WorksPhysics and Astronomy
Dynamics and Statistical Mechanics of Born-Oppenheimer Auxiliary Velocities
Abstract
In Born-Oppenheimer (BO) dynamics, auxiliary variables such as electron densities, induced polarization and electrode charges follow stationarity conditions, yet their time derivatives, such as currents, are physical observables. Using Mass-Zero constrained dynamics, we interpret these derivatives as explicit dynamical variables and derive the invariant measure giving their equilibrium statistical properties. We illustrate the framework in constant-potential molecular dynamics, where the electrode-charge velocities reproduce BO current fluctuations and the frequency-dependent admittance.
Introduction
In molecular dynamics, many physically relevant quantities are fixed by instantaneous stationarity conditions rather than by their own equations of motion. Examples include electron densities in orbital-free [1] and Kohn-Sham electronic-structure dynamics [2], induced polarization in classical force fields [3, 4, 5, 6], and electrode charges in constant-potential simulations [7]. In the Born–Oppenheimer (BO) description these auxiliary variables are functions of the evolving physical coordinates. In particular, their time-derivatives (referred to as “auxiliary velocities” hereafter) follow from the chain rule, but as derived quantities, without a statistical-mechanical framework of their own. These derivatives, however, correspond to physical observables. For example, polarization and electrode currents determine transport and, through fluctuation–dissipation relations, frequency-dependent response [8, 9, 10, 11, 12, 13]. A consistent dynamical and statistical-mechanical description of these derivatives is therefore interesting and useful.
A dynamical framework is introduced in extended-Lagrangian methods that propagate auxiliary variables with fictitious inertia, as in Car–Parrinello dynamics [2, 14], or combine auxiliary dynamics with approximate self-consistent updates [15, 16, 17, 18]. Finite values of the fictitious electronic mass, however, are known to introduce systematic deviations from Born–Oppenheimer dynamics [19, 20], including effective renormalization of the ionic masses and associated biases in dynamical observables such as vibrational frequencies [21, 22]. Mass-Zero (MaZe) constrained dynamics [23, 24] removes these spurious effects via an extended dynamical system that enforces a rigorous null inertia limit and imposes the stationarity conditions as holonomic constraints throughout the propagation. MaZe has been applied to polarizable models [25], orbital-free density functional theory [26], shell models [27] and constant-potential electrodes [24], and its mass-zero limit is proven to reproduce BO dynamics and equilibrium statistics for the physical variables. Here we focus on the auxiliary variables.
We show that (i) in Hamiltonian variables the auxiliary momenta and the constraint multipliers scale linearly with the inertia, so that the mass-zero limit is regular and the auxiliary velocities remain finite and explicit; (ii) these velocities coincide exactly with the time derivatives of the BO stationary solutions along the same physical trajectory; and (iii) the phase-space compressibility and invariant measure of the constrained flow provide the statistical description needed for equilibrium averages and time-correlation functions of the velocities. We illustrate the framework on constant-potential electrodes, where the auxiliary velocities are electrical currents, by comparing MaZe with a BO reference for the current, its autocorrelation, and the frequency-dependent admittance.
Theory We consider a system of physical particles with coordinates , conjugate momenta , and masses collected in the diagonal matrix , coupled to auxiliary variables through a (possibly generalised) potential . In the Born–Oppenheimer (BO) description, the auxiliary variables are determined at each physical configuration by the stationarity conditions
We assume that these conditions define a smooth, locally unique solution so that the corresponding Jacobian with respect to the auxiliary variables, , is nonsingular. MaZe dynamics enforces (1) as a set of holonomic constraints while propagating the auxiliary coordinates. To formulate the dynamics in Hamiltonian variables, we initially assign a finite inertia to the auxiliary variables and introduce their conjugate momenta . The extended Hamiltonian is
The constrained equations of motion are
where the -dimensional denotes the Lagrange multipliers and we have introduced the matrix
The auxiliary potential forces vanish on the constraint manifold because (see Eq. (1)).
Preservation of the stationarity conditions requires the velocity constraint
This relation gives
Thus, the auxiliary momenta scale linearly with , whereas the auxiliary velocities remain finite for all values of the fictitious inertia.
Differentiating the constraints once more and using Eqs. (3), we obtain a linear equation for the multipliers, whose solution can be written as
where
with and , which is finite for by Eq. (5) (). For nonsingular and finite constrained velocities, remains invertible and remains finite as . Consequently, the constraint multipliers also scale linearly with the auxiliary inertia.
Taking the mass-zero limit while retaining the scaled momenta gives
subject to the position and velocity constraints and where is the zero inertia Lagrange multiplier. Since , the equations for the physical variables reduce to
These are the BO equations of motion. For identical initial physical conditions, MaZe and BO therefore generate the same physical trajectories in the mass-zero limit.
We now establish the correspondence for the auxiliary velocities. In BO dynamics, the stationary solution follows the physical trajectory without possessing an independent equation of motion. Its time derivative is determined by the chain rule,
Differentiating the identity with respect to yields , and therefore
Comparison with Eq. (5) shows that the time derivative of the BO stationary solution coincides with the constrained auxiliary velocity,
when evaluated along the same physical trajectory. This identity follows from the geometry of the stationary manifold and holds for any finite . At finite inertia, the constrained and BO physical trajectories generally differ because of the constraint force acting on the physical momenta. In the mass-zero limit, this force vanishes and the physical trajectories coincide, so the correspondence extends to the time evolution as well.
The identity is a direct consequence of the geometry of the stationary manifold (trivial, in that sense), yet it is useful: in MaZe, is a regular phase-space variable of a well-defined dynamical system, produced directly by the constrained integrator detailed in Figure 1, without differentiating the stationary solution [28]. The equilibrium fluctuations of the auxiliary coordinates and velocities, require a consistent statistical description of the extended constrained dynamics. This can be conveniently obtained within the formalism of non-Hamiltonian statistical mechanics [29, 30]. Eliminating the constraint multipliers generally leads to a compressible flow in the ambient phase-space variables . As shown in the Supplementary Material (SM) [31], its compressibility is

Figure 1. Integration cycle of the mass-zero constrained dynamics. Starting from , a velocity half-step is followed by the SHAKE update enforcing ; after recomputing the forces, the RATTLE projection enforces [Eq. (4)].
where denotes the determinant of . The corresponding invariant metric factor can be chosen as
up to a constant factor, which cancels in normalized averages. The constrained equilibrium average of an observable is then
where is the normalization factor, is the stationary density of the extended dynamics, the delta functions enforce the constraints and the velocity constraints, and is the observable.
To obtain a regular statistical description in the mass-zero limit, we introduce the scaled auxiliary momenta . In these variables, the velocity constraint becomes
with no explicit dependence on . The momentum Jacobian cancels between numerator and denominator of the normalized average, and the Hamiltonian becomes , with the regular limit . Since , the limiting average is
where is the corresponding normalization and is the limiting stationary density. Marginalization over and , shows that the physical variables are distributed according to , the BO density for the chosen ensemble. This recovers the known BO limit of the constrained measure [23, 24]. The derivation of this limit in the present variables is given in the SM [31].
In the full extended system, the velocity is slaved to , but remains an explicit phase-space variable. Together with the dynamical correspondence above, the measure allows equilibrium time-correlation functions and linear-response relations to be formulated directly for observables involving the auxiliary velocities [32].
Application to constant-potential electrodes
We illustrate the correspondence using the fluctuating-charge model of constant-potential electrodes, represented in Figure 2. The system is an electrolyte of ions, whose positions are the physical degrees of freedom, confined between two metallic electrodes, labelled and , each composed of fixed atoms. Every electrode atom carries a Gaussian charge distribution of fixed width and fluctuating magnitude (see SM [31]), and the sites of each electrode are held at an imposed potential ( and for the two electrodes, as indicated in the scheme). The electrode-site charges, which adjust to the ionic configuration so as to keep each site at its imposed potential, are the auxiliary coordinates, with an additional auxiliary variable needed to enforce global electroneutrality:

Figure 2. Snapshot of molten LiCl (Li blue, Cl red) between two Al electrodes (grey) held at constant potentials and , with periodic boundary conditions in and . The electrode charges fluctuate under the constant-potential and electroneutrality conditions.
The physical coordinates evolve based on the negative
gradient of
where denotes the non-Coulomb interactions of the electrolyte, is the electrostatic energy of the electrolyte and electrode charges (specified in the SM [31]). The third term in Eq. (19) is the reversible work to charge the electrodes, while the fourth imposes the electroneutrality condition via the Lagrange multiplier , and is the imposed potential at electrode site . The BO stationarity conditions read
They impose the constant-potential conditions and overall electroneutrality. In BO dynamics, the stationary charges are obtained by minimization at each ionic configuration; in MaZe, the same conditions are maintained as holonomic constraints during propagation. In this model, is the (bordered) Hessian of with respect to the electrode charges and , which depends only on the fixed electrode geometry and not on the ionic coordinates. Consequently, the mass-zero compressibility vanishes, . This property simplifies the statistical description but is not required for the general BO–MaZe correspondence.
Let and denote the total charges of the two electrodes. Since , the capacitor charge and its associated current are
In MaZe, is obtained directly from the constrained charge velocities, which are components of . In BO dynamics, it is the derivative of the BO stationary charge along the ionic trajectory, evaluated in the following by central finite differences. The velocity correspondence identifies these currents.
The frequency-dependent admittance , which is the linear response function relating the current flowing through the cell to a small oscillating voltage applied between the two electrodes [8, 9], can be computed from equilibrium fluctuations through two expressions [8]. The current-based Nyquist–Johnson relation is
while its charge-correlation counterpart is
Here , , and . Eq. (22) probes the auxiliary velocities directly, whereas Eq. (23) depends only on the stationary auxiliary coordinates. Comparing both estimators in MaZe and BO therefore tests the dynamical and statistical correspondence at the level of a physical response function. The corresponding impedance is .
In our simulations, we consider molten LiCl between two aluminum electrodes simulated with MetalWalls [33, 34]. Both electrodes are held at the same null potential, V. The cell contains 500 Li and 500 Cl ions and two electrodes of 216 fixed Al atoms each, with Å and Å. Periodic boundary conditions are applied parallel to the electrodes. The distance between the innermost electrode planes is Å. The simulation protocol and interaction models are described in the SM [31]. We compare MaZe (integrated via the algorithm detailed in Figure 1) with a BO reference in which the electrode charges are minimized at every time step using conjugate-gradient.
Figure 3 compares the electrode currents from MaZe charge velocities and from central finite differences of the BO stationary charges, starting from the same initial conditions. For the first 1.2 ps the difference stays at a constant plateau of A ( of the rms current). It then grows exponentially as , with fs, reaches 10% of the rms current at 1.76 ps, and saturates at times the rms current, the value for two uncorrelated signals. The separation time is nearly independent of the timestep (1.76 ps for 1 fs and 1.81 ps for 0.25 fs), so it is not set by the integration error but by the onset of the (Lyapunov-like) amplification of small numerical differences caused by the chaotic dynamics of the liquid. Figure 4 compares the equilibrium current autocorrelation functions: MaZe and BO agree within statistical uncertainty as the divergence of the currents affects times longer than the correlation decay. We finally evaluate the admittance using both Eqs. (22) and (23), with one-sided Fourier–Laplace transforms

Figure 3. Difference between the electrode currents from MaZe and BO/CG for the same initial configuration: (thin line) and its rms in 50 fs bins (dots) over the first 3 ps. The difference is flat ( A, dotted line) up to ps, then grows exponentially as with fs (thick fit window) and saturates at , i.e. the trajectories decorrelate. The dashed line marks .

Figure 4. Normalized autocorrelation function of the current fluctuations, , obtained from MaZe dynamics (solid blue) and from the BO/CG reference (dashed red). The shaded areas indicate standard deviation over non-overlapping 100 ps blocks.
computed by Filon–Lagrange quadrature with a flat-top Hann taper (details in the SM [31]). Figure 5 compares the real part and the absolute value of the imaginary part from MaZe and BO. The four estimates agree within their 95% confidence intervals across the investigated frequency range. At low frequency, the response is that of a blocking capacitor in series with a resistor, with static capacitances μF/cm for MaZe and μF/cm for BO. The figure also shows that the expected quadratic (for the real part) and linear (for the imaginary part) behavior of this quantity is verified in our calculations. At higher frequency, the response is dominated by the bulk properties of the electrolyte, as signaled by the minimum of at rad/s, which corresponds to the change of sign of and lies close to the maximum of (a more detailed comparison with the bulk is given in the SM [31]). Overall, our results confirm the agreement between the current- and charge-correlation estimators, and between the two dynamics, in both the time and frequency domains.

Figure 5. Frequency-dependent admittance of the LiCl–Al cell, (a) real part and (b) absolute value of the imaginary part, normalised by , computed from the current ACF (Eq. (22)) and the charge ACF (Eq. (23)) for MaZe and BO/CG. Black dash-dotted-dotted lines: low-frequency power laws (a) and (b). The minimum in (b) near rad/s is the sign change of . Error bars are 95% confidence intervals. They are omitted around the minimum in (b) because at the sign change, where the interval becomes comparable to the value itself and the bars cannot be displayed on a logarithmic axis.
Conclusions
We have shown that the velocities of the constrained auxiliary variables in MaZe dynamics coincide with the time derivatives of the BO stationary solutions along the same physical trajectory, and that in Hamiltonian variables the mass-zero limit is regular, with auxiliary momenta and multipliers scaling linearly with the inertia and the velocities remaining explicit. The phase-space compressibility and invariant measure give a statistical description of these velocities, including the equilibrium correlation functions and linear-response relations built on them. In constant-potential molecular dynamics, MaZe and BO give consistent electrode currents, current autocorrelations and frequency-dependent admittance, with agreement between estimators based on current and charge correlations.
The correspondence between MaZe and BO is general: it applies whenever auxiliary variables are defined by smooth stationarity conditions. Examples include induced dipoles in polarizable force fields, and the related shell and Drude-oscillator models, whose time derivatives contribute to the polarization current and hence to infrared and dielectric spectra; fluctuating-charge models, where the charge velocities describe intramolecular charge-transfer currents; electron densities in orbital-free density functional theory, whose velocities give access to electronic currents; and constant-potential electrodes of other geometries, such as nanoporous electrodes in supercapacitors, where the charge-flow dynamics is the quantity of interest. For all cases, the results presented in the present work make it possible to interpret and compute the time derivatives of auxiliary BO variables as dynamical variables within a single constrained-dynamics framework.
Acknowledgments
This work has received funding from the European Union’s Horizon 2020 and Horizon Europe research and innovation programs under the Marie Skłodowska-Curie (grant agreement no. 101119598, FLUXIONIC), as well as from the European Research Council (grant agreement no. 863473).
Supplemental Material
Compressibility of the Flow
We start by deriving the phase-space compressibility of the finite-mass constrained flow in the variables . The constraint Jacobian and inverse mass matrix are
with an and a matrix. Writing and , where , the matrix introduced in Eq. (6) is
Multipliers. Differentiating twice along the flow gives , with . From Eqs. (3) and , . Hence with , as in Eq. (7). Since ,
which is Eq. (6); multiplies a matrix that is regular at . Here is regarded as a function of the independent variables ; on the constraint surface , Eq. (5). The phase-space divergence receives contributions only from the momentum equations:
Define
Because is independent of the momenta,
Let . From Eq. (7),
The second term is independent of the momenta, while
It follows that
Thus,
Differentiating at fixed gives
Using the symmetry of and and the cyclic property of the trace, we obtain
where the last equality uses precisely the metric factor of Eq. (14),
The momentum derivatives above are taken at fixed coordinates, treating and as independent variables; the velocity constraint is imposed through in the constrained phase-space measure. If is nonsingular and the constrained velocities remain finite, the mass-zero limit is
For the constant-potential model, is independent of time and hence .
Regular mass-zero limit of the invariant measure
The reduction of the constrained equilibrium measure to the Born–Oppenheimer (BO) density in the mass-zero limit was established in previous work [23, 24]. We include it in this work for completeness, providing a different derivation recast in the Hamiltonian variables used here, where it follows from the regular scaled momenta. We show how the average of Eq. (15) reduces to the BO density as . Change variables to , so that (the factor is common to numerator and denominator and cancels). On the constraint surface the velocity constraint is
with no dependence on . Integrating over and then over with gives, at finite ,
By the matrix determinant lemma, with , so that as . In the limit the physical variables are therefore distributed according to , the BO density for the chosen ensemble, with no residual metric factor, and the auxiliary velocity is the slaved value . The same steps applied to Eq. (17), where , give the weight quoted in the main text. For finite the metric contributes , where the term does not vanish identically for generic . Furthermore, the Hamiltonian contains a modified inertia for the physical variables, reading .
Details of the constant-potential system
The constraints are and , with
and . The matrices and only require derivatives of the Coulomb energy . With an Ewald-type splitting, separates into ion–ion, ion–electrode and electrode–electrode contributions, . In their bare (non-periodic) form these read
with the ionic charges, the (fixed) electrode-site positions, , , the inverse width of the Gaussian charge on electrode site (charge density ), and (= for equal widths ). Here and in the following, Coulomb interactions are written in atomic units (). The periodic (Ewald) form of these terms is given below, with the Ewald splitting parameter. Since does not depend on , is linear in , and does not depend on , only involves , and only involves .
: derivatives with respect to
The auxiliary Hessian is the bordered electrode–electrode Coulomb matrix
whose nontrivial block depends only on the fixed electrode geometry , on the box and on , .
Electrode–electrode contribution . It is split into short-range, long-range ( and ), and self-energy terms,
The double sums run over ordered pairs , which is why the short-range term carries a factor and the Hessians below differ from the prefactors of accordingly.
Short range. With ( the in-plane lattice vector labelled by ; the prime excludes when ) and ,
Long range, . With ,
Long range, . With , which is even in so that the sums below are real,
Self energy.
: derivatives with respect to
The elements of are
because does not depend on .
Electrode–ion contribution .
Each ion–electrode pair appears twice in the sum over ordered pairs of all particles, which gives the factors below.
Short range. With , and ,
Long range, . Using as above,
which does not depend on , and, since ,
Long range, . With , and ,
Simulation Protocol and Comparison with the Bulk
Following Ref. 24, all simulations in this work, including the bulk reference described below, use the same protocol for equilibration and production. After an NVT equilibration at 1500 K with a Nosé–Hoover chain thermostat, we perform NVE production runs of 1 ns with a 1 fs timestep. Independent MaZe and Born–Oppenheimer runs of the cell are started from the same equilibrated configuration, with ion velocities redrawn from a Maxwell–Boltzmann distribution at 1500 K, and are used to collect statistics. The one-sided Fourier–Laplace transforms in Eqs. (22) and (23) were evaluated by Filon–Lagrange quadrature [8], with a flat-top Hann taper applied to suppress long-time statistical noise [9]. We use two windows: 5 ps (flat for the first 1.5 ps) above rad/s, and 20 ps (flat for the first 5 ps) below. Low frequencies are determined by the long-time part of the autocorrelation functions, where the signal has decayed and the statistical noise dominates, the larger window corrects for this. The expected low-frequency behaviour follows from the blocking capacitor in series with a resistor [8], , which gives and for . In the current autocorrelation, is the small remainder of a cancellation between positive and negative parts (the blocking electrodes imply a vanishing zero-frequency conductance, ), which a short taper breaks, so a longer window is needed; at higher frequency the longer window only adds noise. Each 1 ns trajectory was divided into ten non-overlapping blocks of 100 ps. For each method (MaZe and BO/CG) we pooled the blocks of seven independent trajectories, giving 70 blocks per method.
To interpret the response of the cell over the whole frequency range, we compared it with a simulation of the molten salt in the absence of electrodes. The bulk reference is a fully periodic box without electrodes, containing the same 500 Li and 500 Cl ions and the same interaction potential as the cell, in a box of 24.30 24.30 46.87 Å.
To compare the cell with the bulk we need to define an effective slab thickness for the liquid. For dielectric liquids this can be done by locating a dielectric dividing surface (DDS) at each interface, which separates the electrode from a liquid with a well-defined bulk dielectric response [8]. In a dilute electrolyte, the DDS of the pure solvent has been used to define the effective slab thickness, even though in that case the response arises not only from the polar solvent but also from ionic transport [9]. For a conducting liquid such as molten LiCl this construction cannot be used: the liquid screens the electrodes, so the polarization is not localised in the bulk of the slab and no dielectric dividing surface can be defined. We therefore introduce an effective thickness , where is the distance between the innermost electrode planes and is the width of the ion-depleted layer at each interface, taken from the Gibbs dividing surface of the Li and Cl density profiles. In our simulations Å, so that Å for Å. Varying between 1 and 2 Å changes our results by less than 3%, and neither the bulk resistivity nor the static capacitance depends on this choice. The frequency-dependent conductivity of the bulk is obtained from the Green–Kubo relation
where is the volume of the bulk simulation box, and are the charge and velocity of ion , and , with the same procedure used for the cell.
Figure S1 compares the cell impedance rescaled by the geometry of the liquid slab, , with the bulk resistivity of molten LiCl, where is the electrode area, is the effective thickness of the liquid slab defined above, and is the frequency-dependent conductivity of the bulk liquid obtained from the separate bulk simulation. The comparison follows from the picture of the cell as the bulk liquid in series with the interfacial capacitance [9],

Figure S1. Impedance of the LiCl–Al cell (MaZe) compared with bulk LiCl, in a slab of effective thickness (, see the text): (a) and (b) , both multiplied by so that they can be compared directly with the bulk resistivity. The inset in (a) magnifies the low-frequency region on a linear scale. At low frequency of the cell diverges as because of the interfacial capacitance, which is absent in the bulk; above the two curves coincide. The minima of near correspond to the sign change of . Error bars are 95% confidence intervals. The bulk is small and compatible with zero within its 95% confidence interval over most of the range between and , which is why its error bars in (b) extend down past the lower edge of the logarithmic axis.
so that wherever the interfacial term is negligible compared with the bulk term. The capacitance arises only from the electrode–electrolyte interfaces and is absent in the bulk liquid, which has no electrodes and whose response is entirely contained in .
The real part is set by the bulk: equals at low frequency, within about 10% of the bulk value , with an interfacial excess of about 15% below rad/s. This indicates that the resistance of the cell is dominated by the bulk electrolyte. The imaginary part is instead dominated by the interface at low frequency, where reaches at rad/s, more than two orders of magnitude above the bulk value at the same frequency, which is of order or below. The bulk is statistically indistinguishable from zero below rad/s (95% CI), as expected since the bulk behaves as a plain resistor at low frequency, so its impedance is real. The minimum of at rad/s corresponds to the sign change of (and of ). The cell and bulk impedances merge for rad/s, i.e. above the sign change of (and of ), when the capacitive term is smaller than the bulk contribution.
References
- [1]M. Pearson, E. Smargiassi, and P. A. Madden, Ab initio molecular dynamics with an orbital-free density functional, J. Phys.: Condens. Matter 5, 3221 (1993).
- [2]D. Marx and J. Hutter, Ab Initio Molecular Dynamics: Basic Theory and Advanced Methods (Cambridge University Press, Cambridge, 2012).
- [3]B. G. Dick and A. W. Overhauser, Theory of the dielectric constants of alkali halide crystals, Phys. Rev. 112, 90 (1958).DOI
- [4]G. Jacucci, I. R. McDonald, and A. Rahman, Effects of polarization on equilibrium and dynamic properties of ionic systems, Phys. Rev. A 13, 1581 (1976).DOI
- [5]M. Wilson and P. A. Madden, Polarization effects in ionic systems from first principles, J. Phys.: Condens. Matter 5, 2687 (1993).DOI
- [6]M. Wilson, P. A. Madden, N. C. Pyper, and J. H. Harding, Molecular dynamics simulations of compressible ions, J. Chem. Phys. 104, 8068 (1996).DOI
- [7]J. I. Siepmann and M. Sprik, Influence of surface topology and electrostatic potential on water/electrode systems, J. Chem. Phys. 102, 511 (1995).DOI
- [8]G. Pireddu and B. Rotenberg, Frequency-dependent impedance of nanocapacitors from electrode charge fluctuations as a probe of electrolyte dynamics, Phys. Rev. Lett. 130, 098001 (2023).
- [9]G. Pireddu, C. J. Fairchild, S. P. Niblett, S. J. Cox, and B. Rotenberg, Impedance of nanocapacitors from molecular simulations to understand the dynamics of confined electrolytes, Proc. Natl. Acad. Sci. U.S.A. 121, e2318157121 (2024).DOI
- [10]D. Kivelson and P. Madden, Theory of dielectric relaxation, Mol. Phys. 30, 1749 (1975).DOI
- [11]J. M. Caillol, D. Levesque, and J. J. Weis, Electrical properties of polarizable ionic solutions. I. Theoretical aspects, J. Chem. Phys. 91, 5544 (1989).DOI
- [12]J. M. Caillol, D. Levesque, and J. J. Weis, Electrical properties of polarizable ionic solutions. II. Computer simulation results, J. Chem. Phys. 91, 5555 (1989).DOI
- [13]M. Sega, S. S. Kantorovich, A. Arnold, and C. Holm, On the calculation of the dielectric properties of liquid ionic systems, in Recent Advances in Broadband Dielectric Spectroscopy, NATO Science for Peace and Security Series B: Physics and Biophysics, edited by Y. P. Kalm ykov (Springer Netherlands, Dordrecht, 2013) pp. 103–122.DOI
- [14]R. Car and M. Parrinello, Unified approach for molecular dynamics and density-functional theory, Phys. Rev. Lett. 55, 2471 (1985).pubmed.ncbi.nlm.nih.gov/10032153
- [15]J. Kolafa, Time-reversible always stable predictor-corrector method for molecular dynamics of polarizable molecules, J. Comput. Chem. 25, 335 (2004).pubmed.ncbi.nlm.nih.gov/14696069
- [16]T. D. Kühne, M. Krack, F. R. Mohamed, and M. Parrinello, Efficient and accurate Car–Parrinello-like approach to Born–Oppenheimer molecular dynamics, Phys. Rev. Lett. 98, 066401 (2007).pubmed.ncbi.nlm.nih.gov/17358962
- [17]A. M. N. Niklasson, C. J. Tymczak, and M. Challacombe, Time-reversible Born–Oppenheimer molecular dynamics, Phys. Rev. Lett. 97, 123001 (2006).DOI
- [18]A. M. N. Niklasson, Extended Lagrangian Born–Oppenheimer molecular dynamics for orbital-free density-functional theory and polarizable charge equilibration models, J. Chem. Phys. 154, 054101 (2021).
- [19]P. Tangney and S. Scandolo, How well do Car–Parrinello simulations reproduce the Born–Oppenheimer surface? Theory and examples, J. Chem. Phys. 116, 14 (2002).DOI
- [20]P. Tangney, On the theory underlying the Car–Parrinello method and the role of the fictitious mass parameter, J. Chem. Phys. 124, 044111 (2006).DOI
- [21]I.-F. W. Kuo, C. J. Mundy, M. J. McGrath, and J. I. Siepmann, Time-dependent properties of liquid water: A comparison of Car–Parrinello and Born–Oppenheimer molecular dynamics simulations, J. Chem. Theory Comput. 2, 1274 (2006).DOI
- [22]S. W. Ong, E. S. Tok, and H. C. Kang, Vibrational frequencies in Car–Parrinello molecular dynamics, Phys. Chem. Chem. Phys. 12, 14960 (2010).DOI
- [23]S. Bonella, A. Coretti, R. Vuilleumier, and G. Ciccotti, Adiabatic motion and statistical mechanics via mass-zero constrained dynamics, Phys. Chem. Chem. Phys. 22, 10775 (2020).DOI
- [24]A. Coretti, L. Scalfi, C. Bacon, B. Rotenberg, R. Vuilleumier, G. Ciccotti, M. Salanne, and S. Bonella, Mass-zero constrained molecular dynamics for electrode charges in simulations of electrochemical systems, J. Chem. Phys. 152, 194701 (2020).DOI
- [25]A. Coretti, S. Bonella, and G. Ciccotti, Communication: Constrained molecular dynamics for polarizable models, J. Chem. Phys. 149, 191102 (2018).DOI
- [26]A. Coretti, T. J. Baird, R. Vuilleumier, and S. Bonella, Mass-zero constrained dynamics for simulations based on orbital-free density functional theory, J. Chem. Phys. 157, 214110 (2022).DOI
- [27]D. D. Girardier, A. Coretti, G. Ciccotti, and S. Bonella, Mass-zero constrained dynamics and statistics for the shell model in magnetic field, Eur. Phys. J. B 94, 158 (2021).
- [28]We note that previous solutions of MaZe dynamics employed a Verlet algorithm for the auxiliary variables, thus requiring only SHAKE to solve the constraints. The focus on the auxiliary velocities requires the velocity Verlet, combined with SHAKE and RATTLE, integrator introduced in this work.
- [29]M. E. Tuckerman, C. J. Mundy, and G. J. Martyna, On the classical statistical mechanics of non-Hamiltonian systems, Europhys. Lett. 45, 149 (1999).DOI
- [30]M. E. Tuckerman, Y. Liu, G. Ciccotti, and G. J. Martyna, Non-Hamiltonian molecular dynamics: Generalizing Hamiltonian phase space principles to non-Hamiltonian systems, J. Chem. Phys. 115, 1678 (2001).DOI
- [31]See Supplemental Material, included at the end of this document, for the derivation of the phase-space compressibility of the constrained flow, the regular mass-zero limit of the invariant measure, the details of the constant-potential electrode model, the simulation protocol, and the comparison of the cell impedance with the bulk liquid.
- [32]C. Hartmann, C. Schütte, and G. Ciccotti, Communications: On the linear response of mechanical systems with constraints, J. Chem. Phys. 132, 111103 (2010).DOI
- [33]A. Marin-Laflèche, M. Haefele, L. Scalfi, A. Coretti, T. Dufils, G. Jeanmairet, S. K. Reed, A. Serva, R. Berthin, C. Bacon, S. Bonella, B. Rotenberg, P. A. Madden, and M. Salanne. MetalWalls: A classical molecular dynamics software dedicated to the simulation of electrochemical systems, J. Open Source Softw. 5, 2373 (2020).
- [34]A. Coretti, C. Bacon, R. Berthin, A. Serva, L. Scalfi, I. Chubak, K. Goloviznina, M. Haefele, A. Marin-Laflèche, B. Rotenberg, S. Bonella, and M. Salanne, MetalWalls: Simulating electrochemical interfaces between polarizable electrolytes and metallic electrodes, J. Chem. Phys. 157, 184801 (2022).DOI