betaFeedback

WorksPhysics and Astronomy

Dynamic heterogeneity in sodium silicate melts via machine-learning potential

Abstract

We present a comprehensive characterisation of dynamic heterogeneity in sodium silicate melts using molecular dynamics simulation with machine-learning potentials. By studying sodium disilicate, tetrasilicate, and hexasilicate melts across a range of temperatures, mean squared displacement and a time-correlation function computed up to the nanosecond timescale provide a detailed account of how spatial mobility disparities emerge in a realistic multicomponent oxide glass. Within these timescales, the self-part of the van Hove function for sodium displays a bimodality, demonstrating that alkali transport is mediated by discrete displacement events consistent with a hopping mechanism. This distinct hopping allows sodium ions to decouple from the sluggish relaxation of the silicate matrix. Furthermore, evaluation of the non-Gaussian parameter reveals that, although all constituent species exhibit dynamic heterogeneity, the non-Gaussian behaviour is most pronounced for oxygen atoms. This trend reflects the intermittency of structural rearrangements, where framework atoms undergo rare and stochastic events compared to the frequent displacements of mobile ions. Our findings elucidate the microscopic mechanism of ion transport and its connection to dynamic heterogeneity in silicate melts, offering a new avenue to study fundamental glassy physics in realistic vitreous materials.

Introduction

Ion diffusion in alkali-doped silica glasses is a fundamental problem of broad relevance, spanning chemical physics, materials science, and geology[1, 2, 3]. Its importance has been further highlighted by applications in advanced industrial products such as all-solid-state secondary batteries, making it a timely and active area of research[4], where precise control of ion transport is paramount. The mechanism by which alkali ions relax within the disordered silicate network is a key factor governing the macroscopic properties of these materials. At the heart of this context is the preferential pathway picture, which posits that alkali ions do not diffuse uniformly through the SiO2\mathrm{SiO}_{2} matrix but instead migrate through spatially distinct, interconnected channels that form within the network[5]. To date, numerous studies have focused on verifying the static structural aspects of this framework, observing structural changes that support this depiction as sodium density increases[6, 7, 8]. However, since ion transport is inherently a dynamical phenomenon, there is a pressing need to characterise these processes from the perspective of relaxation dynamics through direct observation via computer simulations[9, 10, 11, 12].

Such a characterisation is most naturally pursued within the broader framework of glass physics, where relaxation dynamics have long been a central concern[13, 14, 15]. In supercooled liquids approaching the glass transition, particles of high and low mobility coexist and fluctuate spatially, a phenomenon known as dynamic heterogeneity[16]. These dynamical fluctuations grow as the system cools and relaxation slows, giving rise to hallmark features of the glass transition such as stretched exponential decay of time correlation functions[17, 18, 19], exponential distribution tails of single-particle displacements[20], and four-point susceptibility peaks[21, 22, 23]. Moreover, dynamic heterogeneity is not merely of theoretical interest; it governs transport properties, including macroscopic diffusion[24, 25, 26] and ionic conductivity[27, 28], and is therefore of direct relevance to the ion migration picture outlined above. Nevertheless, detailed insights into dynamic heterogeneity have been predominantly obtained from packing-type glass models, in which interparticle interactions are isotropic and short-ranged[18, 29, 30, 31]. The correlation between this phenomenon and ion diffusion in network glasses such as sodium silicate, where directional covalent bonding gives rise to qualitatively richer local environments, remains to be fully elucidated. It is therefore essential to accurately capture individual particle dynamics within the diverse local environments that are far more pronounced in network glasses, typified by sodium silicate, than in simple packing-type systems. From this perspective, it is crucial to go beyond simple empirical potentials[8, 9, 11] and track relaxation with the precision of first-principles calculations that incorporate the underlying electronic states. However, as is well known, first-principles calculations are computationally intensive, making it increasingly impractical to cover the growing relaxation times near the glass transition as the temperature decreases. More critically, the computational cost of first-principles methods scales as O(N3)O(N^{3}) with respect to the number of particles. This scaling is fundamentally at odds with the requirement for large system sizes necessary to capture dynamic heterogeneity in glass transition research.

In this study, we overcome this computational barrier by utilising machine learning potentials, which have undergone remarkable methodological advances[32, 33]. Such potentials have also enabled detailed and transferable atomistic modelling of the silicon-oxygen system, spanning crystalline, amorphous, and nanostructured environments[34, 35], including alkali silicate glasses closely related to the compositions studied here[36, 37, 38]. Whereas these prior works on alkali silicate glasses have focused primarily on establishing the accuracy and transferability of machine-learning force fields through structural validation, the present study extends this approach to the long-time dynamical regime. Specifically, through the construction of a machine learning potential for sodium silicate and a molecular dynamics (MD) code developed in-house and optimised for graphics processing unit execution, we have succeeded in dynamically tracking the relaxation of both the frozen silica network and the diffusive sodium ions over extended spatial and temporal scales previously inaccessible to first-principles methods. This computational strategy, which leverages the architectural characteristics of modern hardware, enables the application of analytical methods developed in fundamental glass transition research to realistic sodium silicate systems. We provide compelling evidence for the preferential pathway picture from a dynamical perspective, supported by complementary structural evidence presented in the Appendix, establishing a new foundation for the study of realistic glass systems.

Methods

The machine-learning interatomic potential (MLIP) was constructed using a PaiNN-type equivariant graph neural network[39]. Atomic features were embedded in a 60-dimensional latent space, and interatomic interactions were evaluated within a cutoff radius of 5.0 Å using 40 Gaussian radial basis functions combined with a smoothstep cutoff envelope. The network architecture comprised two interaction blocks.

The model parameters were optimised by minimising a weighted loss function incorporating total energies, atomic forces, and cell virials:

L=wELE+wFLF+wΞLΞ(1)\mathcal{L} = w_E \mathcal{L}_E + w_F \mathcal{L}_F + w_\Xi\mathcal{L}_\Xi \tag*{(1)}

where the weighting factors were set to wE=0.01w_E = 0.01, wF=0.99w_F = 0.99, and wΞ=0.01w_\Xi= 0.01. The energy, force, and cell-virial contributions were evaluated as mean squared errors relative to the reference first-principles values. To ensure high-fidelity MD trajectories, a substantially larger weight was assigned to the force-loss term.

The training dataset was generated from ab initio molecular dynamics melt-quench trajectories of sodium silicate glasses. Initial configurations were prepared for four Na2_2O fractions, corresponding to Na6_6Si57_{57}O117_{117}, Na12_{12}Si54_{54}O114_{114}, Na24_{24}Si48_{48}O108_{108}, and Na48_{48}Si36_{36}O96_{96} in 180-atom cells, each prepared at four mass densities of 1.8 g cm−3^{-3}, 2.0 g cm−3^{-3}, 2.2 g cm−3^{-3}, and 2.4 g cm−3^{-3}, yielding 16 density-composition conditions in total.

The ab initio molecular dynamics simulations were performed using VASP[40, 41, 42] with the PBE functional[43] and PAW potentials[44]. The plane-wave cutoff energy was set to 400 eV, and only the Γ\Gamma-point was sampled in kk-space. Each system was first run at 5000 K for 3 ps and subsequently quenched to 300 K over 4 ps using a 1 fs time step.

Atomic configurations were sampled every 10 MD steps from both the high-temperature equilibration and the subsequent quenching trajectories, yielding 699 configurations for each density–composition condition and a total dataset of 11,184 structures. For each sampled configuration, the total energy, atomic forces, and cell virials were recomputed through static single-point DFT calculations with accurate precision settings.

The resulting dataset was randomly partitioned into training (80 %) and validation (20 %) sets. The MLIP was trained for 500 epochs using the AdamW optimiser[45] with an initial learning rate of 10−310^{-3}, which was reduced by a factor of 0.5 every 100 epochs using a step scheduler. The validation accuracy of the trained potential is presented in Appendix A.

Production runs with the trained MLIP were carried out for sodium silicate systems containing 3000 atoms. The linear size of the simulation box was determined by a mass density of 2.30 g cm−3^{-3}. Three distinct compositions were investigated: Na666_{666}Si667_{667}O1667_{1667}, Na400_{400}Si800_{800}O1800_{1800}, and Na284_{284}Si858_{858}O1858_{1858}, which approximately correspond to Na2_2O · 2SiO2_2, Na2_2O · 4SiO2_2, and Na2_2O · 6SiO2_2 (hereafter designated as NS2, NS4, and NS6, respectively). For each composition, three independent initial configurations were generated via random packing.

The systems were initially equilibrated in the liquid state at 4000 K for 20 ps using the Bussi thermostat[46], and subsequently cooled to 300 K at a constant cooling rate of 101210^{12} K s−1^{-1}. During this cooling process, instantaneous configurations were sampled at target temperatures of 4000 K, 3400 K, 2900 K, 2500 K, 2100 K, 1900 K, 1700 K and 1500 K. At each temperature, an additional NVT run of 20 ps was performed using the Nosé–Hoover thermostat[47, 48], followed by production runs in the NVT ensemble. The durations of production runs were: 5.1 ns for 1500 K, 3.2 ns for 1700 K, 1.3 ns for 1900 K and 2100 K, 0.51 ns for 2500 K, 0.32 ns for 2900 K, and 0.13 ns for 3400 K and 4000 K. These extended production windows were selected to cover the relevant structural relaxation timescales, particularly for the network-forming species at lower temperatures.

Bulk relaxations

Mean squared displacement

In this section, we discuss the bulk relaxation of the system and its characteristic relaxation time. We first examine the mean squared displacement (MSD), which is defined as follows:

⟨r2(t)⟩=⟨1Nα∑i=1Nα∣ri(t)−ri(0)∣2⟩(2)\left\langle r^2(t) \right\rangle= \left\langle\frac{1}{N_{\alpha}} \sum_{i=1}^{N_{\alpha}} \left| \mathbf{r}_i(t) - \mathbf{r}_i(0) \right|^2 \right\rangle \tag*{(2)}

where NαN_{\alpha} represents the number of α\alpha-atoms in the system. MSD was evaluated for each atomic species in the system for every composite and temperature.

As illustrated in Figure 1 (a–Figure 1c) for the NS4 composition, at the highest temperature T=4000T = 4000 K, a smooth transition from the short-time ballistic behaviour to long-time diffusive behaviour is observed for all atomic species (Si, O, and Na). Upon cooling, the MSD curves exhibit characteristic changes. For the Si and O atoms composing the SiO2_2 framework, a distinct plateau emerges at intermediate timescales. This feature of MSD represents the two-step relaxation process, a hallmark of the glass transition phenomenon[18]. Even at low temperatures where this plateau is pronounced, our time window covers the diffusive regime of MSD of Si and O atoms at sufficiently long times. In contrast, the relaxation of Na atoms is considerably more rapid than that of the species of SiO2_2 framework. Notably, even at the lowest temperature investigated, T=1500T = 1500 K, an obvious plateau does not emerge in the MSD of Na, unlike the cases of Si or O. Nevertheless, as the temperature decreases, its MSD begins to exhibit incipient signs of a plateau in the intermediate time regime t≈1t \approx1 ps, reflecting the onset of slow dynamics of this species at lower temperatures. Taken together, these results demonstrate that our simulations succeed in characterising the relaxation dynamics of a complex oxide melt over nanosecond timescales. This temporal regime remains entirely inaccessible to ab initio molecular dynamics and has only recently begun to be reached by machine-learning potentials for simpler supercooled liquids, such as toluene[49].

Figure 1. Mean squared displacements ⟨r2(t)⟩\langle r^2(t)\rangle (panels (a–c)) and bond-break correlation CB(t)C_B(t) (panels (d–f))of NS4 composite. Panels (a, d), (b, e), and (c, f) represent data from Si, O, and Na atoms, respectively.

Bond-break correlation function

Next, we present the results for the bond-break correlation function[50, 51]

CB(t)=⟨1Nα∑i=1NαNBi(t∣0)NBi(0)⟩(3)C_B(t) = \left\langle\frac{1}{N_{\alpha}} \sum_{i=1}^{N_{\alpha}} \frac{N_B^i(t|0)}{N_B^i(0)} \right\rangle \tag*{(3)}

where NBi(0)N_B^i(0) represents the number of neighbouring α\alpha-atoms at time t=0t = 0 and NBi(t∣0)N_B^i(t|0) represents the number of the neighbouring α\alpha-atoms that remain bonded at time tt. The neighbouring two particles ii and jj are determined by whether their distance rijr_{ij} is smaller than a threshold value of their species α\alpha: rij<rth,αr_{ij} < r_{\mathrm{th},\alpha}. The thresholds rαr_{\alpha} are selected as the first minimum of the pair correlation function (shown in Supplementary Information), namely rth,Si=3.6r_{\mathrm{th,Si}} = 3.6, rth,O=3.2r_{\mathrm{th,O}} = 3.2, and rth,Na=5.0r_{\mathrm{th,Na}} = 5.0. The threshold at time tt is set to 1.1rth,α1.1r_{\mathrm{th},\alpha}. This function represents the probability that pairs of atoms, initially located within each other’s proximity, remain within a specified threshold distance at time tt. This correlation function is shown

to serve as a good time-correlation function for capturing relaxations of supercooled liquids[52].

CB(t)C_B(t) is shown in Figure 1 (d–Figure 1f) for the NS4 composition in correspondence with the MSD. The temporal evolution of CB(t)C_B(t) exhibits a trend consistent with that of the MSD. At high temperatures, a single-step relaxation is evident, where the correlations for all atomic species (Si, O, and Na) vanish within approximately 10 ps. For Si and O, the substantial slowing down of relaxation at lower temperatures is clearly captured by CB(t)C_B(t). In the low-temperature regime, a well-defined plateau develops for these species, indicative of the two-step relaxation process that is a typical characteristic of the glass transition[18]. At short times, CB(t)C_B(t) undergoes an initial decay associated with the β\beta relaxation process before entering the plateau region. This first decay to the plateau is notably more pronounced for O than for Si. The higher clarity of the first relaxation step and the plateau for oxygens can be attributed to the specific local environment of the species; as oxygen atoms constitute the vertices of the SiO4_4 tetrahedra, they are more directly susceptible to the local cage effect formed by neighbouring atoms of the same type. Consequently, the short-time β\beta relaxation, which is considered as thermal vibrations inside cages[18], and caging by surrounding atoms are more readily observable for O atoms than for central Si atoms. On the other hand, at sufficiently long times, the bond-break correlation eventually decays to zero (α\alpha relaxation), which corresponds to the transition to the diffusive regime observed in the MSD.

In contrast, the slowdown in the relaxation of Na atoms is relatively modest, even as the temperature decreases to the regime where relaxations of Si and O atoms are already slow. At the lowest temperature investigated (T=1500T = 1500 K), while the Si and O correlations persist for roughly 1 ns before decaying, the Na correlation is lost within 10 ps to 100 ps, approximately one to two orders of magnitude faster than the silica framework. Nevertheless, a discernible slowdown in the Na dynamics is still observed at low temperatures (e.g. T=1500T = 1500 K), with incipient signs of a plateau appearing at t≈1t \approx1 ps.

Structural relaxation time

Utilising the bond-break correlation functions, we now define the structural relaxation time τα\tau_{\alpha} as CB(τα)=e−1C_B(\tau_{\alpha}) = e^{-1}. Figure 2 presents the temperature dependence of these relaxation times for each constituent species. Across all investigated compositions, the increase of relaxation times of the network-forming Si and O atoms is manifestly more sluggish than that of the mobile Na ions. The disparate sensitivity of these relaxation times to temperature, evident in the steeper gradients of τα\tau_{\alpha} for Si and O, indicates that the dynamic decoupling between the SiO2_2 framework and the alkali ions becomes increasingly pronounced as the temperature decreases. Within the network itself, a clear hierarchy is observed, with Si atoms consistently exhibiting longer relaxation times than O atoms. Notably, the reliable extraction of τα\tau_\alpha for Si and O down to T=1500T = 1500 K was made possible only by the extended simulation timescale afforded by the present machine-learning potential, as such temperatures correspond to relaxation times beyond the reach of earlier first-principles studies.

Figure 2. Relaxation time τα\tau_\alpha of all compositions. For Si atoms at the lowest temperature, CB(t)C_B(t) does not decay to the threshold value e−1e^{-1} within the simulation window (see Figure 1 (d) for the NS4 case); these data points are therefore absent.

When comparing the dynamics across the various compositions, it is observed that an increase in sodium leads to a systematic acceleration of the dynamics for all atomic species. This compositional influence on the relaxation behaviour is most salient in the low-temperature regime. Conversely, at high temperatures, the disparity between the compositions diminishes; notably, the relaxation times for Na converge, displaying a near-total independence from the specific composition. In contrast, the dynamics of Si and O remain sensitive to the network environment even in the high-temperature limit, maintaining a degree of compositional dependence that is absent in the alkali ion.

Dynamic heterogeneity

The van Hove function

To further elucidate the nature of the relaxation processes, we now turn our attention to the spatial heterogeneity of the relaxation. We first examine the spatial distribution of atomic displacements via the self-part of the van Hove distribution function[53], defined as:

Gsα(r,t)=⟨1Nα∑i=1Nαδ(r−∣ri(t)−ri(0)∣)⟩(4)G_s^\alpha(r,t)=\left\langle\frac{1}{N_\alpha}\sum_{i=1}^{N_\alpha}\delta\left(r-\left|\mathbf{r}_i(t)-\mathbf{r}_i(0)\right|\right)\right\rangle \tag*{(4)}

where δ(x)\delta(x) denotes the Dirac delta function. In the following discussion, the distribution is presented as 4πr2Gsα(r,t)4\pi r^2G_s^\alpha(r,t) to account for the three-dimensional volume element.

Figure 3 illustrates the temporal evolution of this distribution for the NS4 composite at 1500 K. At short times, the distributions show the Gaussian property for each atom. On the other hand, at intermediate and long timescales, the network-forming species, namely Si and O atoms (panels (a) and (b)), exhibit the distributions with pronounced non-Gaussian tails. These tails signify a substantial departure from the Gaussian behaviour typical of simple liquids, indicating that a subpopulation of these atoms undergoes displacements significantly larger than the typical distribution.

Figure 3. Self-part of the van Hove distribution function for the NS4 composite at 1500 K. Panels (a), (b), and (c) represent data for Si, O, and Na species, respectively. Each data corresponds to 0.122 ps, 1.230 ps, 12.299 ps, 122.995 ps, 308.949 ps, 776.046 ps and 1949.339 ps for Si and O atoms and 0.122 ps, 0.195 ps, 0.489 ps, 0.775 ps, 1.230 ps and 4.896 ps for Na atoms.

In contrast, the mobile Na ions (panel (c)) display a markedly different dynamical profile. At extended timescales, the distribution exhibits a discernible bimodality, a hallmark of discrete hopping dynamics, wherein the secondary peak corresponds to ions that have successfully escaped their local coordination cages and migrated to adjacent interstitial sites. This peak is located at approximately r≈3.5r \approx3.5 Å, consistent with the typical Na–Na nearest-neighbour distance derived from the partial pair correlation function (Supplementary Information), confirming that these displacement events correspond to discrete jumps between adjacent sites within the silicate network. The emergence of such bimodality provides direct evidence for the coexistence of relatively immobile, caged ions and highly mobile, jumping ions, and is a key manifestation of dynamic heterogeneity in alkali silicate systems.

Non-Gaussian parameter

In the previous section, our analysis of the van Hove distribution revealed the non-Gaussian nature of the relaxation and the separation of mobile and immobile atoms within the sodium silicate. To quantitatively characterise this heterogeneity, we next evaluate the non-Gaussian parameter α2(t)\alpha_2(t)⁵⁴. In three dimensions, this is defined as:

α2(t)=3⟨r4(t)⟩5⟨r2(t)⟩2−1.(5)\alpha_2(t) = \frac{3\langle r^4(t)\rangle}{5\langle r^2(t)\rangle^2} - 1. \tag*{(5)}

This parameter vanishes for a purely Gaussian process, and its magnitude serves as a probe of the deviation from the Gaussian distribution, thereby reflecting the spatial heterogeneity in particle mobilities.

Figure 4 presents the evolution of α2(t)\alpha_2(t) for Si, O, and Na across all investigated compositions (NS2, NS4, and NS6). For all species and compositions, α2(t)\alpha_2(t) exhibits a characteristic peak at intermediate timescales within our time window. The peak height increases systematically as the temperature is lowered, underscoring the intensification of dynamic heterogeneity as the system approaches the glass transition. The position of this peak shifts towards longer timescales at lower temperatures, mirroring the overall slowing down of the structural relaxation. We note, however, that the characteristic timescale of the α2(t)\alpha_2(t) peak remains consistently shorter than the structural relaxation time τα\tau_\alpha defined from the time-correlation function, as observed in other common glass-forming models¹⁸,⁵⁵⁻⁵⁹.

Figure 4. Temperature and compositional dependence of the non-Gaussian parameter α2(t)\alpha_2(t) for every constituent species. Panels (a–c), (d–f), and (g–i) represent data for Si, O, and Na species, respectively. Panels (a), (d), and (g) represent data for NS2, panels (b), (e), and (h) represent data for NS4, and panels (c), (f), and (i) represent data for NS6. The legend of each panel is the same as that in Figure 1.

A distinct compositional trend is also evident. As the sodium content decreases from NS2 to NS6, corresponding to an increase in the connectivity of the silicate network (see Appendix B), the maximum value of α2(t)\alpha_2(t) increases for all atomic species. This indicates that a reduction in modifier concentration enhances the spatial constraints imposed by the rigid framework, thereby amplifying the intermittency of relaxation events and separation of mobile and immobile particles.

Notably, the peak intensity of α2(t)\alpha_2(t) for oxygen is consistently the most pronounced among all constituent species across the investigated range. In contrast, the magnitudes for silicon and sodium are comparable, with the sodium peaks often reaching or exceeding those of silicon. The elevated non-Gaussianity in the oxygen dynamics stems from the structural and chemical diversity of its local environment. While the central silicon atoms are rigidly anchored within the SiO4_4 units, the vertex oxygen atoms possess greater vibrational and rotational flexibility, allowing for more diverse local relaxation modes. Furthermore, the coexistence of bridging and non-bridging oxygens in the presence of sodium modifiers introduces a stark contrast in local mobilities, thereby broadening the overall distribution of relaxation times for the oxygen species.

For sodium, the robust peak intensity is intrinsically linked to its specific transport mechanism. As evidenced by the bimodality in the self-part of the van Hove functions, which is a feature absent in the framework atoms, sodium ions undergo discrete hopping events between interstitial sites. The significant magnitude of α2(t)\alpha_2(t) for sodium thus reflects the highly intermittent nature of these hops, representing a sharp departure from continuous Gaussian diffusion. Collectively, these results indicate that the dynamical heterogeneity in sodium silicate melts arises from two distinct origins: the infrequent and stochastic rearrangements of the covalent network and the rapid, channel-mediated hopping of the alkali modifiers.

Conclusions

In this study, we have characterised the dynamical heterogeneity in sodium silicate melts across a wide range of temperatures and compositions through the application of machine-learning potentials with ab initio accuracy. The primary significance of this work lies in our ability to extend high-precision simulations to timescales that fully encompass the structural relaxation of the silicate framework. This has enabled a definitive account of how spatial heterogeneities in mobility manifest in a realistic multi-component oxide glass, bridging the gap between first-principles accuracy and the underlying physics behind the complex relaxations near the glass transition. The structural origins of this behaviour are detailed in the Appendix B, where we demonstrate that the partial structure factors and Q(n)Q(n) species distributions are in good agreement with prior experimental and simulation studies of sodium silicate systems[8, 9]. In particular, the growth of the prepeak in SNaNa(k)S_{\mathrm{NaNa}}(k) with increasing sodium content and the systematic shift of the Q(n)Q(n) distribution towards fragmented species provide direct structural evidence for the depolymerisation of the silicate network and the formation of sodium-rich preferential channels, corroborating the dynamical picture presented in the main text.

Our results provide a robust characterisation of dynamic heterogeneity through the systematic evaluation of the non-Gaussian parameter α2(t)\alpha_2(t). We have demonstrated that the intensity and timescale of these dynamical fluctuations are highly species-dependent, with the oxygen atoms consistently exhibiting the most pronounced non-Gaussianity. The magnitude of α2(t)\alpha_2(t) for sodium ions is found to be comparable to, and at certain temperatures exceeds, that of the central silicon atoms. This observation reflects the distinct origins of heterogeneity in this system: the rare, stochastic rearrangements of the rigid silicate framework and the discrete, intermittent hopping events of the alkali modifiers within the interstitial space.

The achievement of ab initio-level accuracy was instrumental in capturing the subtle interplay between the structural relaxation of the matrix and the motion of the modifiers. We identified a bimodality in the self-part of the van Hove functions for sodium, providing evidence of hopping-mediated transport that persists even as the surrounding framework approaches a frozen state. This decoupling, where the alkali dynamics remain non-Gaussian while the framework undergoes slow structural relaxation, represents a fundamental characteristic of the relaxations of modified silicate melts that can be accurately resolved through the high-accuracy molecular dynamics simulation.

In conclusion, our work demonstrates that the characterisation of dynamical heterogeneity in realistic glass-forming systems can be elevated to the level of near-ab initio accuracy. By reaching the structural relaxation regime while maintaining the accuracy of the underlying atomic interactions, we have established a rigorous basis for understanding how local constraints and chemical environments dictate the macroscopic transport properties of multi-component oxide glasses. These insights offer a new standard for the computational design and analysis of ion-conducting materials where the precise control of dynamical fluctuations is of paramount importance.

Acknowledgments

This work is supported by JSPS KAKENHI Grant Number 25H01478, JST FOREST Grant Number JP-MJFR236Q, and a grant from the Inamori Foundation. The computation was performed using computational facilities at the Research Center for Computational Science, Okazaki, Japan (Project: 26-IMS-C117).

Conflict of Interest

The authors have no conflicts to disclose.

Author contributions

K. Shiraishi: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Validation, Visualization, Writing - original draft, Writing - review & editing; R. Nozawa: Software, Validation, Writing - review & editing; E. MinamitanI: Conceptualization, Data curation, Funding acquisition, Methodology, Project administration, Resources, Supervision, Visualization, Writing - review & editing.

Model accuracy

To assess the accuracy of the trained machine-learning interatomic potential, we compare the model predictions against the reference DFT values on the held-out validation set. Figure 5 shows the parity plots for atomic forces, energies per atom, and cell virials. The close agreement between predicted and reference values across the full range of the validation set demonstrates that the trained potential reproduces the underlying first-principles energy, force, and virial surfaces with good accuracy.

Figure 5. Parity plot of predicted versus reference atomic forces, energies per atom, and cell virials (from left to right) on the held-out validation set.

Static structures

In this appendix, we present a structural characterisation of the sodium silicate melts that complements the dynamical analysis presented in the main text. We examine the partial structure factors Sαβ(k)S_{\alpha\beta}(k) and the distribution of bond-bridging oxygen species Q(n)Q(n) across the investigated temperatures and compositions, providing the structural basis for the topological fragmentation and kinetic decoupling discussed in Secs. III and IV.

Static structure factors

We begin our structural analysis by examining the partial structure factors Sαβ(k)S_{\alpha\beta}(k) between α\alpha and β\beta atoms, which provide a quantitative description of the spatial correlations in the sodium silicate melt. The structure factor Sαβ(k)S_{\alpha\beta}(k) is defined as[53]

Sαβ(k)=1NαNβ⟨ρα(k)ρβ(−k)⟩,(6)S_{\alpha\beta}(k) = \frac{1}{\sqrt{N_\alpha N_\beta}} \langle\rho_\alpha(k)\rho_\beta(-k) \rangle, \tag*{(6)}

where ρα(k)=∑j=1Nαexp⁡(ik⋅rj)\rho_\alpha(k) = \sum_{j=1}^{N_\alpha} \exp(ik \cdot r_j) is the Fourier transform of the microscopic density at wave vector kk. We use the Freud package[60] for this calculation. Figure 6 illustrates the evolution of these correlations as a function of temperature and composition.

Figure 6. Partial structure factors Sαβ(k)S_{\alpha\beta}(k) of NS2. Panels (a) and (b) show the temperature dependence of SOO(k)S_{\mathrm{OO}}(k) and SNaNa(k)S_{\mathrm{NaNa}}(k) (panel (a)) and SSiO(k)S_{\mathrm{SiO}}(k) and SNaO(k)S_{\mathrm{NaO}}(k) (panel (b)). Panels (c) and (d) show the corresponding compositional dependence at 1500 K.

The temperature dependence of the partial structure factors, presented in Figure 6 (a) and (Figure 6b) for the NS2 composite, reveals a systematic intensification and sharpening of principal peaks as the system is cooled from 4000 K to 1500 K. This trend signifies a progressive enhancement of structural order and a reduction in thermal fluctuations, allowing the underlying glassy framework to emerge with increasing definition. Notably, the peaks associated with the network-forming species, specifically the SOO(k)S_{\mathrm{OO}}(k) maximum and the pronounced SSiO(k)S_{\mathrm{SiO}}(k) minimum, exhibit a robust development, suggesting that the silicate framework establishes its fundamental topological features well above the glass transition. In contrast, the correlations involving Na ions, SNaNa(k)S_{\mathrm{NaNa}}(k) and SNaO(k)S_{\mathrm{NaO}}(k), whilst also sharpening upon cooling, remain relatively broader than those of the SiO2_2 framework even at the lowest temperature. This persistence of the broad distribution of S(k)S(k) indicates that the alkali modifiers retain a significant degree of local structural flexibility and disordered liquid-like structure within the interstitial voids, even as the surrounding SiO2_2 network proceeds to structural freezing.

The compositional influence on the structural arrangement, shown in Figure 6 (c) and (Figure 6d), highlights the dual role of Na as both a network modifier and a self-organising species. A particularly striking feature is the emergence and systematic growth of a prepeak in SNaNa(k)S_{\mathrm{NaNa}}(k) at k≈1k \approx1 Å−1^{-1} as the sodium fraction increases from NS6 to NS2. This low-kk feature constitutes direct evidence for intermediate-range order, demonstrating that the Na ions do not disperse randomly but instead organise into spatially correlated regions or channels within the silica matrix⁸. Simultaneously, the intensification of the Na–Na prepeak is accompanied by a discernible damping of the SOO(k)S_{\mathrm{OO}}(k) and SSiO(k)S_{\mathrm{SiO}}(k) correlations. This decrease reflects the progressive de-polymerisation of the silicate framework; as the alkali concentration rises, the Na ions act to sever the Si–O linkages, thereby reducing the overall connectivity of the network. This structural degradation is further corroborated by the enhancement of the SNaO(k)S_{\mathrm{NaO}}(k) correlation, which indicates an increased population of Na ions coordinating with oxygens that are not bridging Si. Taken together, these observations suggest a spatial segregation wherein the alkali ions actively carve out their own preferential pathways by locally disrupting the covalent network, a structural precursor to the kinetic decoupling discussed in Secs. III and IV.

Bond-bridging oxygens

To further elucidate the topological nature of the silicate framework, we now examine the distribution of Q(n)Q(n) species. Q(n)Q(n) is defined as the fraction of SiO4_4 tetrahedra possessing nn bridging oxygens and is regularly measured in both experimental⁶¹⁻⁶³ and numerical studies⁹. Figure 7 illustrates a compositional trend in this distribution across the investigated temperatures.

Figure 7. Bond-bridging oxygen Q(n)Q(n) of every composites.

As the sodium concentration increases from NS6 to NS2, there is a pronounced shift from the predominantly fully-connected Q(4)Q(4) environment, typical of pure silica, towards fragmented Q(3)Q(3) and Q(2)Q(2) species. This distribution shift provides quantitative evidence for the systematic de-polymerisation of the silica framework driven by the alkali ions. This fragmented structure, particularly prominent in NS2 with high Q(3)Q(3) and Q(2)Q(2) populations, results in a more open network with increased interstitial volume. These structural features represent the local realisation of the preferential pathways or channels implied by the SNaNa(k)S_{\mathrm{NaNa}}(k) prepeak discussed in the previous section (Figure 6 (c)). Such a framework serves as a direct structural precursor for the accelerated bulk relaxation observed in sodium-rich composites, and critically for the remarkable decoupling of sodium ions from the slower silica matrix. Moreover, the co-existence of disparate local Q(n)Q(n) environments provides a strong structural basis for the high degree of dynamic heterogeneity observed for framework atoms, with a non-uniform spatial distribution of local constraints.

Regarding the temperature dependence, while rapid variations in Q(n)Q(n) populations are evident in the high-temperature regime, there is a clear trend towards a suppression of the variation as the temperature decreases from 4000 K to 1500 K. This behaviour signifies the suppression of dynamic bond rearrangements and the progressive fixation of a topological network, typical of the glass formation process. Although this is natural in the glass formation process upon cooling, there are several differences between the previous report with empirical potential[9]. While the previous data shows that Q(2)Q(2) or Q(4)Q(4) vary their values as lowering temperature at least until T=2100T = 2100 K for NS2, our Q(2)Q(2) or Q(4)Q(4) of NS2 reach their plateaus at around T≈2500T \approx2500 K. We attribute this discrepancy in part to incomplete equilibration at lower temperatures in the present simulations, a limitation that warrants further investigation.

Supplementary Information

Pair correlation function

In this Supplementary Information, we present the pair correlation function gαβ(r)g_{\alpha\beta}(r) (with α\alpha, β\beta = Si, O, Na)

gαβ(r)=VNα(Nβ−δαβ)⟨∑i=1Nα∑j=1Nβ14πr2δ(r−∣ri−rj∣)⟩(7)g_{\alpha\beta}(r) = \frac{V}{N_{\alpha}(N_{\beta}-\delta_{\alpha\beta})}\left\langle\sum_{i=1}^{N_{\alpha}}\sum_{j=1}^{N_{\beta}}\frac{1}{4\pi r^{2}}\delta(r-\lvert r_i-r_j\rvert)\right\rangle \tag*{(7)}

where VV is the total volume of the system[53]. We use the freud package[60] for the calculation of gαβ(r)g_{\alpha\beta}(r). Figure S1–Figure S3 show the all cases of gαβ(r)g_{\alpha\beta}(r) for NS2, NS4, and NS6, respectively.

Figure S1. Pair correlation functions gαβ(r)g_{\alpha\beta}(r) for NS2.

Figure S3. Pair correlation functions gαβ(r)g_{\alpha\beta}(r) for NS6.

Figure S2. Pair correlation functions gαβ(r)g_{\alpha\beta}(r) for NS4.

References

  1. [1]C. A. Angell, K. L. Ngai, G. B. McKenna, P. F. McMillan, and S. W. Martin, Relaxation in glassforming liquids and amorphous solids, Journal of Applied Physics 88, 3113 (2000).DOI
  2. [2]G. N. Greaves and S. Sen, Inorganic glasses, glass-forming liquids and amorphizing solids, Advances in Physics 56, 1 (2007).DOI
  3. [3]M. Micoulaut and M. Bauchy, Topology and Rigidity of Silicate Melts and Glasses, Reviews in Mineralogy and Geochemistry 87, 163 (2022).DOI
  4. [4]Z. A. Grady, C. J. Wilkinson, C. A. Randall, and J. C. Mauro, Emerging Role of Non-crystalline Electrolytes in Solid-State Battery Research, Frontiers in Energy Research 8, 218 (2020).DOI
  5. [5]G. Greaves, EXAFS and the structure of glass, Journal of Non-Crystalline Solids 71, 203 (1985).DOI
  6. [6]G. N. Greaves and K. L. Ngai, Reconciling ionic-transport properties with atomic structure in oxide glasses, Physical Review B 52, 6358 (1995).pubmed.ncbi.nlm.nih.gov/9981865
  7. [7]A. Meyer, H. Schober, and D. B. Dingwell, Structure, structural relaxation and ion diffusion in sodium disilicate melts, Europhysics Letters 59, 708 (2002).DOI
  8. [8]A. Meyer, J. Horbach, W. Kob, F. Kargl, and H. Schober, Channel Formation and Intermediate Range Order in Sodium Silicate Melts and Glasses, Physical Review Letters 93, 027801 (2004).DOI
  9. [9]J. Horbach, W. Kob, and K. Binder, Structural and dynamical properties of sodium silicate melts: an investigation by molecular dynamics computer simulation, Chemical Geology 174, 87 (2001).DOI
  10. [10]P. Jund, W. Kob, and R. Julien, Channel diffusion of sodium in a silicate glass, Physical Review B 64, 134303 (2001).DOI
  11. [11]J. Horbach, W. Kob, and K. Binder, Dynamics of Sodium in Sodium Disilicate: Channel Relaxation and Sodium Diffusion, Physical Review Letters 88, 125502 (2002).pubmed.ncbi.nlm.nih.gov/11909472
  12. [12]S. S. Sørensen, M. M. Smedskjær, and M. Micoulaut, Evidence for Complex Dynamics in Glassy Fast Ion Conductors: The Case of Sodium Thiosilicates, The Journal of Physical Chemistry B 127, 10179–10188 (2023).DOI
  13. [13]M. D. Ediger, C. A. Angell, and S. R. Nagel, Supercooled Liquids and Glasses, The Journal of Physical Chemistry 100, 13200 (1996).DOI
  14. [14]P. G. De Benedetti and F. H. Stillinger, Supercooled liquids and the glass transition, Nature 410, 259 (2001).
  15. [15]L. Berthier and G. Biroli, Theoretical perspective on the glass transition and amorphous materials, Reviews of Modern Physics 83, 587 (2011).DOI
  16. [16]M. D. Ediger, Spatially Heterogeneous Dynamics in Supercooled Liquids, Annual Review of Physical Chemistry 51, 99 (2000).pubmed.ncbi.nlm.nih.gov/11031277
  17. [17]W. Kob and H. C. Andersen, Scaling Behavior in the β-Relaxation Regime of a Supercooled Lennard-Jones Mixture, Physical Review Letters 73, 1376 (1994).DOI
  18. [18]W. Kob and H. C. Andersen, Testing mode-coupling theory for a supercooled binary Lennard-Jones mixture I: The van Hove correlation function, Physical Review E 51, 4626 (1995).DOI
  19. [19]W. Kob and H. C. Andersen, Testing mode-coupling theory for a supercooled binary Lennard-Jones mixture. II. Intermediate scattering function and dynamic susceptibility, Physical Review E 52, 4134 (1995).
  20. [20]P. Chaudhuri, L. Berthier, and W. Kob, Universal Nature of Particle Displacements close to Glass and Jamming Transitions, Physical Review Letters 99, 060604 (2007).DOI
  21. [21]N. Lačević, F. W. Starr, T. B. Schrøder, and S. C. Glotzer, Spatially heterogeneous dynamics investigated via a time-dependent four-point density correlation function, The Journal of Chemical Physics 119, 7372 (2003).DOI
  22. [22]L. Berthier, G. Biroli, J.-P. Bouchaud, L. Cipelletti, D. El Masri, D. L’Hôte, F. Ladieu, and M. Pierno, Direct Experimental Evidence of a Growing Length Scale Accompanying the Glass Transition, Science 310, 1797 (2005).DOI
  23. [23]C. Toninelli, M. Wyart, L. Berthier, G. Biroli, and J.-P. Bouchaud, Dynamical susceptibility of glass formers: Contrasting the predictions of theoretical scenarios, Physical Review E 71, 041505 (2005).pubmed.ncbi.nlm.nih.gov/15903675
  24. [24]M. T. Cicerone and M. D. Ediger, Enhanced translation of probe molecules in supercooled o-terphenyl: Signature of spatially heterogeneous dynamics?, The Journal of Chemical Physics 104, 7210 (1996).DOI
  25. [25]S. Sengupta, S. Karmakar, C. Dasgupta, and S. Sastry, Breakdown of the Stokes-Einstein relation in two, three, and four dimensions, The Journal of Chemical Physics 138, 12A548 (2013).DOI
  26. [26]T. Kawasaki and K. K. Kim, Identifying time scales for violation/preservation of Stokes-Einstein relation in supercooled water, Science Advances 3, e1700399 (2017).DOI
  27. [27]J. Horbach and W. Kob, Static and dynamic properties of a viscous silica melt, Physical Review B 60, 3169 (1999).DOI
  28. [28]J. Habasaki and K. L. Ngai, The mixed alkali effect in ionically conducting glasses revisited: A study by molecular dynamics simulation, Physical Chemistry Chemical Physics 9, 4673 (2007).
  29. [29]W. Kob, C. Donati, S. J. Plimpton, P. H. Poole, and S. C. Glotzer, Dynamical Heterogeneities in a Supercooled Lennard-Jones Liquid, Physical Review Letters 79, 2827 (1997).DOI
  30. [30]C. Donati, J. F. Douglas, W. Kob, S. J. Plimpton, P. H. Poole, and S. C. Glotzer, Stringlike Cooperative Motion in a Supercooled Liquid, Physical Review Letters 80, 2338 (1998).DOI
  31. [31]E. R. Weeks, J. C. Crocker, A. C. Levitt, A. Schofield, and D. A. Weitz, Three-Dimensional Direct Imaging of Structural Relaxation Near the Colloidal Glass Transition, Science 287, 627 (2000).pubmed.ncbi.nlm.nih.gov/10649991
  32. [32]J. Behler and M. Parrinello, Generalized Neural-Network Representation of High-Dimensional Potential-Energy Surfaces, Physical Review Letters 98, 146401 (2007).pubmed.ncbi.nlm.nih.gov/17501293
  33. [33]V. L. Deringer, M. A. Caro, and G. Csányi, Machine Learning Interatomic Potentials as Emerging Tools for Materials Science, Advanced Materials 31, 1902765 (2019).DOI
  34. [34]L. C. Erhard, J. Rohrer, K. Albe, and V. L. Deringer, A machine-learned interatomic potential for silica and its relation to empirical models, npj Computational Materials 8, 90 (2022).
  35. [35]L. C. Erhard, J. Rohrer, K. Albe, and V. L. Deringer, Modelling atomic and nanoscale structure in the silicon–oxygen system through active machine learning, Nature Communications 15, 1927 (2024).
  36. [36]M. Bertani, T. Charpentier, F. Faglioni, and A. Pedone, Accurate and Transferable Machine Learning Potential for Molecular Dynamics Simulation of Sodium Silicate Glasses, Journal of Chemical Theory and Computation 20, 1358 (2024).DOI
  37. [37]A. Pedone, M. Bertani, and M. Benassi, Revisiting Machine Learning Potentials for Silicate Glasses: The Missing Role of Dispersion Interactions, Journal of Chemical Theory and Computation 21, 4769 (2025).DOI
  38. [38]S. Ganissetti, T. Du, N. M. A. Krishnan, and M. M. Smedskjaer, On-the-fly machine learning force fields for alkali silicate glasses, Physical Review Materials 9, 115601 (2025).
  39. [39]K. Schütt, O. Unke, and M. Gastegger, Equivariant message passing for the prediction of tensorial properties and molecular spectra, in Proceedings of the 38th International Conference on Machine Learning, Vol. 139 (PMLR, 2021) pp. 9377–9388.arxiv.org/abs/2102.03150
  40. [40]G. Kresse and J. Hafner, Ab initio molecular dynamics for liquid metals, Physical Review B 47, 558(R) (1993).
  41. [41]G. Kresse and J. Furthmüller, Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set, Computational Materials Science 6, 15 (1996).DOI
  42. [42]G. Kresse and J. Furthmüller, Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set, Physical Review B 54, 11169 (1996).pubmed.ncbi.nlm.nih.gov/9984901
  43. [43]J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized Gradient Approximation Made Simple, Physical Review Letters 77, 3865 (1996).pubmed.ncbi.nlm.nih.gov/10062328
  44. [44]G. Kresse and D. Joubert, From ultrasoft pseudopotentials to the projector augmented-wave method, Physical Review B 59, 1758 (1999).DOI
  45. [45]L. Loshchilov and F. Hutter, Decoupled Weight Decay Regularization, in International Conference on Learning Representations (2019).
  46. [46]G. Bussi, D. Donadio, and M. Parrinello, Canonical sampling through velocity rescaling, The Journal of Chemical Physics 126, 014101 (2007).pubmed.ncbi.nlm.nih.gov/17212484
  47. [47]S. Nosé, A unified formulation of the constant temperature molecular dynamics methods, The Journal of Chemical Physics 81, 511 (1984).DOI
  48. [48]W. G. Hoover, Canonical dynamics: Equilibrium phase-space distributions, Physical Review A 31, 1695 (1985).pubmed.ncbi.nlm.nih.gov/9895674
  49. [49]F. Pabst and S. Baroni, Glassy dynamics in a glass-forming liquid: A first-principles study of toluene, Physical Review E 111, L023401 (2025).DOI
  50. [50]R. Yamamoto and A. Onuki, Kinetic Heterogeneities in a Highly Supercooled Liquid, Journal of the Physical Society of Japan 66, 2545 (1997).DOI
  51. [51]R. Yamamoto and A. Onuki, Dynamics of highly supercooled liquids: Heterogeneity, rheology, and diffusion, Physical Review E 58, 3515 (1998).DOI
  52. [52]H. Shiba, T. Kawasaki, and A. Onuki, Relationship between bond-breakage correlations and four-point correlations in heterogeneous glassy dynamics: Configuration changes and vibration modes, Physical Review E 86, 041504 (2012).DOI
  53. [53]J.-P. Hansen and I. R. McDonald, Theory of Simple Liquids, 4th ed. (Academic Press, 2013).
  54. [60]V. Ramasubramani, B. D. Dice, E. S. Harper, M. P. Spellings, J. A. Anderson, and S. C. Glotzer, A Software Suite for High Throughput Analysis of Particle Simulation Data, Computer Physics Communications 254, 107275 (2020).

Paper details

vlnc:W_53sgh96v
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents