Fact-checked by Grok 2 weeks ago

Water model

A water model is an empirical mathematical approximation used in computational simulations, particularly (MD), to represent the structure, interactions, and behavior of molecules at the atomic level. These models simplify the quantum mechanical nature of into classical force fields, typically treating molecules as rigid or flexible geometries with point charges and Lennard-Jones potentials to capture electrostatic and van der Waals interactions, respectively. Developed to reproduce essential bulk properties of liquid —such as , self-diffusion coefficient, heat of vaporization, dielectric constant, isobaric , thermal expansion coefficient, and isothermal —they enable realistic modeling of effects in complex systems like biomolecules, materials, and chemical reactions. Water models have evolved since the 1970s, starting with simple three-point representations like the model, which places partial charges on the oxygen and atoms with fixed lengths and angles. Subsequent refinements addressed limitations in reproducing experimental data; for instance, the extended model incorporates a correction to better match and coefficients, while four-point models like TIP4P shift charges to a dummy site on the bisector of the H-O-H angle for improved tetrahedral geometry and properties. Polarizable models, such as those using oscillators or inducible dipoles, further enhance accuracy by accounting for electronic , though they increase computational cost; examples include and SWM4-NDP, which perform well in biomolecular contexts but require specialized software. Among non-polarizable models, TIP3P and remain widely used due to their balance of efficiency and fidelity, with TIP3P favoring faster simulations despite slight underestimation of . The choice of water model profoundly impacts simulation outcomes, as discrepancies in properties like hydrogen bonding or solvation can alter predicted structures and in applications ranging from to . Recent advancements, including machine learning-derived potentials, aim to bridge classical and quantum accuracy, but classical models dominate due to scalability in large-scale studies. Overall, water models underscore the interplay between computational tractability and physical realism, continually refined through benchmarking against experimental data to support interdisciplinary research in , , and .

Fundamentals

Definition and scope

Water models are simplified mathematical representations of the H₂O molecule employed in to approximate its . These models use empirical parameters, such as partial atomic charges and van der Waals coefficients, which are derived from quantum mechanical calculations, experimental measurements, and optimization against bulk properties including , radial distribution functions, and . The scope of water models encompasses simulations of liquid , ice phases, small clusters, and aqueous solutions containing solutes like ions or biomolecules, primarily through and methods. Key components include electrostatic interactions modeled via Coulombic potentials between charged sites on the molecules and van der Waals attractions/repulsions captured by Lennard-Jones potentials. Non-bonded interactions, which dominate intermolecular forces in these models, exclude covalent bonds within the molecule itself and focus on long-range electrostatic and effects between separate molecules. The general form of the pairwise U is: U = \sum_{i < j} \frac{q_i q_j}{4 \pi \epsilon_0 r_{ij}} + \sum_{i < j} 4 \epsilon \left[ \left( \frac{\sigma}{r_{ij}} \right)^{12} - \left( \frac{\sigma}{r_{ij}} \right)^6 \right] where q_i and q_j are partial charges on interaction sites i and j, r_{ij} is the inter-site distance, \epsilon_0 is the vacuum permittivity, and \epsilon and \sigma are Lennard-Jones energy and size parameters, respectively. Water models facilitate large-scale classical simulations of complex aqueous systems where ab initio quantum mechanical approaches are computationally prohibitive due to the need to treat thousands of molecules over extended timescales. This enables detailed studies of solvation dynamics, hydrogen bonding networks, and thermodynamic properties that underpin biological and chemical processes in water. In contrast to continuum solvation models, which approximate the solvent as a homogeneous dielectric continuum without explicit molecules, water models treat water as discrete particles to capture molecular-level details. The first computational use of such a model dates to the 1971 molecular dynamics simulation of liquid water by Rahman and Stillinger.

Historical development

The development of water models began in the early 20th century with theoretical efforts to describe the structure of ice and liquid water based on geometric considerations. In 1933, Bernal and Fowler proposed the first explicit model for water, representing the molecule as a rigid tetrahedron with two hydrogen atoms and a lone pair, emphasizing hydrogen bonding in ice structures without computational simulations. This geometric approach laid foundational insights into water's anomalous properties but lacked dynamic treatment due to the absence of computing resources. The advent of molecular dynamics (MD) simulations in the 1970s marked a pivotal shift toward computational modeling of liquid water. The first MD simulation of liquid water was performed by Rahman and Stillinger in 1971, using a three-site potential to study structural and dynamic properties at ambient conditions, demonstrating the feasibility of simulating water's hydrogen-bond network. By 1974, Stillinger and Rahman refined this with the ST2 model, a six-site rigid water potential incorporating Lennard-Jones interactions and electrostatic charges to better capture short-range repulsion and hydrogen bonding, enabling more accurate reproduction of liquid water's radial distribution functions. Concurrently, the MCY potential, derived from ab initio configuration interaction calculations on water dimers—a pairwise additive model—addressed limitations in transferability across phases. These early efforts were motivated by the need to balance quantum-derived accuracy with the computational demands of MD, though models often overestimated melting points or struggled with dielectric properties. The 1980s saw simplifications for broader applicability, particularly in biomolecular simulations where efficiency was paramount. Jorgensen et al. introduced the three-site and four-site models in 1983, using fixed partial charges and Lennard-Jones sites to approximate water's electrostatics and dispersion with fewer parameters than ST2, facilitating Monte Carlo and MD studies of aqueous solutions. In 1987, Berendsen et al. developed the model, an extension of the earlier SPC three-site potential, by adding a polarization correction via a scaling factor to improve the dielectric constant and radial distribution at liquid densities. These rigid, non-polarizable models prioritized computational speed over detailed many-body effects, responding to challenges in simulating large systems like proteins in water. From the 1990s onward, models evolved to incorporate polarizability and optimize phase behavior. Polarizable models, such as the in the 1990s, allowed charges to fluctuate in response to electric fields, enhancing accuracy for interfaces and ionic solutions. Further ab initio-based refinements continued, while the 2000s introduced , a polarizable atomic multipole model that explicitly includes higher-order electrostatics for better thermodynamic properties across temperatures. Optimized rigid models emerged, including in 2005, which improved reproduction of water's phase diagram and anomalies like density maximum, and in 2014, focusing on global optimization of bulk properties using experimental targets. Throughout, motivations centered on reconciling accuracy in reproducing experimental observables—such as the phase diagram, diffusion coefficients, and dielectric response—with computational efficiency, while addressing transferability issues across vapor, liquid, and solid phases. Post-2020, machine learning-based potentials, trained on quantum mechanical data, have begun to enable highly accurate, flexible representations of water's potential energy surface for large-scale simulations.

Classification

By interaction sites

Water models are classified by the number of interaction sites, which are points representing partial charges and van der Waals interactions within the water molecule, including atomic positions (oxygen and hydrogens) or virtual (dummy) sites for lone pairs. These sites enable the modeling of electrostatic forces via Coulombic interactions between partial charges q and Lennard-Jones potentials for dispersion and repulsion. Increasing the number of sites enhances the accuracy of the electrostatic charge distribution, better capturing the molecular dipole moment and hydrogen bonding geometry, but at the computational expense of more pairwise interactions per molecule pair. Two-site models represent the simplest configuration, with partial charges placed only on the oxygen and a single effective site for the merged hydrogens, omitting explicit hydrogen positions to minimize complexity. This approach is particularly suited for applications requiring efficient predictions of dielectric properties, such as in coarse-grained simulations of bulk water behavior. Three-site models assign partial charges directly to the oxygen and two hydrogen atoms, maintaining a rigid geometry that approximates the experimental structure, for example, with O-H bond lengths of 0.9572 Å and H-O-H angles of 104.52°. This setup balances computational efficiency with reasonable electrostatic representation, making it widely applicable for simulating liquid water properties. Four-site models extend the three-site framework by introducing an off-center virtual negative charge site (M-site) along the angle bisector, typically displaced 0.15 Å from the oxygen, to improve the dipole moment and hydrogen bonding directionality without altering atomic positions. This addition refines the electrostatics for better agreement with experimental solvation and structural data. Five-site models incorporate two positive charges on the hydrogens and two negative charges on virtual sites mimicking the tetrahedral lone pairs, with no charge on the oxygen atom, enhancing the tetrahedral coordination and density anomaly reproduction in liquid water. Six-site models combine elements of four- and five-site designs, adding extra virtual sites to further detail charge distribution, though they are less common and primarily used for specialized simulations of ice-water interfaces near the melting point. Overall, three- and four-site models predominate due to their optimal trade-off between accuracy and computational cost; the number of sites directly influences the pairwise interaction count, such as nine site-site distances for a three-site model pair versus more for higher-site variants.

By molecular flexibility and polarizability

Water models can be classified according to their treatment of molecular flexibility and polarizability, which extend beyond the static geometry and fixed charges of rigid, non-polarizable representations to better capture the dynamic behavior of water molecules in condensed phases. Flexibility refers to the ability of the model to account for intramolecular vibrations, such as oscillations in bond lengths and angles, which are absent in rigid models that enforce fixed geometries using constraint algorithms like . In flexible models, these degrees of freedom are governed by intramolecular potential energy functions, typically harmonic forms for stretching and bending: U_{\text{bond}} = \frac{k_b}{2} (r - r_0)^2 U_{\text{angle}} = \frac{k_\theta}{2} (\theta - \theta_0)^2 where k_b and k_\theta are the respective force constants, and r_0 and \theta_0 are the equilibrium bond length and angle. This approach enables the simulation of vibrational modes, improving the representation of dynamic properties like diffusion and spectroscopic signatures compared to constrained rigid baselines. Flexible models, such as variants of the simple point charge (SPC) potential developed in the 1990s, allow bond lengths to vary around experimental gas-phase values, enhancing realism in liquid-state simulations. Polarizability addresses the redistribution of a molecule's electron density in response to the local electric field from neighboring molecules, an effect ignored in fixed-charge models that assume invariant partial charges. In polarizable models, this induction is incorporated either implicitly or explicitly. Implicit schemes approximate the average polarization in bulk water by adjusting fixed charges to an effective value that mimics the enhanced molecular dipole moment, as in the where the charge is set to -0.8476 e on oxygen and +0.4238 e on each hydrogen to account for liquid-phase polarization without dynamic response. Explicit methods, however, directly compute the response, often via inducible point dipoles satisfying \vec{\mu}_{\text{ind}} = \alpha \vec{E}_{\text{local}} with the molecular polarizability \alpha \approx 1.4 \, \AA^3, matching the experimental gas-phase value of water. One prevalent explicit approach uses Drude oscillators, where a lightweight negative charge is harmonically bound to an atomic site to represent electron cloud displacement under the field, enabling iterative or extended Lagrangian propagation in simulations. Seminal explicit polarizable models include AMOEBA, introduced in 2003, which combines atomic multipoles with inducible dipoles for accurate many-body electrostatics. Incorporating flexibility and polarizability enhances the fidelity of simulations for properties sensitive to molecular dynamics, such as infrared (IR) spectra, where flexible models better reproduce experimental vibrational frequencies and widths by allowing explicit OH stretching and HOH bending modes. These advanced models also improve hydrogen bonding networks and dielectric responses in heterogeneous environments. However, they incur a computational overhead of 2-10 times compared to rigid non-polarizable counterparts due to additional degrees of freedom and field iterations, limiting their use to refined or smaller-scale studies despite the gains in accuracy.

Rigid non-polarizable models

Two-site models

Two-site models are the simplest rigid non-polarizable representations of water molecules in molecular simulations, consisting of two interaction sites: the oxygen atom and a single positive charge site that effectively merges the two hydrogen atoms along the molecular symmetry axis. This formulation simplifies the electrostatic interactions to a point-charge model, with the oxygen site assigned a negative charge (q_O) and the hydrogen site an equal-magnitude positive charge (q_H = -q_O) to maintain molecular neutrality. The potential is restricted to oxygen-oxygen interactions to capture short-range dispersion and repulsion, while electrostatics dominate intersite forces involving the hydrogen site. Such models are particularly suited for theoretical analyses using integral equation methods, where the reduced number of sites facilitates analytical tractability. The intermolecular pair potential in these models is expressed as u(\mathbf{r}, \boldsymbol{\Omega}_1, \boldsymbol{\Omega}_2) = \sum_{i,j = \mathrm{O,H}} u_{ij}(r_{ij}), where \mathbf{r} is the vector between molecular centers of mass, \boldsymbol{\Omega}_1 and \boldsymbol{\Omega}_2 denote molecular orientations, and r_{ij} is the distance between sites i and j on different molecules. The site-site potential u_{ij} combines Coulombic electrostatics, q_i q_j / (4\pi \epsilon_0 r_{ij}), for all pairs, with an additional LJ term only for O-O interactions: $4\epsilon [(\sigma / r_{\mathrm{OO}})^{12} - (\sigma / r_{\mathrm{OO}})^6]. Representative parameters, derived to approximate experimental properties like the gas-phase dipole moment of approximately 1.85 D, include \epsilon \approx 0.156 kcal/mol, \sigma \approx 3.17 Å, and charges on the order of |q| \approx 0.82 e (with site separation adjusted accordingly, e.g., around 0.7–1.0 Å along the bisector). These values stem from early basic formulations in the 1970s, refined in later theoretical works to balance simplicity and fidelity to bulk properties. A seminal example is the two-site model proposed in 2009 for theoretical studies using integral equation methods, derived from the . This approach places the negative charge at the oxygen position and the positive charge at a displaced site to mimic the dipole, with no intramolecular flexibility or polarizability. The model's low computational cost arises from the minimal number of interactions (only four per pair: O-O LJ + Coulomb, O-H Coulomb twice, H-H Coulomb), making it ideal for large-scale simulations or analytical derivations in gas-phase studies and dielectric response calculations. For instance, site-renormalized molecular fluid theory using this model accurately predicts the static dielectric constant of liquid water near 300 K (ε ≈ 78) by treating sites as effective simple fluids. Despite these advantages, two-site models exhibit significant limitations in reproducing the structural properties of liquid water, such as radial distribution functions and hydrogen bonding networks, due to the absence of explicit angular dependence from separate hydrogen sites. Simulations often overestimate liquid density by 10–20% at ambient conditions and fail to capture tetrahedral coordination, leading to unrealistic diffusion coefficients and solvation free energies for polar solutes. Consequently, these models are rarely employed standalone for condensed-phase simulations today, serving instead as baselines for comparison with more elaborate three- or four-site variants or in preliminary theoretical explorations.

Three-site models

Three-site models represent a class of rigid, non-polarizable water potentials that place partial charges on the oxygen atom and the two hydrogen atoms, enabling explicit representation of hydrogen bonding interactions. These models typically employ a fixed geometry derived from experimental gas-phase values, with bond lengths of OH = 0.9572 Å and the HOH angle = 104.52°, while the Lennard-Jones (LJ) interaction is centered solely on the oxygen site. The partial charges generally range from q_O = -0.8 to -1.04 e and q_H = +0.4 to +0.52 e, balancing the molecular dipole moment (approximately 1.85 D) and electrostatic interactions. Among the earliest three-site models is the Transferable Intermolecular Potential with Superimposed charges (TIPS), introduced as a precursor to later variants, which uses charges of q_O = -0.80 e and q_H = +0.40 e, along with LJ parameters ε = 0.15 kcal/mol and σ = 3.12 Å to approximate liquid properties in Monte Carlo simulations. The Simple Point Charge (SPC) model, developed in 1981, refines this approach with q_O = -0.82 e, q_H = +0.41 e, ε = 0.110 kcal/mol, and σ = 3.166 Å, prioritizing computational efficiency for molecular dynamics (MD) studies of protein hydration. Concurrently, the Transferable Intermolecular Potential 3-Point (TIP3P) model from 1983 employs slightly adjusted parameters: q_O = -0.834 e, q_H = +0.417 e, ε = 0.152 kcal/mol, and σ = 3.1507 Å, offering improved reproduction of vapor pressure and solvation free energies compared to SPC. To address limitations in dielectric screening, the SPC/E (extended SPC) model was proposed in 1987, scaling the charges to q_O = -0.8476 e and q_H = +0.4238 e while retaining the original LJ parameters and geometry; this adjustment mimics average polarization effects without explicit polarizability, enhancing agreement with experimental radial distribution functions. These models evolved from two-site representations by incorporating explicit hydrogen sites, which better capture directional hydrogen bonding essential for biomolecular environments.
ModelYearq_O (e)q_H (e)ε (kcal/mol)σ (Å)
TIPS1981-0.80+0.400.153.12
SPC1981-0.82+0.410.1103.166
TIP3P1983-0.834+0.4170.1523.1507
SPC/E1987-0.8476+0.42380.1103.166
The simplicity of three-site models makes them highly efficient for large-scale MD simulations, as implemented in popular packages like and , where they routinely reproduce liquid water density near 1 g/cm³ at 298 K with errors under 1%. However, they generally underestimate the static dielectric constant (around 70 versus the experimental 78) due to the absence of polarizability, and performance degrades at low temperatures, overestimating melting points and failing to capture the density maximum near 277 K. Despite these shortcomings, their balance of accuracy and speed has made them staples in biomolecular simulations for decades.

Four-site models

Four-site water models extend the three-site rigid non-polarizable framework by introducing a fourth massless site (M) bearing a negative charge, positioned along the bisector of the H-O-H angle and displaced from the oxygen atom by 0.15–0.18 Å. This configuration places zero charge on the oxygen (q_O ≈ 0 e), positive charges on the hydrogens (q_H ≈ +0.52 e), and a balancing negative charge on the M site (q_M ≈ -1.04 e), with Lennard-Jones interactions assigned solely to the oxygen site. The electrostatic interactions are modeled using Coulombic potentials between all charged sites, yielding an effective molecular dipole moment of approximately 2.18 D, which better approximates the enhanced polarity in condensed phases compared to three-site models. The seminal TIP4P model, introduced in 1983, uses OH bond length of 0.9572 Å, H-O-H angle of 104.52°, M-site displacement of 0.15 Å, q_H = +0.52 e, q_M = -1.04 e, and oxygen Lennard-Jones parameters σ = 3.1536 Å, ε = 0.155 kcal/mol. This model provides reasonable descriptions of liquid water's structure and thermodynamics at ambient conditions, with oxygen-oxygen radial distribution functions aligning well with neutron scattering data. Subsequent refinements addressed limitations in phase behavior and electrostatic treatment. The TIP4P/2005 model (2005) reparameterizes the original with q_H = +0.5564 e, q_M = -1.1128 e, M-site displacement of 0.1546 Å, and oxygen Lennard-Jones parameters σ = 3.1589 Å, ε/k_B = 93.2 K (≈0.185 kcal/mol), optimizing for liquid density anomalies, isothermal compressibility, and the phase diagram. It predicts a melting point of ice I_h at approximately 250 K (versus experimental 273 K) and accurately reproduces multiple ice polymorphs and the liquid-vapor coexistence curve. The TIP4P-Ew variant (2004) adjusts for compatibility with Ewald summation methods used in biomolecular simulations, featuring q_H = +0.52422 e, q_M = -1.04844 e, M-site displacement of 0.125 Å, and oxygen Lennard-Jones parameters σ = 3.16435 Å, ε = 0.16275 kcal/mol; it improves bulk properties like density maximum near 1 °C and enthalpy of vaporization across the liquid range. The OPC model (2014) employs an optimization of charge placement without rigid geometric constraints beyond symmetry, yielding q_H = +0.6791 e, q_M = -1.3582 e, OH bond length of 0.8724 Å, H-O-H angle of 103.6°, M-site displacement of 0.1594 Å, and oxygen Lennard-Jones parameters σ = 3.166 Å, ε = 0.2128 kcal/mol; this enhances transferability across properties like solvation free energies and dielectric constant. Notable variants include TIP4P/Ice (2005), tailored for supercooled water and ice phases with adjusted Lennard-Jones ε/k_B = 106 K to better match ice densities and melting behavior at 272 K. The TIP4P-D model (2015) incorporates an empirical dispersion correction to the Lennard-Jones term, improving predictions of solvation properties and structural features in protein environments without altering core electrostatics. These models excel over three-site counterparts by mitigating inaccuracies in tetrahedral coordination and dipole representation through the off-center negative charge, yielding superior liquid structure, phase equilibria, and ice properties while maintaining computational efficiency.
ModelYearq_H (e)M displacement (Å)σ_O (Å)ε_O (kcal/mol)Key Optimization
TIP4P1983+0.520.153.15360.155Liquid structure/thermodynamics
TIP4P/20052005+0.55640.15463.15890.185Phase diagram/anomalies
TIP4P-Ew2004+0.524220.1253.164350.16275Ewald sums/biomolecules
OPC2014+0.67910.15943.1660.2128Transferability/solvation

Five- and six-site models

Five-site models of water explicitly incorporate two additional interaction sites to represent the lone pairs of electrons on the oxygen atom, enabling a more accurate depiction of the directional nature of hydrogen bonds compared to models with fewer sites. In these rigid, non-polarizable formulations, the oxygen atom carries no charge (q_O = 0), while each hydrogen bears a positive charge of approximately +0.4 e, and each lone pair site carries -0.4 e. The lone pair sites are positioned about 0.7 Å from the oxygen along the bisector of the H-O-H angle, maintaining tetrahedral geometry with bond angles near 104.5° for the hydrogens and ~109.5° for the lone pairs; the Lennard-Jones parameters are typically assigned solely to the oxygen site. This configuration enhances the model's ability to capture the angular specificity of hydrogen bonding in condensed phases like ice and clusters, though at the expense of increased computational demands. The ST2 model, introduced in 1974, represents an early and influential five-site approach designed to simulate liquid water with strong emphasis on hydrogen bond directionality through a stiff square-well potential between lone pairs and hydrogens. Its Lennard-Jones parameters are ε = 0.067 kcal/mol and σ = 3.10 Å on oxygen, with lone pair sites at 0.8 Å from oxygen. The ST2 model excels in reproducing the Bernal-Fowler rules for ice structure, ensuring each oxygen has exactly four neighboring hydrogens—two covalently bonded and two hydrogen-bonded at longer distances—thus providing excellent agreement for water clusters and solid phases. However, its rigidity leads to overestimated cohesive energies and a melting point around 285 K, higher than the experimental 273 K. A more modern five-site model is , developed in 2000 to improve upon three- and four-site models in reproducing liquid water properties like density across a wide temperature range (-37 to 100 °C) and pressure up to 10,000 atm. It uses Lennard-Jones parameters ε = 0.236 kcal/mol and σ = 3.56 Å on oxygen, with lone pair sites at 0.669 Å from oxygen and charges of +0.241 e on hydrogens and -0.241 e on lone pairs. yields a melting point of approximately 274 K, close to experiment, and provides good self-diffusivity values (around 2.4 × 10^{-5} cm²/s at 298 K, near the experimental 2.3 × 10^{-5} cm²/s), making it suitable for studies of liquid water dynamics and hydrogen bonding in clusters. An extension, from 2004, reoptimizes the parameters (ε = 0.243 kcal/mol, σ = 3.562 Å) for compatibility with Ewald summation methods to handle long-range electrostatics, maintaining similar performance in thermodynamic and transport properties while improving accuracy in periodic simulations. Six-site models build on the five-site framework by adding an extra charge site, often on the oxygen or another virtual position, to refine electrostatics for specific interfaces or phases. A notable example is the 2003 model by , tailored for simulating the ice-water interface near the melting point, featuring six charge sites (q_H ≈ +0.33 e, q_O ≈ -0.66 e split across sites, with additional negative charges for lone pairs) and Lennard-Jones on oxygen (ε = 0.155 kcal/mol, σ = 3.166 Å). This model accurately predicts the melting behavior of ice Ih at 273 K and the structure of the solid-liquid interface, with low interfacial free energy (~28 mJ/m², close to experimental estimates). Such models offer enhanced precision for phase boundaries but amplify computational costs, as each molecular pair involves up to 36 electrostatic interactions plus Lennard-Jones terms. Overall, five- and six-site models provide superior hydrogen bonding fidelity for ice and clusters relative to four-site variants, which rely on off-center charges without explicit lone pairs, but their higher complexity—requiring 17–26 interactions per pair—limits use in large-scale simulations. While effective for targeted applications like supercooled water or interfaces, they often overestimate melting points in some variants (e.g., ST2) or require careful parameterization to avoid artifacts in liquid properties.

Flexible and polarizable models

Flexible variants

Flexible variants of water models extend rigid non-polarizable models by incorporating intramolecular degrees of freedom, allowing bond lengths and angles to fluctuate according to harmonic or anharmonic potentials. This is achieved by adding an intramolecular potential energy term to the intermolecular interactions, typically expressed as U_{\text{intra}} = \sum U_{\text{bond}} + U_{\text{angle}} + U_{\text{improper}}, where the bond stretching and angle bending terms are often harmonic functions: U_{\text{bond}} = \frac{1}{2} k_{\text{bond}} (r - r_0)^2 and U_{\text{angle}} = \frac{1}{2} k_{\text{angle}} (\theta - \theta_0)^2. The improper dihedral term constrains out-of-plane distortions. These modifications enable the models to capture vibrational motions without relying on constraint algorithms like . A seminal example is the flexible SPC (SPC/Fw) model, developed in 2006 as a three-site extension of the rigid SPC model by removing geometric constraints and tuning intramolecular parameters to match experimental diffusion and dielectric properties. In this model, the O-H bond stretching uses k_{\text{bond}} = 450 kcal mol^{-1} Å^{-2} and equilibrium distance r_0 = 1.0 Å, while the H-O-H angle bending employs k_{\text{angle}} = 55 kcal mol^{-1} rad^{-2} and equilibrium angle \theta_0 = 109.47^\circ. Similarly, the flexible TIP3P model adapts the rigid three-site TIP3P geometry with comparable intramolecular potentials, often using k_{\text{bond}} \approx 450 kcal mol^{-1} Å^{-2} and k_{\text{angle}} \approx 100 kcal mol^{-1} rad^{-2}, to study proton transfer and solvation dynamics. Four-site flexible models, such as introduced in 2011, build on the rigid by adding flexibility tuned to reproduce vibrational spectra, using Morse potentials for O-H stretches (D_r = 432.581 kJ mol^{-1}, \beta = 22.87 nm^{-1}, r_{eq} = 0.9419 Å; harmonic limit k \approx 1081 kcal mol^{-1} Å^{-2}) and harmonic bending (k_{\theta} = 367.810 kJ mol^{-1} rad^{-2} \approx 87.9 kcal mol^{-1} rad^{-2}, \theta_{eq} = 107.4^\circ ). Another notable variant is from 2009, a four-site model incorporating quasi-harmonic O-H stretches via Morse functions and a cubic angle term to facilitate charge redistribution and flexibility, enhancing quantum nuclear effects in classical simulations. These models maintain fixed partial charges but allow geometric relaxation. The primary strengths of flexible variants lie in their ability to reproduce infrared (IR) and Raman spectra, as well as vibrational dynamics, by explicitly modeling intramolecular modes that rigid models approximate. For instance, and accurately predict the O-H stretching frequencies around 3400 cm^{-1} and bending modes near 1600 cm^{-1}, improving agreement with experimental spectroscopy. They are particularly valuable in quantum mechanics/molecular mechanics (QM/MM) hybrid simulations, where flexibility aids in describing reactive processes like proton hopping without constraints. However, these models increase computational demands by necessitating smaller integration timesteps (typically 0.5–1 fs versus 2 fs for rigid models) to maintain energy conservation during bond vibrations, potentially limiting efficiency in large-scale simulations. Additionally, lacking polarizability, they do not account for electronic response to environments, restricting accuracy in highly polar or anisotropic settings.

Explicitly polarizable models

Explicitly polarizable water models incorporate inducible atomic polarizabilities to capture the redistribution of electron density in response to varying electrostatic environments, such as those encountered in dense liquids or at interfaces. These models treat polarization explicitly by allowing atomic charges or dipoles to adjust dynamically during simulations, improving accuracy over fixed-charge representations for properties sensitive to environmental fluctuations. Two primary formulations are used to implement explicit polarization. In the inducible point dipole approach, induced dipoles \boldsymbol{\mu}_{\ind} on each atom are computed as \boldsymbol{\mu}_{\ind} = \alpha \mathbf{E}, where \alpha is the atomic polarizability and \mathbf{E} is the local electric field; these dipoles are solved self-consistently via iterative methods to account for mutual induction between molecules. The associated polarization energy is given by U_{\pol} = -\frac{1}{2} \sum_i \boldsymbol{\mu}_{\ind, i} \cdot \mathbf{E}_i, which reflects the energy gained from aligning induced dipoles with the field. Alternatively, the Drude oscillator method attaches an auxiliary charged particle (Drude particle) to each polarizable site, connected by a harmonic spring with potential U_{\Drude} = \frac{k}{2} r^2, where r is the displacement from equilibrium and k is the spring constant; the position of the Drude particle equilibrates to produce an effective induced dipole proportional to the local field. Prominent examples include the Atomic Multipole Optimized Energetics for Biomolecular Applications (AMOEBA) model, introduced in 2003, which combines permanent atomic multipoles up to the quadrupole level with inducible dipoles on oxygen (\alpha_O = 0.837 ų) and hydrogens (\alpha_H = 0.496 ų). Another key model is SWM4-NDP (Simple Water Model with Negative Drude Particle), a four-site rigid water representation from 2006 featuring Drude oscillators on the oxygen atom and negative charges on hydrogens to mimic polarization effects. Earlier contributions include the Gaussian charge (GK) model from 1994, which employs delocalized Gaussian charge distributions for inducible polarization, and the POLIR model from the 1990s, which uses inducible point dipoles for intermolecular interactions. Recent developments include the OPC3-pol model (2023), which provides efficient explicit polarizability for improved solvation and interface properties with minimal added cost. These models excel in reproducing key bulk properties, such as the dielectric constant of approximately 78 for liquid water at ambient conditions, and provide more accurate solvation free energies compared to non-polarizable alternatives. They also perform better at interfaces, where polarization effects are pronounced, yielding improved descriptions of surface tension and ion adsorption. However, the need for iterative self-consistent field calculations to converge induced dipoles increases computational cost by a factor of 5–10 relative to fixed-charge models, and parameterization demands high-level quantum mechanical data for polarizabilities and interaction damping. Such models can be combined with flexible water variants to further enhance vibrational accuracy, though this adds complexity.

Many-body interaction models

Many-body interaction models for water extend pairwise potentials by incorporating explicit three-body and higher-order terms to account for cooperative effects in hydrogen-bonded networks, such as charge transfer and non-additive dispersion forces. These models decompose the total potential energy as U_{\text{total}} = U_1 + U_2 + U_3 + \cdots, where U_1 represents one-body terms (e.g., intramolecular vibrations), U_2 captures pairwise interactions, and U_3 includes short-range three-body contributions often modeled via or empirical charge-transfer functions. Higher-order terms are typically truncated at three-body for computational feasibility, though they can approximate longer-range effects through induction. One seminal model is the MCY potential, developed in 1977 from ab initio calculations on water dimers and extended to clusters with explicit three-body corrections to capture non-additive effects in small aggregates. The MCY approach fits pair potentials to configuration interaction data while adding three-body terms derived from supermolecule calculations, enabling accurate descriptions of dimer binding energies and cluster geometries. In the 1990s, the NCC model by Niesar, Corongiu, and Clementi introduced explicit many-body polarization and three-body dispersion terms based on ab initio data for water trimers, building on polarizable pair interactions to better represent cooperative polarization in liquid-like environments. This model uses inducible point dipoles and anisotropic polarizabilities, with three-body contributions fitted to Hartree-Fock energies for 7,533 trimer configurations, improving transferability to ionic solutions. A more recent advancement is the MB-Pol model (2013), a transferable many-body potential derived entirely from coupled-cluster singles, doubles, and perturbative triples (CCSD(T)) calculations at the complete basis set limit, incorporating one-, two-, and three-body terms with short-range charge transfer and Axilrod-Teller-Muto dispersion. MB-Pol achieves root-mean-square deviations of 0.05 kcal/mol for dimer energies and has been extended to ion-water interactions, accurately reproducing water's phase diagram, including the density maximum at 4°C and viscosity anomalies. For instance, simulations with MB-Pol in 2023 confirmed experimental liquid-vapor coexistence curves across wide temperature and pressure ranges. A 2023 refinement, MB-pol(2023), achieves sub-chemical accuracy (better than 0.1 kcal/mol) for properties from gas to liquid phases. Water-specific refinements include the BK3 model (2013), which adds a three-body correction to TIP4P-like pairwise potentials via polarizable Drude oscillators to model non-additive induction and dispersion in bulk water. This approach enhances predictions of structural properties like radial distribution functions compared to non-polarizable baselines. These models excel in capturing water's thermodynamic anomalies, such as the density maximum at 4°C and elevated viscosity, which arise from cooperative three-body effects absent in pairwise approximations. However, their computational cost—often 10–50 times higher than rigid pairwise models due to explicit n-body evaluations—limits applications to small systems or clusters, with larger simulations requiring specialized algorithms or approximations. Despite this, MB-Pol serves as a benchmark for validating emerging potentials in phase diagram studies.

Emerging approaches

Machine learning-based potentials

Machine learning-based potentials represent a class of data-driven models that approximate the quantum mechanical potential energy surface of water systems by training on high-fidelity ab initio data, such as density functional theory (DFT) calculations, to achieve near-quantum accuracy at the computational cost of classical simulations. These models typically use neural networks or kernel methods, where atomic coordinates serve as input to predict energies and forces, enabling efficient molecular dynamics simulations of large-scale water systems. For instance, the ANI framework employs deep neural networks trained on DFT datasets to deliver transferable potentials with force field-like speed for organic molecules including water. Similarly, the Deep Potential Molecular Dynamics (DeepMD) approach, implemented in the DeePMD-kit package, constructs many-body potentials from extensive DFT trajectories, capturing complex interactions in liquid water and aqueous solutions. Gaussian Approximation Potentials (GAP), based on kernel ridge regression, provide a sparse representation suitable for water's irregular energy landscape by fitting to selected DFT configurations. A key mathematical formulation for these potentials is the learned energy function: E = f(\theta; \mathbf{R}) where \mathbf{R} denotes the atomic coordinates, f is the machine learning model (e.g., a neural network), and \theta are the optimized parameters from training on ab initio energies and forces. This surrogate approach bridges the gap between expensive simulations and empirical force fields, allowing for the study of water's structural and dynamic properties over extended timescales. Significant developments since 2020 include the machine learning reparameterization of the TIP4P four-site water model in 2021, which utilized automated molecular dynamics simulations to optimize geometric and charge parameters, improving predictions of density, diffusion, and radial distribution functions across temperatures from 273 to 373 K. In 2024, DeepMD was applied to saline solutions, where the model was trained on ab initio data for NaCl-water systems and demonstrated diffusion coefficients and structural properties matching the polarizable AMOEBA model while enabling simulations of anomalous water dynamics in electrolytes. That same year, a machine learning atomistic potential was developed specifically for the hematite-water interface, trained on DFT trajectories to accurately reproduce adsorption energies, water orientation, and proton transfer events at the mineral surface, facilitating studies of geochemical processes. Also in 2024, comprehensive benchmarks compared multiple machine learning potentials for water, such as kernel-based and Behler-Parrinello neural networks, highlighting their performance in reproducing thermodynamic properties like vapor-liquid coexistence and phase diagrams against DFT references. In November 2025, fine-tuning of machine learning interatomic potentials based on the MACE framework enabled accurate simulations of water and aqueous solutions, reproducing experimental thermodynamic and dynamical properties. These potentials excel in delivering ab initio-level accuracy for properties like self-diffusion coefficients (e.g., within 5% of for liquid water at 300 K) and vibrational spectra, while operating at speeds comparable to classical models, thus enabling simulations of complex environments such as interfaces and biomolecules. They also handle many-body effects implicitly through training data, surpassing explicit polarizable models in scalability for solvated systems. However, limitations include potential biases from finite training datasets, which may lead to inaccuracies in unexplored configurations like high-pressure phases, and the black-box nature of neural networks, complicating physical interpretability and transferability beyond the training regime.

Ab initio methods

Ab initio methods in water modeling rely on quantum mechanical calculations to describe the electronic structure and interactions without empirical parameters, providing a fundamental approach to simulating water's behavior at the molecular level. Density functional theory (DFT) is the most commonly employed framework for generating potential energy surfaces (PES) of water clusters and liquid phases, offering a balance between accuracy and computational feasibility. For instance, second-order Møller-Plesset perturbation theory (MP2) has been used to compute high-accuracy PES for small water systems, capturing correlation effects beyond mean-field approximations. In liquid water molecular dynamics (MD) simulations, functionals such as BLYP have been pivotal in early ab initio studies, while more recent meta-generalized gradient approximations like SCAN provide improved descriptions of hydrogen bonding and dispersion interactions. A primary application of these methods is ab initio molecular dynamics (AIMD), pioneered by the Car-Parrinello approach in 1985, which couples with fictitious Lagrangian dynamics to propagate both nuclear and electronic degrees of freedom efficiently. This technique has enabled simulations of liquid water's structural and dynamical properties, revealing insights into hydrogen bond networks and vibrational spectra. AIMD trajectories also serve as reference data for parameterizing higher-level models; for example, the many-body MB-Pol potential is trained on approximately 100,000 ab initio configurations derived from coupled-cluster calculations on water clusters, ensuring transferability to bulk phases. Recent advancements include comprehensive 2023 assessments of DFT functionals for reproducing liquid water's radial distribution functions and diffusion coefficients, highlighting the superior performance of hybrid functionals like over generalized gradient approximations in capturing short-range correlations. Hybrid quantum mechanics/molecular mechanics (QM/MM) methods extend ab initio accuracy to solvated protein environments by treating the reactive region quantum mechanically while modeling surrounding water classically, facilitating studies of enzyme mechanisms in aqueous media. The strengths of ab initio methods lie in their parameter-free nature, derived directly from the Schrödinger equation, and their ability to explicitly account for electronic effects such as charge transfer and polarization in hydrogen bonds, which are challenging for classical models. However, these approaches face significant limitations due to their high computational cost, which is roughly 10^6 times greater than classical MD, restricting simulations to systems of a few hundred water molecules over picosecond timescales. Emerging developments include the 2025 Open Molecules (OMol25) dataset, comprising over 100 million DFT calculations on diverse molecular systems including water-inclusive biomolecules and electrolytes, designed to train machine learning potentials with ab initio fidelity.

Performance and applications

Computational efficiency

The computational efficiency of water models in molecular dynamics simulations is largely determined by the scaling of pairwise interactions, which follow an O(N^2) complexity for N molecules due to the need to evaluate non-bonded forces between all pairs. Per water molecule pair, the cost increases with the number of sites, as electrostatic interactions require calculating distances between all charged sites (scaling as the square of charged site count), while Lennard-Jones (LJ) terms are typically limited to oxygen atoms (1 per pair) but still contribute to overall distance computations. In rigid non-polarizable models, three-site variants (e.g., TIP3P or SPC) involve 9 electrostatic distance calculations (3 charged sites × 3) and 1 LJ interaction (O-O) per pair, providing a baseline efficiency suitable for large-scale simulations. Four-site models (e.g., TIP4P) require 9 electrostatic terms but additional computations for the dummy site, increasing the per-pair cost by roughly 30-40% overall. Five-site models (e.g., TIP5P) demand 25 electrostatic calculations (5 charged sites × 5) and 1 LJ term, leading to ~2-3 times the cost of three-site models for equivalent system sizes. Six-site models involve even more terms (e.g., >36 electrostatic), further elevating costs. Flexible variants add intramolecular computations for bonds and angles, alongside high-frequency vibrations that necessitate smaller timesteps of 1-2 fs (versus 2 fs for rigid models) to ensure , effectively doubling or quadrupling the total computational demand for equivalent wall-clock time. Explicitly polarizable models, such as or SWM4-NDP, incorporate iterative self-consistent field procedures (typically 3-5 iterations per timestep) to compute induced multipoles, resulting in 5-20 times the cost of comparable fixed-charge rigid models, though optimized variants like OPC3-pol, which has computational efficiency between that of three- and four-site non-polarizable models, support larger timesteps (e.g., 4 fs). Many-body interaction models extend this further by including explicit (and higher) terms, requiring additional nested loops over molecular triplets or clusters, which can elevate costs by 10 times or more relative to pairwise rigid models. Machine learning-based potentials, such as representations of data, impose substantial upfront training expenses but achieve inference speeds roughly 10-100 times slower than classical fields, allowing access to quantum-level accuracy at reduced for production simulations compared to methods. Ab initio methods, relying on on-the-fly quantum mechanical evaluations (e.g., ), are 10^4 to 10^6 times slower than classical due to the exponential scaling of electronic structure calculations per timestep, limiting them to small systems and short trajectories. Common optimizations mitigate these demands across model types, including spherical cutoffs (8-12 ) for short-range van der Waals and , particle-mesh for long-range interactions, and GPU-accelerated implementations, which can accelerate simulations by 10-100 times depending on hardware and system scale.

Validation against experiments

Validation of water models against experimental data is essential to assess their reliability for simulating physical properties of liquid water, ice, and related phases. Key experimental benchmarks include the density (ρ ≈ 1 g/cm³ at 298 K), self-diffusion coefficient (D ≈ 2.3 × 10^{-9} m²/s at 298 K), static dielectric constant (ε ≈ 78 at 298 K), melting temperature (T_m = 273 K), viscosity (η ≈ 0.89 × 10^{-3} Pa·s at 298 K), and the oxygen-oxygen radial distribution function g(r) derived from neutron scattering experiments. These properties probe structural, thermodynamic, dynamic, and electrostatic aspects, with discrepancies highlighting model limitations in capturing hydrogen bonding, polarity, and many-body effects. Rigid non-polarizable models, such as three- and four-site variants, provide computational efficiency but often trade off accuracy in specific properties. The TIP3P model reproduces the liquid density well at ambient conditions (ρ ≈ 0.96 g/cm³ at 298 K) but overestimates the constant (ε ≈ 97) and severely underpredicts the (T_m ≈ 146 K), leading to poor description of phase behavior. In contrast, the TIP4P/2005 model excels in capturing density anomalies, such as the temperature of maximum density (≈ 277 K vs. experimental 277 K), and provides a more realistic (T_m ≈ 250 K), though its constant remains underestimated (ε ≈ 60). The five-site ST2 model accurately describes ice structure and hydrogen bonding via g(r), but its is excessively high (T_m ≈ 380 K), limiting applicability to low-temperature phases. Flexible and polarizable models address some shortcomings of rigid models by allowing intramolecular vibrations and induced dipoles, improving agreement with spectroscopic and electrostatic data. The polarizable model matches the experimental dielectric constant closely (ε ≈ 78 at 298 K) and reproduces () spectra, including vibrational frequencies and intensities for O-H stretching and bending modes, due to its explicit treatment of multipolar interactions. The MB-Pol many-body model, refined through data-driven approaches, provides the most accurate to date, correctly predicting the liquid-vapor coexistence curve, , and ice-liquid boundary within 1-2% of experiments in 2023 simulations, outperforming classical models in thermodynamic consistency. Emerging machine learning-based potentials, trained on data, enhance transferability and accuracy for complex environments. For instance, deep potential (DPMD) models from 2024 simulations match self-diffusion coefficients within 5% in bulk and solutions, capturing dynamic anomalies better than classical potentials. A 2025 benchmark of 44 models found classical four-site TIP4P-type models provide the best structural agreement with and scattering data, while hybrid machine learning-classical approaches performed poorly in reproducing total structure factors. As of 2025, benchmarks emphasize the continued relevance of optimized classical models for structural accuracy, with ML approaches showing promise in dynamics but requiring further efficiency gains. Recent benchmarks highlight improved g(r) predictions for top classical models, aligning closely with experimental in the first . Validation typically involves benchmarking against neutron diffraction for g(r), Raman/IR spectroscopy for vibrational properties, and dielectric relaxation measurements for ε, with assessments of transferability to ionic solutions revealing how models handle shells (e.g., TIP4P/2005 performs well for NaCl hydration but overestimates ion-water binding in polarizable variants like ). Persistent gaps include supercooled regimes below 250 K, where rigid models fail to capture enhanced fluctuations, and interfacial properties at vapor-liquid surfaces, though post-2020 models have improved predictions for these by incorporating quantum effects.

References

  1. [1]
  2. [2]
    Water models – Practical considerations for Molecular Dynamics
    A good water model must faithfully reproduce six bulk properties of water: Self diffusion coefficient, D Isobaric heat capacity, C p
  3. [3]
  4. [4]
    Systematic Comparison of the Structural and Dynamic Properties of ...
    Aug 18, 2021 · A Polarizable Model of Water for Molecular Dynamics Simulations of Biomolecules. ... Neural Network-Based Molecular Dynamics Simulation of Water ...Introduction · Methods · Results and Discussion · Supporting Information
  5. [5]
    Building Water Models: A Different Approach - PMC - NIH
    Simplified classical water models are currently an indispensable component in practical atomistic simulations. Yet, despite several decades of intense research, ...
  6. [6]
    Molecular Dynamics Study of Liquid Water - AIP Publishing
    Oct 1, 1971 · A sample of water, consisting of 216 rigid molecules at mass density 1 gm/cm 3 , has been simulated by computer using the molecular dynamics technique.
  7. [7]
  8. [8]
  9. [9]
    Comparison of water models for structure prediction - ScienceDirect
    Sep 1, 2025 · Describing the interactions of water molecules is one of the most common, yet critical, tasks in molecular dynamics simulations.Comparison Of Water Models... · 4. Results And Discussions · 4.3. Total Scattering...
  10. [10]
    Coarse-grained simulation of water: A comparative study and overview
    Apr 8, 2025 · Based on the number of interaction sites, these models are mainly classified into three point,28,29 four point,30–33 five point,34,35 and six ...I. Introduction · Ii. Simulation Details · Iii. Results
  11. [11]
    Site-renormalised molecular fluid theory: on the utility of a two ... - NIH
    We propose a simple, two-site model of water, using the familiar three-site Simple Point Charge (SPC) model as a guide.
  12. [12]
    A five-site model for liquid water and the reproduction of the density ...
    The ability of simple potential functions to reproduce accurately the density of liquid water from −37 to 100 °C at 1 to 10 000 atm has been further explored.
  13. [13]
    An intermolecular potential model for the simulation of ice and water ...
    Apr 22, 2003 · An intermolecular potential model of H 2 O with six interaction sites is proposed. The model is developed for the simulation of ice and water near the melting ...
  14. [14]
    How good are polarizable and flexible models for water
    We present a systematic analysis of state-of-the-art polarizable and flexible water models from a many-body perspective.Polarizable models · Analysis of many-body... · Harmonic frequencies of water...
  15. [15]
    A simple polarizable model of water based on classical Drude ...
    Sep 8, 2003 · A simple polarizable water model is developed and optimized for molecular dynamics simulations of the liquid phase under ambient conditions.Missing: paper | Show results with:paper
  16. [16]
    Quantum, intramolecular flexibility, and polarizability effects on the ...
    These simple potential function models represent a water molecule by a set of sites, either rigid or flexible relative to each other, that interact with a ...<|control11|><|separator|>
  17. [17]
    [PDF] Polarizable force fields Thomas A Halgren* and Wolfgang Damm
    In this article, we will review progress made in developing polarizable molecular force fields that seek to overcome limitations inherent in such approaches and ...
  18. [18]
  19. [19]
    TIP4P model of water page on SklogWiki - a wiki for statistical ...
    Jan 20, 2011 · The TIP4P model is a rigid planar four-site interaction potential for water, with a Lennard-Jones site for oxygen and three charge sites.
  20. [20]
  21. [21]
    An Empirical Polarizable Force Field Based on the Classical Drude ...
    Jan 27, 2016 · We review the classical Drude oscillator model, in which electronic degrees of freedom are modeled by charged particles attached to the nuclei of their core ...
  22. [22]
    Polarizable Water Models Conforming to Thole Polarization ... - NIH
    In the Drude-oscillator models, Drude particles are harmonically restrained to oscillate around heavy atoms to approximate polarization. The Drude-oscillator ...
  23. [23]
    Accurate predictions of the properties of water by a Gaussian charge ...
    From dimer to condensed phases at extreme conditions: Accurate predictions of the properties of water by a Gaussian charge polarizable model Available. Patrice ...INTRODUCTION · Description of the model · Predicted microstructure and...
  24. [24]
    Modeling Molecular Interactions in Water: From Pairwise to Many ...
    May 17, 2016 · (82) The PE model describes each water molecule as a single site carrying both the experimental static dipole and quadrupole moments along with ...Figure 3 · Figure 10 · 5 Conclusions And Future...
  25. [25]
    Realistic phase diagram of water from “first principles” data-driven ...
    Jun 8, 2023 · Fully derived from CCSD(T) data, MB-pol accurately predicts structural, thermodynamic, dynamical, and spectroscopic properties of water from gas ...
  26. [26]
    A systematic development of a polarizable potential of water
    May 31, 2013 · The bulk properties calculated with the BK3 model were obtained with a code applying the quaternion formalism65 but any other bond fixing method ...
  27. [27]
    Potential energy surfaces and dynamic properties via ab initio ...
    Feb 20, 2024 · This study examines the efficacy of density functional theory (DFT) and the correlation consistent Composite Approach (ccCA) in generating anharmonic ...
  28. [28]
    Computation of the methane–water potential energy hypersurface ...
    Ab initio methods were found to be more accurate than DFT methods, which do not account for the substantial dispersion interactions that exist between methane ...
  29. [29]
    Car–Parrinello molecular dynamics simulation of liquid water
    Jun 15, 2002 · New Car–Parrinello molecular dynamics simulations of the structural and dynamical properties of liquid water are reported.
  30. [30]
    Elevating density functional theory to chemical accuracy for water ...
    Nov 4, 2021 · We present density-corrected SCAN (DC-SCAN) calculations for water which, minimizing density-driven errors, elevate the accuracy of the SCAN functional.<|separator|>
  31. [31]
    Unified Approach for Molecular Dynamics and Density-Functional ...
    Nov 25, 1985 · We present a unified scheme that, by combining molecular dynamics and density-functional theory, profoundly extends the range of both concepts.Missing: water | Show results with:water
  32. [32]
    On the accuracy of the MB-pol many-body potential for water ...
    The MB-pol many-body potential has recently emerged as an accurate molecular model for water simulations from the gas to the condensed phase.
  33. [33]
    Structure and dynamics of liquid water from ab initio simulations
    Feb 15, 2024 · Here we assess the current status of describing liquid water using ab initio molecular dynamics, with a special focus on the performance of all the later ...
  34. [34]
    Evaluating the Chemical Reactivity of DFT-Simulated Liquid Water ...
    Oct 15, 2024 · We use the DD to explore the reactivity of water simulated using three different DFT functionals: the LDA functional (LDA), a hybrid-GGA functional (PBE0), and ...
  35. [35]
    Hybrid Quantum Mechanical/Molecular Mechanical Methods For ...
    May 9, 2023 · In this article, we briefly review the basic methodological details of QM/MM approaches and discuss their applications to various energy ...
  36. [36]
    Ab initio theory and modeling of water - PNAS
    The difficulty in modeling liquid water with DFT arises from the delicate nature of the H-bond network. An H bond is a directional attractive force between the ...
  37. [37]
    [PDF] Pushing the limit of molecular dynamics with ab initio accuracy to ...
    However, most AIMD applications are limited by computational cost to systems with thousands of atoms at most. We report that a machine learning- based ...
  38. [38]
    Computational Chemistry Unlocked: A Record-Breaking Dataset to ...
    May 14, 2025 · Open Molecules 2025, or OMol25, is a collection of more than 100 million 3D molecular snapshots whose properties have been calculated with density functional ...
  39. [39]
    Explicit Water Models Affect the Specific Solvation and Dynamics of ...
    The merits of the TIP3P model include its high and overestimated mobility, which accelerates simulation processes and thus effectively increases sampling. ACS ...
  40. [40]
    Simulating a flexible water model as rigid: Best practices and ...
    Apr 3, 2023 · It would be desirable to simulate these flexible water models as rigid, as such a practice will greatly improve computational efficiency.
  41. [41]
    What is the state-of-the-art in current flexible water models?
    Sep 30, 2020 · There are many rigid, fixed-charge models of water which rely on the Lennard-Jones potential (or something similar) to describe intermolecular interactions.Missing: k_bond k_angle
  42. [42]
    A fast polarizable water model for atomistic simulations - PMC - NIH
    We have developed a globally optimal polarizable water model, OPC3-pol, that explicitly accounts for electronic polarizability with minimal impact on ...
  43. [43]
    Reduced computational cost of polarizable force fields by a ...
    Apr 19, 2019 · With 3 fs time steps, two additional DC-JI/DIIS SCF iterations are required to maintain energy drift compared to that of 1 fs time steps.
  44. [44]
    Accurate and efficient molecular dynamics based on machine ...
    May 9, 2022 · We propose a molecular dynamics (MD) methodology which can simultaneously achieve both AIMD-level high accuracy and CMD-level high efficiency.
  45. [45]
    Comparing machine learning potentials for water: Kernel-based ...
    Mar 20, 2024 · In this paper, we investigate the performance of different machine learning potentials (MLPs) in predicting key thermodynamic properties of water using RPBE + ...
  46. [46]
    Static and Dynamic Correlations in Water: Comparison of Classical ...
    The computational cost of a PIMD simulation of liquid water significantly increases when the underlying Born-Oppenheimer potential energy surface is calculated ...
  47. [47]
    Fast, approximation-free molecular simulation of the SPC/Fw water ...
    Jul 16, 2024 · SPC/Fw water model. This paper benchmarks the event-chain Monte Carlo algorithm, molecular dynamics, and the Metropolis algorithm for molecular ...
  48. [48]
    Large-scale molecular dynamics simulations of bubble collapse in ...
    Dec 14, 2023 · The simulations for the TIP4P/2005 water model, which accurately predicts the water viscosity, yield much elongated collapse dynamics where ...Ii. Methods · B. Md Simulations · D. Bubble Dynamics In...<|control11|><|separator|>
  49. [49]
    Transport and dielectric properties of water and the influence of ...
    Feb 12, 2014 · Particularly for TIP3P, widely differing values for the dielectric constant can be found in the literature;18,25–27 in order to address these ...
  50. [50]
    The melting temperature of the most common models of water
    For TIP5P, Nada and van der Eerden51 have reported a melting temperature of T = 274 K and recently Koyama et al.97 reported T = 268 K ⁠. Thus, our results are ...
  51. [51]
    General Purpose Model for Condensed Phases of Water: TIP4P/2005
    TIP4P/2005 is a rigid four site model which consists of three fixed point charges and one Lennard-Jones center.
  52. [52]
    [PDF] Melting points of water models: Current situation
    We revisited the melting point of TIP4P/2005 and TIP5P obtaining Tm = 250 and 274 K, respectively.
  53. [53]
    Second critical point in two realistic models of water - Science
    Jul 17, 2020 · Here, we use histogram reweighting and large-system scattering calculations to investigate computationally two molecular models of water, TIP4P/ ...
  54. [54]
    A new method to calculate broadband dielectric spectra of solvents ...
    Aug 12, 2024 · Total dielectric spectrum (black) of water using the AMOEBA water model and dissected into contributions from induced (red) and static (blue) ...
  55. [55]
    Spatiotemporal characterization of water diffusion anomalies in ...
    Dec 11, 2024 · Our research used DPMD to study how salts affect water by comparing its results with ab initio molecular dynamics, SPC/Fw, AMOEBA, and MB-Pol models.
  56. [56]
    Transferable performance of machine learning potentials across ...
    Nov 19, 2024 · This study evaluates the transferability of the deep potential (DP) and neural equivariant interatomic potential (NequIP) models for graphene– ...
  57. [57]
    [2505.23446] Comparison of water models for structure prediction
    May 29, 2025 · Forty-four classical water potential models are compared to identify those that can accurately describe the structure in alignment with experimental results.