Midpoint cell method for hybrid (MPI+OpenMP) parallelization of molecular dynamics simulations

We have developed a new hybrid (MPI+OpenMP) parallelization scheme for molecular dynamics (MD) simulations by combining a cell‐wise version of the midpoint method with pair‐wise Verlet lists. In this scheme, which we call the midpoint cell method, simulation space is divided into subdomains, each of which is assigned to a MPI processor. Each subdomain is further divided into small cells. The interaction between two particles existing in different cells is computed in the subdomain containing the midpoint cell of the two cells where the particles reside. In each MPI processor, cell pairs are distributed over OpenMP threads for shared memory parallelization. The midpoint cell method keeps the advantages of the original midpoint method, while filtering out unnecessary calculations of midpoint checking for all the particle pairs by single midpoint cell determination prior to MD simulations. Distributing cell pairs over OpenMP threads allows for more efficient shared memory parallelization compared with distributing atom indices over threads. Furthermore, cell grouping of particle data makes better memory access, reducing the number of cache misses. The parallel performance of the midpoint cell method on the K computer showed scalability up to 512 and 32,768 cores for systems of 20,000 and 1 million atoms, respectively. One MD time step for long‐range interactions could be calculated within 4.5 ms even for a 1 million atoms system with particle‐mesh Ewald electrostatics. © 2014 Wiley Periodicals, Inc.     

The implementation of a fast and accurate QM/MM potential method in Amber

Version 9 of the Amber simulation programs includes a new semi-empirical hybrid QM/MM functionality. This includes support for implicit solvent (generalized Born) and for periodic explicit solvent simulations using a newly developed QM/MM implementation of the particle mesh Ewald (PME) method. The code provides sufficiently accurate gradients to run constant energy QM/MM MD simulations for many nanoseconds. The link atom approach used for treating the QM/MM boundary shows improved performance, and the user interface has been rewritten to bring the format into line with classical MD simulations. Support is provided for the PM3, PDDG/PM3, PM3CARB1, AM1, MNDO, and PDDG/MNDO semi-empirical Hamiltonians as well as the self-consistent charge density functional tight binding (SCC-DFTB) method. Performance has been improved to the point where using QM/MM, for a QM system of 71 atoms within an explicitly solvated protein using periodic boundaries and PME requires less than twice the cpu time of the corresponding classical simulation.     

Interactive quantum chemistry: a divide-and-conquer ASED-MO method

We present interactive quantum chemistry simulation at the atom superposition and electron delocalization molecular orbital (ASED-MO) level of theory. Our method is based on the divide-and-conquer (D&C) approach, which we show is accurate and efficient for this non-self-consistent semiempirical theory. The method has a linear complexity in the number of atoms, scales well with the number of cores, and has a small prefactor. The time cost is completely controllable, as all steps are performed with direct algorithms, i.e., no iterative schemes are used. We discuss the errors induced by the D&C approach, first empirically on a few examples, and then via a theoretical study of two toy models that can be analytically solved for any number of atoms. Thanks to the precision and speed of the D&C approach, we are able to demonstrate interactive quantum chemistry simulations for systems up to a few hundred atoms on a current multicore desktop computer. When drawing and editing molecular systems, interactive simulations provide immediate, intuitive feedback on chemical structures. As the number of cores on personal computers increases, and larger and larger systems can be dealt with, we believe such interactive simulations-even at lower levels of theory-should thus prove most useful to effectively understand, design and prototype molecules, devices and materials.     

Record Atomistic Simulation of Crystalline Silicon: Bridging Microscale Structures and Macroscale Properties

Based on the molecular dynamics software package CovalentMD 2.0, the fastest molecular dynamics simulation for covalent crystalline silicon with bond-order potentials has been implemented on the third highest performance supercomputer “Sunway TaihuLight” in the world (before June 2019), and already obtained 16.0 Pflops (10¹⁵ floating point operation per second) in double precision for the simulation of crystalline silicon, which is recordly high for rigorous atomistic simulation of covalent materials. The simulations used up to 160,768 64-core processors, totally nearly 10.3 million cores, to simulate more than 137 billion silicon atoms, where the parallel efficiency is over 80% on the whole machine. The running performance on a single processor reached 15.1% of its theoretical peak at highest. The longitudinal dimension of the simulated system is far beyond the range with scale-dependent properties, while the lateral dimension significantly exceeds the experimentally measurable range. Our simulation enables virtual experiments on real-world nanostructured materials and devices for predicting macroscale properties and behaviors from microscale structures directly, bringing about many exciting new possibilities in nanotechnology, information technology, electronics and renewable energies, etc. © 2019 Wiley Periodicals, Inc.     

Efficient lookup table using a linear function of inverse distance squared

The major bottleneck in molecular dynamics (MD) simulations of biomolecules exist in the calculation of pairwise nonbonded interactions like Lennard‐Jones and long‐range electrostatic interactions. Particle‐mesh Ewald (PME) method is able to evaluate long‐range electrostatic interactions accurately and quickly during MD simulation. However, the evaluation of energy and gradient includes time‐consuming inverse square roots and complementary error functions. To avoid such time‐consuming operations while keeping accuracy, we propose a new lookup table for short‐range interaction in PME by defining energy and gradient as a linear function of inverse distance squared. In our lookup table approach, densities of table points are inversely proportional to squared pair distances, enabling accurate evaluation of energy and gradient at small pair distances. Regardless of the inverse operation here, the new lookup table scheme allows fast pairwise nonbonded calculations owing to efficient usage of cache memory. © 2013 Wiley Periodicals, Inc.     

Liquid Structure of 1-Propanol by Molecular Dynamics Simulations and X-Ray Scattering

Molecular dynamics (MD) simulations and X-ray scattering experiments have been carried out on liquid 1-propanol. The radial distribution functions obtained from these two methods were in good agreement with each other. On the basis of the hydrogen-bond number and the angular correlation functions of the four sequentially hydrogen-bonded O atoms derived from the MD calculation, it was found that the hydrogen-bonded O atoms preferentially form a planar zigzag chain structure, but that the plane undulates like a wave.     

Theoretical study of the stability of lithium atoms in alpha-rhombohedral boron

The stability of lithium atoms in alpha-rhombohedral boron was investigated by first-principles calculations of total energies and molecular dynamics (MD) simulations. In the case of a low concentration (1.03 at. %), Li at the center of the icosahedral B12 site (the I-site) had a negative binding energy, which suggests Li at the I-site is unstable. However, MD simulations at temperatures below 750 K indicated that Li is still confined in the B12 cage under these conditions, which means Li at the I-site is metastable. Over 800 K, Li began to move away from the B12 site and settled at the tetrahedral site (the T-site) or at the octahedral site (the O-site). Li at the T-site also had a negative binding energy, but MD simulations indicated it was metastable up to 1400 K and did not move to other sites. Li at the O-site was energetically the most favorable, having a positive binding energy. In the case of a high concentration (7.69 at. %), the I-site changed to an unstable saddle point. At this concentration, the T-site was metastable and the O-site became the most stable. In MD simulations at 1400 K, Li atoms at the O-site never jumped to other sites regardless of concentration. Considering these facts, the diffusion coefficient of Li in alpha-rhombohedral boron would have to be very small below 1400 K.     

An N log N approximation based on the natural organization of biomolecules for speeding up the computation of long range interactions

Presented here is a method, the hierarchical charge partitioning (HCP) approximation, for speeding up computation of pairwise electrostatic interactions in biomolecular systems. The approximation is based on multiple levels of natural partitioning of biomolecular structures into a hierarchical set of its constituent structural components. The charge distribution in each component is systematically approximated by a small number of point charges, which, for the highest level component, are much fewer than the number of atoms in the component. For short distances from the point of interest, the HCP uses the full set of atomic charges available. For long‐distance interactions, the approximate charge distributions with smaller sets of charges are used instead. For a structure consisting of N charges, the computational cost of computing the pairwise interactions via the HCP scales as O(N log N), under assumptions about the structural organization of biomolecular structures generally consistent with reality. A proof‐of‐concept implementation of the HCP shows that for large structures it can lead to speed‐up factors of up to several orders of magnitude relative to the exact pairwise O(N²) all‐atom computation used as a reference. For structures with more than 2000–3000 atoms the relative accuracy of the HCP (relative root‐mean‐square force error per atom), approaches the accuracy of the particle mesh Ewald (PME) method with parameter settings typical for biomolecular simulations. When averaged over a set of 600 representative biomolecular structures, the relative accuracies of the two methods are roughly equal. The HCP is also significantly more accurate than the spherical cutoff method. The HCP has been implemented in the freely available nucleic acids builder (NAB) molecular dynamics (MD) package in Amber tools. A 10 ns simulation of a small protein indicates that the HCP based MD simulation is stable, and that it can be faster than the spherical cutoff method. A critical benefit of the HCP approximation is that it is algorithmically very simple, and unlike the PME, the HCP is straightforward to use with implicit solvent models. © 2009 Wiley Periodicals, Inc. J Comput Chem, 2010     

Message passing interface and multithreading hybrid for parallel molecular docking of large databases on petascale high performance computing machines

A mixed parallel scheme that combines message passing interface (MPI) and multithreading was implemented in the AutoDock Vina molecular docking program. The resulting program, named VinaLC, was tested on the petascale high performance computing (HPC) machines at Lawrence Livermore National Laboratory. To exploit the typical cluster‐type supercomputers, thousands of docking calculations were dispatched by the master process to run simultaneously on thousands of slave processes, where each docking calculation takes one slave process on one node, and within the node each docking calculation runs via multithreading on multiple CPU cores and shared memory. Input and output of the program and the data handling within the program were carefully designed to deal with large databases and ultimately achieve HPC on a large number of CPU cores. Parallel performance analysis of the VinaLC program shows that the code scales up to more than 15K CPUs with a very low overhead cost of 3.94%. One million flexible compound docking calculations took only 1.4 h to finish on about 15K CPUs. The docking accuracy of VinaLC has been validated against the DUD data set by the re‐docking of X‐ray ligands and an enrichment study, 64.4% of the top scoring poses have RMSD values under 2.0 Å. The program has been demonstrated to have good enrichment performance on 70% of the targets in the DUD data set. An analysis of the enrichment factors calculated at various percentages of the screening database indicates VinaLC has very good early recovery of actives. © 2013 Wiley Periodicals, Inc.     

Inherent speedup limitations in multiple time step/particle mesh Ewald algorithms

Multiple time step (MTS) algorithms present an effective integration approach to reduce the computational cost of dynamics simulations. By using force splitting to allow larger time steps for the more slowly varying force components, computational savings can be realized. The Particle-Mesh-Ewald (PME) method has been independently devised to provide an effective and efficient treatment of the long-range electrostatics interactions. Here we examine the performance of a combined MTS/PME algorithm previously developed for AMBER on a large polymerase beta/DNA complex containing 40,673 atoms. Our goal is to carefully combine the robust features of the Langevin/MTS (LN) methodology implemented in CHARMM-which uses position rather than velocity Verlet with stochasticity to make possible outer time steps of 150 fs-with the PME formulation. The developed MTS/PME integrator removes fast terms from the reciprocal-space Ewald component by using switch functions. We analyze the advantages and limitations of the resulting scheme by comparing performance to the single time step leapfrog Verlet integrator currently used in AMBER by evaluating different time-step protocols using three assessors for accuracy, speedup, and stability, all applied to long (i.e., nanosecond) simulations to ensure proper energy conservation. We also examine the performance of the algorithm on a parallel, distributed shared-memory computer (SGI Origin 2000 with 8 300-MHz R12000 processors). Good energy conservation and stability behavior can be demonstrated, for Newtonian protocols with outer time steps of up to 8 fs and Langevin protocols with outer time steps of up to 16 fs. Still, we emphasize the inherent limitations imposed by the incorporation of MTS methods into the PME formulation that may not be widely appreciated. Namely, the limiting factor on the largest outer time-step size, and hence speedup, is an intramolecular cancellation error inherent to PME. This error stems from the excluded-nonbonded correction term contained in the reciprocal-space component. This cancellation error varies in time and introduces artificial frequencies to the governing dynamics motion. Unfortunately, we find that this numerical PME error cannot be easily eliminated by refining the PME parameters (grid resolution and/or order of interpolating polynomial). We suggest that methods other than PME for fast electrostatics may allow users to reap the full advantages from MTS algorithms.     

Routine Microsecond Molecular Dynamics Simulations with AMBER on GPUs. 1. Generalized Born

We present an implementation of generalized Born implicit solvent all-atom classical molecular dynamics (MD) within the AMBER program package that runs entirely on CUDA enabled NVIDIA graphics processing units (GPUs). We discuss the algorithms that are used to exploit the processing power of the GPUs and show the performance that can be achieved in comparison to simulations on conventional CPU clusters. The implementation supports three different precision models in which the contributions to the forces are calculated in single precision floating point arithmetic but accumulated in double precision (SPDP), or everything is computed in single precision (SPSP) or double precision (DPDP). In addition to performance, we have focused on understanding the implications of the different precision models on the outcome of implicit solvent MD simulations. We show results for a range of tests including the accuracy of single point force evaluations and energy conservation as well as structural properties pertainining to protein dynamics. The numerical noise due to rounding errors within the SPSP precision model is sufficiently large to lead to an accumulation of errors which can result in unphysical trajectories for long time scale simulations. We recommend the use of the mixed-precision SPDP model since the numerical results obtained are comparable with those of the full double precision DPDP model and the reference double precision CPU implementation but at significantly reduced computational cost. Our implementation provides performance for GB simulations on a single desktop that is on par with, and in some cases exceeds, that of traditional supercomputers.     

Shared-memory parallelization of the TURBOMOLE programs AOFORCE, ESCF, and EGRAD: how to quickly parallelize legacy code

The programs ESCF, EGRAD, and AOFORCE are parts of the TURBOMOLE program package and compute excited-state properties and ground-state geometric hessians, respectively, for Hartree-Fock and density functional methods. The range of applicability of these programs has been extended by allowing them to use all CPU cores on a given node in parallel. The parallelization strategy is not new and duplicates what is standard today in the calculation of ground-state energies and gradients. The focus is on how this can be achieved without needing extensive modifications of the existing serial code. The key ingredient is to fork off worker processes with separated address spaces as they are needed. Test calculations on a molecule with about 80 atoms and 1000 basis functions show good parallel speedup up to 32 CPU cores.     

Molecular nitrogen-N2 properties: the intermolecular potential and the equation of state

Quantum mechanical (QM) high precision calculations were used to determine N(2)-N(2) intermolecular interaction potential. Using QM numerical data the anisotropic potential energy surface was obtained for all orientations of the pair of the nitrogen molecules in the rotation invariant form. The new N(2)-N(2) potential is in reasonably good agreement with the scaled potential obtained by van der Avoird et al. using the results of Hartree-Fock calculations [J. Chem. Phys. 84, 1629 (1986)]. The molecular dynamics (MD) of the N(2) molecules has been used to determine nitrogen equation of state. The classical motion of N(2) molecules was integrated in rigid rotor approximation, i.e., it accounted only translational and rotational degrees of freedom. Fincham [Mol. Simul. 11, 79 (1993)] algorithm was shown to be superior in terms of precision and energy stability to other algorithms, including Singer [Mol. Phys. 33, 1757 (1977)], fifth order predictor-corrector, or Runge-Kutta, and was therefore used in the MD modeling of the nitrogen pressure [S. Krukowski and P. Strak, J. Chem. Phys. 124, 134501 (2006)]. Nitrogen equation of state at pressures up to 30 GPa (300 kbars) and temperatures from the room temperature to 2000 K was obtained using MD simulation results. Results of MD simulations are in very good agreement (the error below 1%) with the experimental data on nitrogen equation of state at pressures below 1 GPa (10 kbars) for temperatures below 1800 K [R. T. Jacobsen et al., J. Phys. Chem. Ref. Data 15, 735 (1986)]. For higher temperatures, the deviation is slightly larger, about 2.5% which still is a very good agreement. The slightly larger difference may be attributed to the vibrational motion not accounted explicitly by rigid rotor approximation, which may be especially important at high temperatures. These results allow to obtain reliable equation of state of nitrogen for pressures up to 30 GPa (300 kbars), i.e., close to molecular nitrogen stability limit, determined by Nellis et al. [Phys. Rev. Lett. 53, 1661 (1984)].     

Computation of binding free energy with molecular dynamics and grand canonical Monte Carlo simulations

The binding of a ligand to a receptor is often associated with the displacement of a number of bound water molecules. When the binding site is exposed to the bulk region, this process may be sampled adequately by standard unbiased molecular dynamics trajectories. However, when the binding site is deeply buried and the exchange of water molecules with the bulk region may be difficult to sample, the convergence and accuracy in free energy perturbation (FEP) calculations can be severely compromised. These problems are further compounded when a reduced system including only the region surrounding the binding site is simulated. To address these issues, we couple molecular dynamics (MD) with grand canonical Monte Carlo (GCMC) simulations to allow the number of water to fluctuate during an alchemical FEP calculation. The atoms in a spherical inner region around the binding pocket are treated explicitly while the influence of the outer region is approximated using the generalized solvent boundary potential (GSBP). At each step during thermodynamic integration, the number of water in the inner region is equilibrated with GCMC and energy data generated with MD is collected. Free energy calculations on camphor binding to a deeply buried pocket in cytochrome P450cam, which causes about seven water molecules to be expelled, are used to test the method. It concluded that solvation free energy calculations with the GCMC/MD method can greatly improve the accuracy of the computed binding free energy compared to simulations with fixed number of water.     

MOIL-opt: Energy-Conserving Molecular Dynamics on a GPU/CPU system

We report an optimized version of the molecular dynamics program MOIL that runs on a shared memory system with OpenMP and exploits the power of a Graphics Processing Unit (GPU). The model is of heterogeneous computing system on a single node with several cores sharing the same memory and a GPU. This is a typical laboratory tool, which provides excellent performance at minimal cost. Besides performance, emphasis is made on accuracy and stability of the algorithm probed by energy conservation for explicit-solvent atomically-detailed-models. Especially for long simulations energy conservation is critical due to the phenomenon known as "energy drift" in which energy errors accumulate linearly as a function of simulation time. To achieve long time dynamics with acceptable accuracy the drift must be particularly small. We identify several means of controlling long-time numerical accuracy while maintaining excellent speedup. To maintain a high level of energy conservation SHAKE and the Ewald reciprocal summation are run in double precision. Double precision summation of real-space non-bonded interactions improves energy conservation. In our best option, the energy drift using 1fs for a time step while constraining the distances of all bonds, is undetectable in 10ns simulation of solvated DHFR (Dihydrofolate reductase). Faster options, shaking only bonds with hydrogen atoms, are also very well behaved and have drifts of less than 1kcal/mol per nanosecond of the same system. CPU/GPU implementations require changes in programming models. We consider the use of a list of neighbors and quadratic versus linear interpolation in lookup tables of different sizes. Quadratic interpolation with a smaller number of grid points is faster than linear lookup tables (with finer representation) without loss of accuracy. Atomic neighbor lists were found most efficient. Typical speedups are about a factor of 10 compared to a single-core single-precision code.     

Dissolution of solid NaCl nanoparticles embedded in supersaturated water vapor probed by molecular dynamic simulations

The dissolution process for small, on the order of 1000 atoms, crystalline NaCl particles with defects embedded in highly supersaturated water vapor was studied by the molecular dynamics (MD) simulation method. We found that a breakdown of the crystal lattice does not occur unless (1) the thickness of water layer covering the surface of salt particles exceeds several molecular layers and (2) there are a considerable number of defects in the crystal. The collapse of the crystal lattice starts when the amount of water taken up by a salt particle reaches about half ( approximately 50%) of the amount of salt in this particle. The number of defects required to initiate subsequent dissolution of the NaCl crystal on the time scale accessible by our simulations ( approximately 40 ns) is in the range of 10 to 12%. We also report the estimates for the time required to form supersaturated aqueous solutions of NaCl from originally crystalline particles as a function of the number of defects in the crystal.     

Adaptive partitioning in combined quantum mechanical and molecular mechanical calculations of potential energy functions for multiscale simulations

In many applications of multilevel/multiscale methods, an active zone must be modeled by a high-level electronic structure method, while a larger environmental zone can be safely modeled by a lower-level electronic structure method, molecular mechanics, or an analytic potential energy function. In some cases though, the active zone must be redefined as a function of simulation time. Examples include a reactive moiety diffusing through a liquid or solid, a dislocation propagating through a material, or solvent molecules in a second coordination sphere (which is environmental) exchanging with solvent molecules in an active first coordination shell. In this article, we present a procedure for combining the levels smoothly and efficiently in such systems in which atoms or groups of atoms move between high-level and low-level zones. The method dynamically partitions the system into the high-level and low-level zones and, unlike previous algorithms, removes all discontinuities in the potential energy and force whenever atoms or groups of atoms cross boundaries and change zones. The new adaptive partitioning (AP) method is compared to Rode's "hot spot" method and Morokuma's "ONIOM-XS" method that were designed for multilevel molecular dynamics (MD) simulations. MD simulations in the microcanonical ensemble show that the AP method conserves both total energy and momentum, while the ONIOM-XS method fails to conserve total energy and the hot spot method fails to conserve both total energy and momentum. Two versions of the AP method are presented, one scaling as O(2N) and one with linear scaling in N, where N is the number of groups in a buffer zone separating the active high-level zone from the environmental low-level zone. The AP method is also extended to systems with multiple high-level zones to allow, for example, the study of ions and counterions in solution using the multilevel approach.     

Development of hardware accelerator for molecular dynamics simulations: a computation board that calculates nonbonded interactions in cooperation with fast multipole method

Evaluation of long-range Coulombic interactions still represents a bottleneck in the molecular dynamics (MD) simulations of biological macromolecules. Despite the advent of sophisticated fast algorithms, such as the fast multipole method (FMM), accurate simulations still demand a great amount of computation time due to the accuracy/speed trade-off inherently involved in these algorithms. Unless higher order multipole expansions, which are extremely expensive to evaluate, are employed, a large amount of the execution time is still spent in directly calculating particle-particle interactions within the nearby region of each particle. To reduce this execution time for pair interactions, we developed a computation unit (board), called MD-Engine II, that calculates nonbonded pairwise interactions using a specially designed hardware. Four custom arithmetic-processors and a processor for memory manipulation ("particle processor") are mounted on the computation board. The arithmetic processors are responsible for calculation of the pair interactions. The particle processor plays a central role in realizing efficient cooperation with the FMM. The results of a series of 50-ps MD simulations of a protein-water system (50,764 atoms) indicated that a more stringent setting of accuracy in FMM computation, compared with those previously reported, was required for accurate simulations over long time periods. Such a level of accuracy was efficiently achieved using the cooperative calculations of the FMM and MD-Engine II. On an Alpha 21264 PC, the FMM computation at a moderate but tolerable level of accuracy was accelerated by a factor of 16.0 using three boards. At a high level of accuracy, the cooperative calculation achieved a 22.7-fold acceleration over the corresponding conventional FMM calculation. In the cooperative calculations of the FMM and MD-Engine II, it was possible to achieve more accurate computation at a comparable execution time by incorporating larger nearby regions.     

Data parallel large-scale molecular dynamics for liquids

An efficient data parallel computational scheme is presented for large-scale molecular dynamics (MD ) simulations of liquids with short-range interactions. The method is based on decomposition of the simulation cell into equally sized subcells, with the shortest side length equal to the cutoff radius. Inter- and intracell interactions are calculated in a coarse-grained manner. A geometric sorting procedure, based on particle distances to subcell boundaries, is used to minimize the overall computations and the nonproductive communications. Using only nearest-neighbor communications, an efficient scheme is developed for periodic updates of the contents of subcells due to the migration of particles. Special “null-particles” are introduced, which act as buffers during the periodic updates and allow for a globally uniform algorithm during the calculations. Communication cost is about 7% of the total CPU time. The method is found to be linearly scalable with the number of particles, performing better as the ratio of virtual to physical processors increases. The MD code is written in Fortran 90 and implemented on a CM-200. The overall speed is approximately 5.9 μs. per MD step and per particle for 1 million particles and 5.5 μs for 5 million particles. The method should be easily transferred to other massively parallel computers of SIMD and MIMD type. © 1993 John Wiley & Sons, Inc.     

Isomerization dynamics and thermodynamics of ionic argon clusters

The dynamics and thermodynamics of small Ar(n) (+) clusters, n=3, 6, and 9, are investigated using molecular dynamics (MD) and exchange Monte Carlo (MC) simulations. A diatomic-in-molecule Hamiltonian provides an accurate model for the electronic ground state potential energy surface. The microcanonical caloric curves calculated from MD and MC methods are shown to agree with each other, provided that the rigorous conservation of angular momentum is accounted for in the phase space density of the MC simulations. The previously proposed projective partition of the kinetic energy is used to assist MD simulations in interpreting the cluster dynamics in terms of inertial, internal, and external modes. The thermal behavior is correlated with the nature of the charged core in the cluster by computing a dedicated charge localization order parameter. We also perform systematic quenches to establish a connection with the various isomers. We find that the Ar(3) (+) cluster is very stable in its linear ground state geometry up to about 300 K, and then isomerizes to a T-shaped isomer in which a quasineutral atom lies around a charged dimer. In Ar(6) (+) and Ar(9) (+), the covalent trimer core is solvated by neutral atoms, and the weakly bound solvent shell melts at much lower energies, occasionally leading to a tetramer or pentamer core with weakly charged extremities. At high energies the core itself becomes metastable and the cluster transforms into Ar(2) (+) solvated by a fluid of neutral argon atoms.