structure to an equilibrated, strongly bound species within 1.5 ps. In contrast, in cold, glassy clusters the relaxation processes is not completed and the electron becomes trapped in a metastable surface state with an intermediate binding energy. These results question the validity of extrapolations of the properties of solvated electrons from cold clusters of increasing size to the liquid bulk.

Benzene Radical Anion in the Context of the Birch Reduction: When Solvation Is the Key.

The benzene radical anion is an important intermediate in the Birch reduction of benzene by solvated electrons in liquid ammonia. Beyond organic chemistry, it is an intriguing subject of spectroscopic and theoretical studies due to its rich structural and dynamical behavior. In the gas phase, the species appears as a metastable shape resonance, while in the condensed phase, it remains stable. Here, we approach the system by ab initio molecular dynamics in liquid ammonia and demonstrate that the inclusion of solvent is crucial and indeed leads to stability. Beyond the mere existence of the radical anion species, our simulations explore its condensed-phase behavior at the molecular level and offer new insights into its properties. These include the dynamic Jahn-Teller distortions, vibrational spectra in liquid ammonia, and the structure of the solvent shell, including the motif of a π-hydrogen bond between ammonia molecules and the aromatic ring.