betaFeedback

WorksPhysics and Astronomy

A Local Structural Basis to Resolve Amorphous Ices

Abstract

Phases with distinct thermodynamic properties must differ in their underlying microscopic configurations. While ordered phases are readily distinguished by unit cells and space groups, the local structural basis differentiating amorphous phases is less apparent. Here, using a new probabilistic data-driven framework applied to molecular simulations of water, we identify local collective variables that discriminate low-density and high-density amorphous (LDA and HDA) ices and characterize pressure-induced transitions between them. As expected, descriptors related to local density effectively distinguish LDA and HDA; however, phase identity is surprisingly encoded within the first coordination shell. Furthermore, the pressure-induced LDA–HDA transformation proceeds through redistribution between LDA- and HDA-like local environments with no evidence for intermediate structures, consistent with a first-order-like phase transition. This contrasts with the gradual structural evolution observed in other amorphous systems, such as metallic glasses. Critically, local hydrogen density reveals pronounced structural hysteresis between compression and decompression pathways, which is not apparent in orientational order parameters, demonstrating that the microscopic interpretation of amorphous transformations depends fundamentally on descriptor choice. These findings are robust across force fields and provide a general strategy for characterizing disordered phases lacking obvious distinguishing features.

The thermodynamic properties that define macroscopic materials ultimately emerge from distributions of microscopic configurations [1]. In crystalline materials, the structural basis of a particular phase is unambiguously in its unit cell and space group. For amorphous phases, there is no obvious analogous structural scaffold. If two amorphous phases are thermodynamically distinct, with different densities or response functions, their underlying distributions of local environments must also differ [2, 3]. What is far less clear is how they differ, over what length scales such differences are encoded, and whether they can be expressed in terms of a small number of collective variables that admit a microscopic interpretation.

Water’s amorphous ices provide a compelling arena to address these questions. Low-density amorphous (LDA) and high-density amorphous (HDA) ice represent two glassy states with markedly different densities and thermodynamic behavior [4, 5]. Their interconversion has long been discussed in the context of water’s putative liquid–liquid phase transition and associated polyamorphism [6, 7, 8]. LDA and HDA are viewed as glassy analogs of the low-density and high-density liquids in supercooled water [9, 10], making the structural basis directly relevant to long-standing questions about water’s anomalous behavior.

Prior studies have established that LDA exhibits an open, tetrahedrally ordered network, whereas HDA is denser and more disordered [7, 11, 12]. This picture has emerged primarily from physically motivated descriptors, including tetrahedral order parameters [13, 14, 15], bond-orientational order (BOO) metrics [16, 17, 18], coordination numbers, and local structure indices [14, 19, 20]. However, which descriptors most effectively encode this distinction, and whether other key collective variables remain undiscovered or more informative, is unclear. Studies of supercooled liquid water suggest that distinguishing low- and high-density environments requires correlations extending beyond the first coordination shell [17, 20], but whether this nonlocality reflects fundamental structural physics rather than descriptor limitations has not been assessed.

A separate open question concerns the microscopic character of the LDA–HDA transformation. These glasses exhibit first-order-like transformations under pressure [4, 12, 21, 22, 23, 24], yet whether the transition proceeds through structurally distinct intermediates, as in some metallic glasses [25, 26], or through redistribution between LDA-like and HDA-like motifs without intermediates [17, 27, 28], remains unresolved at the molecular scale.

Existing approaches struggle to address these questions simultaneously. Simple low-dimensional order parameters are interpretable but can be biased by their construction, highlighting anticipated features while obscuring others. High-dimensional descriptors [29, 30] can encode rich local information, and machine-learning classifiers built on these representations can achieve high discriminative accuracy [18, 31, 32, 33, 34, 35, 36]. However, such models are often opaque, classifying without revealing which descriptors matter and lacking mechanisms to detect out-of-distribution (OOD) configurations. One therefore cannot identify the minimal discriminative features, assess how local the distinction is, or determine whether transformation intermediates represent novel structures.

Here, we use an interpretable probabilistic framework applied to molecular simulations of water (Figure 1) to systematically identify the local structural basis distinguishing LDA and HDA and to characterize the microscopic nature of their pressure-induced interconversion. Mutual-information (MI) analysis ranks a large set of local structural descriptors by their discriminative power, and the most informative are combined into a calibrated classifier that also detects configurations belonging to neither phase. We find that interstitial hydrogen density dominates the discrimination and that phase identity is encoded within the first coordination shell. Analysis of transformation trajectories reveals no intermediate structures; rather, the transition proceeds through redistribution of discrete LDA-like and HDA-like environments. Strikingly, local hydrogen density exposes structural hysteresis between compression and decompression pathways that is entirely invisible to orientational order parameters, revealing that the choice of descriptor has fundamental consequences for the physical interpretation of amorphous transformations.

Figure 1. Probabilistic framework for resolving local structural differences between amorphous phases. (1) Local atomic environments are sampled from MD simulations and represented using ACSF and BOO descriptors. (2) A two-stage feature-selection procedure removes redundant descriptors and identifies those most informative for distinguishing LDA and HDA. (3) For each selected descriptor and class, one-dimensional probability densities are estimated using Gaussian kernel density estimation. (4) The joint probability for each class is computed as the product of the descriptor-wise probabilities, providing an interpretable probabilistic classification and enabling explicit detection of out-of-distribution configurations.

Probabilistic classification effectively distinguishes amorphous states and novel structures

We consider probabilistic classification of data generated by DP_MBpol, a Deep Potential model trained on the many-body polarizable (MB-pol) potential parameterized against coupled-cluster calculations [24, 37]. A second model, DP_SCAN, trained on DFT data with the SCAN functional [38, 39], is analyzed in the SI and employed later to assess sensitivity to the force field choice. Local environments are represented using atom-centered symmetry functions (ACSFs) [29] and bond-orientational order (BOO) parameters [16], constructed from 16 nearest neighbors, consistent with prior work [17, 18, 20], though we systematically vary environment size for analysis. The framework uses mutual information to rank descriptors by discriminative power, selects an informative and approximately uncorrelated subset, and combines class-conditional density estimates into a joint probability that enables molecule-by-molecule phase assignment and detection of out-of-distribution configurations. Importantly, this procedure automatically discovers which descriptors best distinguish phases without requiring a priori knowledge of the relevant order parameters. Complete methodological details are provided in Methods and SI.

We first assess how accurately and confidently our framework can classify LDA and HDA. To evaluate performance, we benchmark against three approaches with distinct strategies: a neural network trained on bond-orientational order parameters (BOO-NN) [17, 18], a coordinate-level architecture that learns representations directly from atomic positions (PointNet) [31], and an autoencoder combined with Gaussian mixture modeling (AE-GMM) [35]. All models are trained using environments constructed from 16 nearest neighbors; full precision, recall, and accuracy metrics for DP MBpol and DP SCAN are reported in the SI (Table S1 and Table S2).

ModelHDA-PrecisionHDA-RecallLDA-PrecisionLDA-RecallAccuracy
BOO-NN0.997 ± 0.0010.997 ± 0.0020.997 ± 0.0020.997 ± 0.0010.997 ± 0.001
PointNet0.937 ± 0.0080.988 ± 0.0030.988 ± 0.0030.934 ± 0.0080.961 ± 0.003
AE-GMM0.938 ± 0.0080.982 ± 0.0030.983 ± 0.0030.935 ± 0.0080.959 ± 0.003
Present Work (BOO)0.997 ± 0.0010.968 ± 0.0100.967 ± 0.0100.997 ± 0.0010.982 ± 0.005
Present Work (ACSF)0.973 ± 0.0120.988 ± 0.0090.987 ± 0.0100.972 ± 0.0130.980 ± 0.003
Present Work (BOO+ACSF)0.996 ± 0.0010.989 ± 0.0030.989 ± 0.0030.996 ± 0.0010.993 ± 0.001

Table S1. Comparison of model performance for classification of DP_MBpol-generated LDA/HDA environments. Values represent the mean and standard deviation over five-fold cross-validation.

ModelHDA-PrecisionHDA-RecallLDA-PrecisionLDA-RecallAccuracy
BOO-NN0.947 ± 0.0070.997 ± 0.0010.997 ± 0.0010.945 ± 0.0070.971 ± 0.003
PointNet0.937 ± 0.0080.988 ± 0.0030.988 ± 0.0030.934 ± 0.0080.961 ± 0.003
AE-GMM0.938 ± 0.0080.982 ± 0.0030.983 ± 0.0030.935 ± 0.0080.959 ± 0.003
Present Work (BOO)0.951 ± 0.0060.977 ± 0.0080.978 ± 0.0080.950 ± 0.0050.964 ± 0.006
Present Work (ACSF)0.935 ± 0.0100.997 ± 0.0010.998 ± 0.0010.931 ± 0.0100.964 ± 0.005
Present Work (BOO+ACSF)0.948 ± 0.0060.987 ± 0.0040.988 ± 0.0040.945 ± 0.0050.967 ± 0.004

Table S2. Comparison of model performance for classification of DP_SCAN-generated LDA/HDA environments. Values represent the mean and standard deviation over five-fold cross-validation.

All approaches achieve high classification accuracy, confirming that LDA and HDA environments are microscopically distinguishable. Our method achieves 99.3% binary classification accuracy, matching or surpassing all benchmarks and outperforming classifiers trained on BOO parameters or ACSFs alone, indicating that combining complementary descriptor types improves classification. However, high classification accuracy alone does not guarantee

reliability when applied to transformation trajectories, where configurations belonging to neither LDA nor HDA may arise. To evaluate OOD detection, we examine how models classify ice Ih environments, which share LDA’s tetrahedral coordination but possess long-range crystalline order. BOO-NN and PointNet assign high LDA probabilities to many Ih configurations (Figure 2A,B), reflecting overconfidence characteristic of neural-network classifiers that partition descriptor space without recognizing inputs outside the training distribution. In contrast, AE-GMM and our method assign Ih probabilities orders of magnitude lower than genuine LDA structures (Figure 2C,D), enabling reliable detection of novel configurations. This capability is essential for analyzing transformation pathways where intermediate environments may exist.

Local hydrogen density and orientational symmetry resolve amorphous and crystalline ice environments

Figure 2. Local hydrogen density and orientational symmetry resolve amorphous and crystalline ice environments. (A-D) Predicted LDA log-probabilities for true LDA environments (red) and hexagonal ice Ih (purple) for (A) BOO-NN, (B) PointNet, (C) AE-GMM, and (D) this work. ‘Density*’ refers to probability densities that are normalized so that their maximum value is 1.0. (E) Mutual information for the 50 most informative descriptors, where green indicates selected descriptors and gray indicates unselected descriptors. Descriptors are selected for both their mutual information values and their correlation with previously chosen descriptors. Schematics of descriptors and corresponding probability distributions from LDA and HDA environments for (F, G) GH1G_{\mathrm{H}}^{1}, (H, I) q3q_3, and (J, K) qˉ11\bar{q}_{11}. In the schematics, darker colors indicate higher descriptor values.

To test this more broadly, we evaluate OOD detection against 14 additional ice polymorphs spanning a wide range of pressures (SI Section S4). Using the complete set of weakly correlated descriptors, rather than only the five selected for LDA/HDA classification, substantially improves detection for most polymorphs, with several phases moving from near-zero to near-complete detection, though a subset of high-pressure phases remain only partially distinguishable from LDA/HDA even with the full descriptor set. This residual structural overlap does not imply that these crystalline phases are present within LDA/HDA. Previous work has found no evidence of ice-like domains in computational LDA or HDA [40], indicating that the classifier ambiguity for these phases instead reflects local structural similarity.

Hierarchy of descriptor importance highlights key structural differences

To identify which structural descriptors most effectively distinguish LDA and HDA, we examine the descriptors selected by mutual-information and correlation analysis in our probabilistic framework (Figure 2E). Notably, the ACSF descriptor GH1G_{\mathrm{H}}^{1} exhibits the highest MI, demonstrating that radial density information is more discriminating than the orientational symmetry captured by BOO parameters, which have been the focus of prior work [17, 18]. Many additional ACSF descriptors are nearly as informative as GH1G_{\mathrm{H}}^{1} but are excluded due to their high correlation with it. The remaining selected features are BOO parameters, which are consistent with those previously identified as changing significantly during LDA/HDA transitions [18]. These descriptors enable accurate classification for multiple water force fields (Table S1 and Table S2), indicating that they capture structural differences intrinsic to the amorphous ices rather than force-field-specific artifacts.

To understand the physical significance of these descriptors, we examine how they encode structural differences between LDA and HDA. The most informative descriptor, GH1G_{\mathrm{H}}^{1}, can be interpreted as a weighted hydrogen count in the local environment of an oxygen, up to Rc=5.0 A˚R_{\mathrm{c}} = 5.0\,\mathrm{\mathring{A}} or ca. 16 water molecules (Figure 2F). HDA environments exhibit substantially larger GH1G_{\mathrm{H}}^{1} values than LDA (Figure 2G), consistent with HDA’s higher bulk density. Notably, local hydrogen density is more informative than the analogous oxygen-based descriptor GO1G_{\mathrm{O}}^{1}; this distinction emerges automatically from the descriptor selection. The remaining selected descriptors consist of BOO parameters, which quantify spherical symmetries of varying order across different length scales (Figure 2H,J). For all selected BOO descriptors, LDA environments exhibit larger values than HDA (Figure 2I,K). This behavior is similar to that of the conventional tetrahedral order parameter qq, for which average values are q≈0.91q \approx0.91 for LDA and q≈0.74q \approx0.74 for HDA [24]. The HDA value is comparable to that of ambient liquid water, whose tetrahedrality is typically q≈0.75q \approx0.75 [14], whereas LDA lies much closer to the ordered tetrahedral structure of ice Ih, which exhibits q≈0.98q \approx0.98 [13, 14].

Taken together, these descriptors reveal that local environments are distinguished primarily through variations in local density, with orientational symmetry playing a secondary but complementary role. The prominence of density-based features is consistent with prior work identifying local density as correlated with effective order parameters for liquid water near its critical point [19]. Within this low-dimensional descriptor space, individual molecular environments are classified as LDA or HDA based on their position along these axes, while environments falling outside the characteristic ranges of both phases are naturally identified as distinct.

LDA and HDA are resolvable within the first coordination shell due to interstitial hydrogen

A common feature of prior classification approaches for LDA and HDA is the use of structural information that extends beyond the first coordination shell, incorporating 16 nearest neighbors [17, 18, 20]. This perspective is reinforced by studies of supercooled liquid water, where clear separation between low- and high-density environments emerges only when structural information is averaged beyond the first shell [20]. Motivated by these findings, we ask whether the same level of structural nonlocality is required to distinguish the glassy states. To this end, we compute ACSFs and BOO parameters for local environments containing between 2 and 16 nearest neighbors. Here, we retain only local descriptors that do not incorporate information beyond the defined environment (i.e., excluding neighbor-averaged quantities such as qˉ\bar{q}) to directly quantify the minimal structural information required for classification.

Classification accuracy increases monotonically with environment size (Figure 3A), at approximately 86% with only two neighboring water molecules and plateauing above 97% for eight or more neighbors. These results indicate that the distinction between glassy LDA and HDA is encoded primarily in local packing motifs and not nanometer-scale domains. That substantial phase-specific information is encoded within the immediate coordination shell is unexpected given prior work highlighting the necessity of second-shell information [17, 18, 20]. Rather, Figure 3A shows that the first solvation shell contains much of the structural information needed to distinguish LDA and HDA, with larger environments extending toward the second solvation shell offering a modest further improvement in classification accuracy.

Classification accuracy and descriptor distributions for amorphous ices

Figure 3. Amorphous state identity is encoded by length-scale dependent features. (A) Classification accuracy as a function of environment size, where size is the number of neighboring water molecules. Means and standard deviations are obtained from five-fold cross validation. (B) Functional form of the per-neighbor contribution to GH3G_{\mathrm{H}}^{3} as a function of interatomic distance. Distributions from HDA and LDA environments characterized by (C) GH3(κ=1.5)G_{\mathrm{H}}^{3}(\kappa= 1.5) using environments of size 3, (D) GH3(κ=1.5)G_{\mathrm{H}}^{3}(\kappa= 1.5) using environments of size 16, and (E) GH1G_{\mathrm{H}}^{1} using environments of size 3.

These conclusions are necessarily bounded by the maximum environment size considered and do not exclude the possibility that structural differences between LDA and HDA persist at longer range. The length-scale scan of Figure 3A is computationally inexpensive and constitutes a useful diagnostic for any new application of this framework, revealing the environment size beyond which additional structural information ceases to improve discrimination; this saturation behavior may be system-specific. More broadly, the question addressed here is inherently local. LDA and HDA are readily distinguishable by their bulk properties, and the non-trivial result is that this distinction is already encoded at the level of individual molecular environments.

To understand what structural features most strongly delineate LDA and HDA at the smallest scales, we examine the most informative descriptors when considering only three neighboring water molecules. We find that GH3(κ=1.5)G_{\mathrm{H}}^{3}(\kappa= 1.5) is the most informative descriptor at this scale. The functional form (Figure 3B) takes negative values at distances between 1.5 and 3 Å, corresponding to hydrogen-bonded neighbors. HDA environments exhibit more negative values than LDA (Figure 3C), indicating higher hydrogen density within hydrogen-bonding distance of the central molecule. This is consistent with densification disrupting the tetrahedral hydrogen-bond network of LDA through insertion of interstitial water molecules into the first coordination shell. The classifier effectively detects these interstitial hydrogens through GH3G_{\mathrm{H}}^{3}, enabling accurate LDA–HDA discrimination using only first-shell information.

However, the quality of the information provided by GH3G_{\mathrm{H}}^{3} depends on spatial scale. This is demonstrated by computing its value using 16 neighboring water molecules and observing the substantial overlap between distributions of LDA and HDA configurations (Figure 3D). Thus, the ability of GH3G_{\mathrm{H}}^{3} to discriminate LDA and HDA worsens with increasing environment size. Conversely, GH1G_{\mathrm{H}}^{1}, the most informative descriptor for these larger environments (Figure 2F), provides little separation for environments of three neighbors (Figure 3E). These contrasting behaviors indicate that different structural correlations dominate at different length scales. In particular, interstitial hydrogen insertion is most clearly distinctive at local scales, while the cumulative density differences captured by GH1G_{\mathrm{H}}^{1} emerge only when averaging over larger environments.

Redistribution of local environments underlies transition and highlights first-order character

We next consider the microscopic character of the LDA–HDA transformation. Specifically, does it proceed through structurally distinct intermediate environments or via redistribution between locally LDA-like and HDA-like motifs? Experimentally, LDA and HDA can be interconverted via isothermal compression and decompression, with signatures commonly interpreted as first-order-like phase transitions [4, 12, 21]. We apply our classifier to isothermal compression and decompression trajectories at T=80T = 80 K for DP_MBpol water [24], labeling each molecule as LDA, HDA, or outlier based on its local environment.

Figure 4A–C summarizes the transformation behavior. The LDA→HDA transition unfolds continuously over approximately 8–16 kbar during compression. The reverse transformation requires negative pressures, with LDA and HDA populations coexisting over a broad pressure range before LDA is fully recovered, which is indicative of pronounced hysteresis. Crucially, the transformation proceeds entirely through changes in the relative populations of LDA- and HDA-like environments, with no intermediate motifs observed. This population-based picture is consistent with LDA and HDA corresponding to separate regions on the underlying potential energy landscape [41, 42] and provides direct molecular-scale evidence for the first-order-like phase transition character of the LDA-HDA transformation.

Compression and decompression pathways of amorphous ices

Figure 4. Compression and decompression of amorphous ices follow structurally distinct microscopic pathways without intermediate local environments. (A) Evolution of LDA (red), HDA (blue), and outlier (gray) populations during isothermal compression at T=80T = 80 K for rates of 3 kbar/ns (solid) and 0.3 kbar/ns (dashed). Vertical lines indicate transition pressures, defined by equal LDA and HDA populations. (B) Density–pressure equation of state extracted from compression (circles) and decompression (squares) trajectories for rates of 3 kbar/ns (filled) and 0.3 kbar/ns (open). (C) Analogous evolution of microscopic populations during decompression. (D) Out-of-equilibrium phase diagram deduced from local structural descriptors (markers with same meaning as in B) based on compression/decompression runs at different temperatures. Lines indicate LDA-to-HDA transition pressures from a thermodynamic criterion applied to the equation of state [24], which is reliable for compression but not decompression. (E) Distributions of GH1G_{\mathrm{H}}^{1} (top) and q3q_3 (bottom) for environments sampled during compression (blue) and decompression (red) within a common density window (1.15–1.25 g/cm3^3). (F) Mean local-environment descriptors (q3q_3, q4q_4, GH1G_{\mathrm{H}}^{1}) along compression (blue) and decompression (red) pathways for separate isothermal MD simulations at temperatures from 60 to 140 K. The upper set of curves shows trajectories in the full q3q_3–q4q_4–GH1G_{\mathrm{H}}^{1} space; the lower set shows the same trajectories projected onto the q3q_3–q4q_4 plane. Stars mark transition points; darker shades correspond to lower temperatures. See SI Figures S2 and S3 for 2D projections.

The population-redistribution picture also connects to the phenomenology of fluid polyamorphism developed by Anisimov and coworkers [43], which distinguishes between a “discrete” mechanism, driven by interconversion between two structurally distinct molecular states, and a “continuous” mechanism arising from nonideality in the Gibbs energy, both of which

can produce identical macroscopic signatures. The molecule-resolved probabilities PLDA,iP_{\mathrm{LDA},i} and PHDA,iP_{\mathrm{HDA},i} computed here are analogous microscopic realizations in the glass state, of the interconversion fractions of the liquid state, central to that framework. The absence of intermediate environments along the LDA-HDA transformation pathway indicates that the discrete mechanism operates in water’s amorphous ices. Our results are consistent with theoretical descriptions of water based on two-state models [44, 45].

Local structural motifs reliably map LDA/HDA boundaries

We next use molecule-resolved classification to identify LDA/HDA transition pressures across a range of temperatures, defining the transition as the point of equal LDA and HDA populations. Conventionally, the LDA–HDA transition pressures are estimated from the rate-dependent density–pressure equation of state (EoS), where the LDA→\toHDA transformation appears as a maximum in the rate of density change with respect to pressure (an inflection point in the density-pressure EoS). However, the reverse HDA→\toLDA transition lacks clear thermodynamic signatures in the EoS, precluding identification of its transition pressure [24].

For compression trajectories, classifier-derived and EoS-based transition pressures agree quantitatively (Figure 4D, circles and lines), reflecting strong coupling between local structural reorganization and macroscopic densification. For decompression, the classifier robustly identifies HDA→\toLDA transition pressures even where the EoS varies continuously without inflection (Figure 4D, squares), enabling construction of complete out-of-equilibrium phase diagrams, including the decompression spinodal previously inaccessible from thermodynamic analysis [24]. Furthermore, substantial hysteresis is observed, with decompression transitions occurring at negative pressures and differing from compression transition pressures by 15–25 kbar, consistent with large kinetic barriers characteristic of first-order-like transformations.

Structural pathways of pressure-induced transformations exhibit hysteresis

Given the substantial hysteresis in transition pressures, we next ask whether compression and decompression follow identical structural pathways by tracking mean values of GH1G_{\mathrm{H}}^{1}, q3q_{3}, and q4q_{4} along trajectories across temperatures from 60 to 140 K. The distributions of GH1G_{\mathrm{H}}^{1} and q3q_{3} sampled within a common density window of 1.15–1.25 g/cm3^{3} (Figure 4E) directly demonstrate structural hysteresis. For q3q_{3}, compression and decompression distributions are clearly separated, demonstrating that the forward and reverse transformations traverse structurally distinct regions of local configuration space. More subtly, GH1G_{\mathrm{H}}^{1} also differs between pathways despite identical bulk densities, a nontrivial result given that GH1G_{\mathrm{H}}^{1} reflects local interstitial hydrogen density and might naively be expected to track global density alone.

In the q3q_{3}–GH1G_{\mathrm{H}}^{1} plane (Figure 4F), compression and decompression follow clearly distinct paths across all temperatures. Compression proceeds in two stages: (i) initial densification with little loss of orientational order (increasing GH1G_{\mathrm{H}}^{1} at nearly constant q3q_{3}), reflecting elastic compression of the hydrogen-bond network, followed by (ii) collapse of tetrahedral order as densification continues into the HDA basin. Temperature primarily affects the second stage, with higher temperatures allowing tetrahedral order to collapse at slightly higher hydrogen coordination. Decompression follows a qualitatively different sequence: density decreases substantially before any recovery of tetrahedral order, with q3q_{3} increasing only after the system has expanded well into the low-density regime.

Strikingly, when GH1G_{\mathrm{H}}^{1} is replaced by q4q_{4}, compression and decompression trajectories collapse onto nearly identical curves, and the transformation appears reversible. This pattern holds for any combination of BOO parameters alone (SI Figures S2 and S3), demonstrating that the microscopic interpretation of amorphous transformations depends critically on descriptor choice; local hydrogen density reveals pronounced hysteresis that orientational order parameters alone entirely obscure.

Different force fields produce locally distinct LDA phases

As a further test, we now ask whether the same titular amorphous phase, generated by different force fields, is structurally distinguishable. A classifier trained to distinguish DP_SCAN and DP_MBpol LDA configurations achieves 77.5 ±\pm 0.3% accuracy on unseen environments, confirming that the two models produce measurably different local structures. Structural differences between HDA configurations are considered in the SI, Figure S4. The most discriminating features are exclusively ACSFs, with no BOO parameters within the top five (Table I), indicating that the models differ primarily in radial packing rather than orientational symmetry. This is consistent with oxygen–oxygen radial distribution functions at T=80T = 80 K, which show a slightly higher short-range peak for DP_SCAN (Figure 5A), and with ACSF descriptors that confirm systematically higher local densities for both oxygen and hydrogen in DP_SCAN environments (Figure 5B,C), reflecting the tendency of SCAN-based potentials to overstabilize close-contact configurations [24, 39].

FeatureMutual Information
GO2(η=2.5,Rs=1.0)G_{\mathrm{O}}^{2}(\eta= 2.5, R_{\mathrm{s}} = 1.0)0.220
GH2(η=1.0,Rs=1.0)G_{\mathrm{H}}^{2}(\eta= 1.0, R_{\mathrm{s}} = 1.0)0.188
GH,O5(η=4.5,ζ=0.5,λ=−1.0)G_{\mathrm{H,O}}^{5}(\eta= 4.5, \zeta= 0.5, \lambda= -1.0)0.142
GH3(κ=0.5)G_{\mathrm{H}}^{3}(\kappa= 0.5)0.141
GO3(κ=2.0)G_{\mathrm{O}}^{3}(\kappa= 2.0)0.117

Table I. Most informative local structural descriptors for resolving DP_SCAN and DP_MBpol local environments.

Figure 5. Radial packing variations highlight differences between quantum-chemistry-derived water models. (A) Oxygen–oxygen radial distribution functions for LDA at T=80T = 80 K and P=1P = 1 bar. (B–E) Distributions of select ACSF and BOO descriptors for LDA generated by each model across temperatures from 60 to 140 K at P=1P = 1 bar. All probability densities are normalized such that the maximum value of either distribution is 1.0, with both distributions scaled by the same factor (denoted as Density*). The blue and orange colors represent the DP SCAN- and DP MBpol models, respectively.

Figure S4. Differences between local environments of HDA generated from DP_SCAN and DP_MBpol. (A–D) Distributions of select ACSF and BOO descriptors for HDA generated by each model. All probability densities are normalized such that the maximum value of either distribution is 1.0 (denoted as Density*). The blue and orange colors represent the DP_SCAN and DP_MBpol models, respectively.

Orientational descriptors reveal more subtle differences. DP_MBpol produces slightly lower tetrahedral order and greater variability in local orientational environments (Figure 5D,E), indicating a more distorted hydrogen-bond network despite both models capturing the overall tetrahedral topology of LDA. These results demonstrate that environment-resolved structural analysis provides a sensitive lens for comparing machine-learned potentials as such models proliferate.

Implications

The findings reported here advance the structural picture of water’s amorphous ices in several respects. The locality of the LDA–HDA distinction, encoded within the first coordination shell, challenges the prevailing view that second-shell correlations are necessary to resolve amorphous water environments [17, 20]. In

vitrified glasses, density fluctuations are arrested in distinct configurational basins, and differences in interstitial occupancy become sufficiently pronounced that first-shell descriptors alone suffice, suggesting that the requisite length scale for structural classification is state-dependent. This state-dependence suggests that extending the framework to equilibrium liquid environments where broader thermal fluctuations are expected to increase configuration overlap may require a larger or different descriptor set. Classification of LDL and HDL environments as the liquid–liquid critical point is approached represents a compelling direction for future work. The absence of intermediate structural populations along transformation pathways, established here with a classifier capable of explicitly detecting OOD configurations, constrains the energy landscape picture of glass–glass interconversion [42] and provides molecular-scale evidence for first-order character even where macroscopic observables vary continuously. Additionally, the molecule-resolved classification demonstrated here offers a general microscopic strategy for distinguishing the discrete and continuous mechanisms of polyamorphism [43].

More broadly, the finding that orientational order parameters alone mask the hysteresis revealed by local hydrogen density carries a fundamental cautionary implication. The microscopic interpretation of amorphous transformations depends critically on descriptor choice, and analyses based on a single class of order parameter risk overlooking essential structural physics. This descriptor-dependence extends to OOD detection itself. Broadening the descriptor set used to flag anomalous environments substantially improved sensitivity to several high-pressure ice polymorphs. The probabilistic framework introduced here is applicable to any system in which distinct amorphous or disordered phases lack obvious microscopic distinguishing features. The environment-resolved comparison of machine-learned potentials illustrates one such application, revealing force-field-specific structural signatures invisible to conventional metrics. As data-driven interatomic potentials become increasingly prevalent, such analyses will be essential for connecting microscopic structure to emergent condensed-phase behavior.

Methods

We analyze MD trajectories of LDA and HDA ice generated using two machine-learned interatomic potentials: DP SCAN, trained on DFT data with the SCAN functional [38, 39], and DP MBpol, trained on the many-body polarizable MB-pol potential parameterized against coupled-cluster calculations [37, 46, 47, 48]. All simulations were performed using LAMMPS (2 Aug 2023) [49] interfaced with DeePMD-kit (v3.0.0) [50, 51]; full simulation protocols are provided in Ref. 24.

The training and evaluation dataset comprises 10,000 molecular environments, equally divided between LDA and HDA. Configurations were deliberately selected from thermodynamic conditions where phase identity is unambiguous, avoiding the transition region. LDA environments were drawn from isobaric quenches at temperatures at least 30 K below the glass transition and from isothermal compression of LDA at pressures at least 3 kbar below its transformation pressure to HDA. HDA environments were taken from compression trajectories at least 3 kbar above the LDA-to-HDA transformation pressure, as well as from decompression trajectories restricted to densities exceeding that of recovered HDA at 80 K and 1 bar, to exclude partially transformed structures. An additional 1,000 hexagonal ice (Ih) configurations, which share the tetrahedral hydrogen-bond network of LDA but possess long-range crystalline order, were curated as a benchmark for out-of-distribution (OOD) detection. All classification results are validated using five-fold cross-validation.

Each molecule’s local environment is characterized using two complementary, rotationally invariant descriptor families. Atom-centered symmetry functions (ACSFs) [29] encode radial and angular correlations of interatomic distances through a set of radial (G1G^1, G2G^2, G3G^3) and angular (G4G^4, G5G^5) functions, computed separately for oxygen and hydrogen neighbors to retain sensitivity to hydrogen-bond motifs and proton arrangements not visible from oxygen positions alone. Bond-orientational order (BOO) parameters [16], including locally averaged variants [52], quantify the orientational symmetry of the surrounding molecular arrangement via spherical harmonics. Unless otherwise stated, local environments are defined by the 16 nearest neighboring water molecules, with a cutoff radius of Rc≈5.0R_{\mathrm{c}} \approx5.0 Å, consistent with prior work [17, 18, 20]. Full descriptor definitions and parameter sets are provided in the SI.

The most discriminative descriptors are identified through a two-stage selection procedure. First, each descriptor is ranked by its mutual information (MI) with the class label [53], retaining those with MI values at least 10% of the maximum observed value. Second, a greedy correlation filter removes redundant descriptors. Proceeding in order of decreasing MI, a descriptor is retained only if its Pearson correlation with all previously retained descriptors satisfies ∣r∣<0.8|r| < 0.8. This yields a compact set of informative, approximately uncorrelated descriptors without requiring prior assumptions about which structural motifs are important. For classification, we use the five most informative of these descriptors (Fig. 2E); for OOD detection (below), we instead use the complete correlation-filtered descriptor set, as additional weakly correlated descriptors provide independent axes along which an anomalous environment may be identified and thereby maximize outlier detection sensitivity.

Classification proceeds by modeling the class-conditional distribution of each selected descriptor independently using Gaussian kernel density estimation (KDE), with bandwidths optimized by five-fold cross-validation on the provided training data. For a new environment, the joint likelihood for each class is computed as the product of per-descriptor likelihoods under a naïve Bayes assumption, and the environment is assigned to the highest-likelihood class. A key advantage of this formulation is explicit OOD detection. If an environment falls outside of the LDA and HDA reference distributions along any of the descriptors retained by the correlation filter (defined by a kernel density estimate below 5×10−45 \times10^{-4}), then it is flagged as structurally novel rather than forced into a known category. We note that OOD detection in this framework is inherently relative to the structural descriptors provided. Structures that are anomalous along alternative descriptor axes may not be reliably flagged. Five-fold cross-validation is used to obtain robust estimates of classification accuracy. Full mathematical details of the probabilistic model are provided in the SI.

Code Availability

All code required to recreate the results above can be obtained at https://github.com/webbtheosim/amorphous-ice.

Acknowledgments

Q.M.G. acknowledges support from the National Science Foundation Graduate Research Fellowship Program under Grant No. DGE-2039656. R.J.S., P.G.D., and M.A.W acknowledge support from the “Chemistry in Solution and at Interfaces” (CSI) Center funded by the U.S. Department of Energy through Award No. DE-SC0019394. N.G. is thankful to the NSF-CREST Center for Interface Design and Engineered Assembly of Low Dimensional systems (IDEALS; grant numbers HRD-1547380 and HRD-2112550). Simulations and analyses were performed using resources from Princeton Research Computing at Princeton University, which is a consortium led by the Princeton Institute for Computational Science and Engineering (PICSciE) and Office of Information Technology’s Research Computing.

Supporting Information for “A Local Structural Basis to Resolve Amorphous Ices”

Quinn M. Gallagher; Ryan J. Szukalo; Nicolas Giovambattista; Pablo G. Debenedetti; Michael A. Webb

Methods

Figure 1 in the main text illustrates our framework that combines interpretable structural descriptors with probabilistic classification to identify local collective variables that distinguish amorphous phases. First, each molecule’s local environment is represented using atom-centered symmetry functions (ACSFs) and bond-orientational order (BOO) parameters, which together provide a comprehensive, symmetry-invariant characterization of local structure (Step 1). Second, a two-stage feature selection procedure uses mutual information (MI) to rank descriptors by discriminative power and removes redundant descriptors through correlation analysis (Step 2). Third, class-conditional probability distributions of the selected descriptors are then modeled, enabling transparent probabilistic classification (Step 3). Finally, the joint probability for each class is used for molecule-by-molecule state assignment and explicit detection of out-of-distribution configurations, i.e., molecular environments that cannot be classified as LDA-like or HDA-like (Step 4). This approach automatically identifies which structural descriptors best distinguish the amorphous ices without requiring a priori knowledge of the relevant order parameters, providing both high classification accuracy and physical interpretability. The following sections detail each step of this procedure.

Structural Descriptors

We consider two complementary, interpretable descriptor families that together provide a comprehensive characterization of microscopic environments. ACSFs encode the local environment through radially and angularly resolved correlations of interatomic distances [29]. BOO parameters [16] provide a complementary description that quantifies the orientational symmetry of the surrounding molecular arrangement and have been widely used to characterize amorphous and crystalline ices [17, 18]. Together, these descriptors form a symmetry-invariant and physically grounded representation of each molecule’s local environment (Figure 1, Step 1).

Atom-Centered Symmetry Functions

ACSFs are computed using several distinct functional forms that capture complementary aspects of local atomic environments through radially and angularly resolved correlations. All ACSF functions are centered on oxygen atoms (central oxygen atom ii), with contributions included from both oxygen and hydrogen neighbors to ensure sensitivity to variations in hydrogen-bond motifs and proton arrangements not visible from oxygen positions alone. Throughout, we denote rij\mathbf{r}_{ij} as the displacement vector from central atom ii to neighboring atom jj and Rij=∥rij∥R_{ij} = \lVert\mathbf{r}_{ij} \rVert as the corresponding scalar distance. All ACSF functions employ a smooth cutoff function fc(Rij)f_{\mathrm{c}}(R_{ij}):

fc(Rij)={0.5[cos⁡(πRijRc)+1]if Rij≤Rc0if Rij>Rc,(1)f_{\mathrm{c}(R_{ij})} = \begin{cases} 0.5 \left[\cos\left(\dfrac{\pi R_{ij}}{R_{\mathrm{c}}}\right) + 1\right] & \text{if } R_{ij} \le R_{\mathrm{c}} \\ 0 & \text{if } R_{ij} > R_{\mathrm{c}}, \end{cases} \tag*{(1)}

This cutoff function smoothly decays from 1 at Rij=0R_{ij} = 0 to 0 at Rij=RcR_{ij} = R_{\mathrm{c}}, with both the function and its first derivative vanishing continuously at the cutoff radius. We define the local environment around a central oxygen ii by the 16 nearest neighboring water molecules. Let Nmol(i)\mathcal{N}_{\mathrm{mol}}(i) denote this set of 16 neighbor molecules. We then define the corresponding sets of neighboring atoms as NO(i)\mathcal{N}_{\mathrm{O}}(i) and NH(i)\mathcal{N}_{\mathrm{H}}(i), the set of the 16 neighboring oxygen atoms and 32 neighboring hydrogen atoms belonging to the molecules in Nmol(i)\mathcal{N}_{\mathrm{mol}}(i). Unless otherwise stated, we set Rc≈5.0R_{\mathrm{c}} \approx5.0 Å and always ensure that all atoms in these neighbor sets satisfy Rij≤RcR_{ij} \le R_{\mathrm{c}} (i.e., the neighbor list is defined by neighbor count).

Three types of radial symmetry functions provide complementary representations of the local density. The simplest radial function, G1G^{1}, provides a weighted count of neighboring atoms:

G1=∑j∈Nγ(i)fc(Rij),(2)G^{1} = \sum_{j \in\mathcal{N}_{\gamma}(i)} f_{\mathrm{c}}(R_{ij}), \tag*{(2)}

where γ∈{O,H}\gamma\in\{\mathrm{O}, \mathrm{H}\} specifies the neighbor species set. In particular, we compute separate G1G^{1} functions for oxygen and hydrogen neighbors, denoted GO1G_{\mathrm{O}}^{1} and GH1G_{\mathrm{H}}^{1}.

The next function, G2G^{2}, focuses on the local density at a particular distance from the central atom,

G2(η,Rs)=∑j∈Nγ(i)exp⁡[−η(Rij−Rs)2]fc(Rij),(3)G^{2}(\eta,R_s)=\sum_{j\in N_\gamma(i)}\exp\left[-\eta(R_{ij}-R_s)^2\right]f_c(R_{ij}), \tag*{(3)}

where η\eta controls the width of the Gaussian and RsR_s shifts its center. By using multiple G2G^{2} functions with different η\eta and RsR_s values, these functions can construct a radially resolved representation of the local density. The third function, G3G^{3}, utilizes cosine damping:

G3(κ)=∑j∈Nγ(i)cos⁡(κRij)fc(Rij),(4)G^{3}(\kappa)=\sum_{j\in N_\gamma(i)}\cos(\kappa R_{ij})f_c(R_{ij}), \tag*{(4)}

where κ\kappa adjusts the period length. Using multiple G3G^{3} functions with different κ\kappa values provides a Fourier-like decomposition of the radial distribution.

Two types of angular functions capture information about three-body configurations and local orientational order. This is achieved by considering angles centered at atom ii formed by neighbor atoms jj and kk,

θjik=arccos⁡(rij⋅rikRijRik).(5)\theta_{jik}=\arccos\left(\frac{\mathbf{r}_{ij}\cdot\mathbf{r}_{ik}}{R_{ij}R_{ik}}\right). \tag*{(5)}

The two functions take the form (summing over unordered neighbor pairs j<kj<k):

G4(η,ζ,λ)=21−ζ∑j∈Nα(i)∑k∈Nβ(i)k>j(1+λcos⁡θjik)ζ×exp⁡[−η(Rij2+Rik2+Rjk2)]fc(Rij)fc(Rik)fc(Rjk),(6)G^{4}(\eta,\zeta,\lambda)=2^{1-\zeta}\sum_{j\in N_\alpha(i)}\sum_{\substack{k\in N_\beta(i)\\k>j}}(1+\lambda\cos\theta_{jik})^\zeta\times\exp\left[-\eta(R_{ij}^{2}+R_{ik}^{2}+R_{jk}^{2})\right]f_c(R_{ij})f_c(R_{ik})f_c(R_{jk}), \tag*{(6)}

and

G5(η,ζ,λ)=21−ζ∑j∈Nα(i)∑k∈Nβ(i)k>j(1+λcos⁡θjik)ζ×exp⁡[−η(Rij2+Rik2)]fc(Rij)fc(Rik).(7)G^{5}(\eta,\zeta,\lambda)=2^{1-\zeta}\sum_{j\in N_\alpha(i)}\sum_{\substack{k\in N_\beta(i)\\k>j}}(1+\lambda\cos\theta_{jik})^\zeta\times\exp\left[-\eta(R_{ij}^{2}+R_{ik}^{2})\right]f_c(R_{ij})f_c(R_{ik}). \tag*{(7)}

Here, α,β∈{O,H}\alpha,\beta\in\{\mathrm{O},\mathrm{H}\} specify the neighbor species sets used for jj and kk. The parameter ζ\zeta controls the angular resolution, with larger values emphasizing specific angular arrangements, while λ∈{−1,1}\lambda\in\{-1,1\} shifts the sensitivity to linear (λ=−1\lambda=-1) versus bent (λ=1\lambda=1) configurations. The parameter η\eta controls the radial extent over which angular correlations are evaluated. The key distinction between G4G^{4} and G5G^{5} lies in their radial dependencies. G4G^{4} includes the neighbor–neighbor distance RjkR_{jk}, restricting contributions to compact triplets where all three interatomic distances are small, while G5G^{5} omits this constraint and thus captures angular correlations over a wider range of geometries. These angular functions are computed over all possible species combinations of the three atoms involved, with parameter sets chosen to capture angular correlations at different radial extents.

Bond-Orientational Order Parameters

BOO parameters provide a complementary description of local structure by quantifying orientational symmetry through spherical harmonics [16]. Unlike ACSFs which are explicitly constructed as sums over neighbors, BOO parameters characterize the angular distribution of the local environment in a rotationally invariant manner. For each oxygen atom ii, we use the same oxygen neighbor list as above and define Nb=∣NO(i)∣=16N_b=\lvert N_{\mathrm{O}}(i)\rvert=16 neighboring oxygen atoms. The complex-valued order parameter qlmq_{lm} is computed as:

qlm(i)=1Nb∑j∈NO(i)Ylm(θij,ϕij),(8)q_{lm}(i)=\frac{1}{N_b}\sum_{j\in N_{\mathrm{O}(i)}}Y_{lm}(\theta_{ij},\phi_{ij}), \tag*{(8)}

where YlmY_{lm} are spherical harmonics of degree ll and order mm and (θij,ϕij)(\theta_{ij},\phi_{ij}) are the polar and azimuthal angles of the vector rij\mathbf{r}_{ij} connecting oxygen atoms ii and jj. The Steinhardt BOO parameter qlq_l is obtained by combining all mm components for a given ll:

ql(i)=4π2l+1∑m=−ll∣qlm(i)∣2.(9)q_l(i)=\sqrt{\frac{4\pi}{2l+1}\sum_{m=-l}^{l}|q_{lm}(i)|^2}. \tag*{(9)}

Low values of ll (e.g., l=3,4l=3,4) are sensitive to tetrahedral and other low-symmetry arrangements, while higher ll values capture more complex orientational motifs.

We also include third-order descriptors wlw_l, which provide additional sensitivity to the shape of the orientational distribution:

wl(i)=∑m1+m2+m3=0Clm1m2m3qlm1(i)qlm2(i)qlm3(i)(∑m=−ll∣qlm(i)∣2)3/2,(10)w_l(i)=\frac{\sum_{m_1+m_2+m_3=0}C_l^{m_1m_2m_3}q_{lm_1}(i)q_{lm_2}(i)q_{lm_3}(i)}{\left(\sum_{m=-l}^{l}|q_{lm}(i)|^2\right)^{3/2}}, \tag*{(10)}

where Clm1m2m3C_l^{m_1m_2m_3} are the Wigner 33-jj coupling coefficients and the sum runs over all triplets (m1,m2,m3)(m_1,m_2,m_3) satisfying m1+m2+m3=0m_1+m_2+m_3=0. The wlw_l parameters are particularly useful for distinguishing structures with similar qlq_l values but different higher-order orientational correlations.

Additionally, we compute locally averaged variants that incorporate information from the neighborhood surrounding each molecule [52]. The locally averaged complex order parameter qˉlm(i)\bar{q}_{lm}(i) is defined as:

qˉlm(i)=1Nb+1[qlm(i)+∑j∈NO(i)qlm(j)].(11)\bar{q}_{lm}(i)=\frac{1}{N_b+1}\left[q_{lm}(i)+\sum_{j\in N_O(i)}q_{lm}(j)\right]. \tag*{(11)}

This averaging procedure includes contributions from both the central molecule ii and its NbN_b oxygen neighbors. From qˉlm(i)\bar{q}_{lm}(i), we compute the rotationally invariant

qˉl(i)=4π2l+1∑m=−ll∣qˉlm(i)∣2.(12)\bar{q}_l(i)=\sqrt{\frac{4\pi}{2l+1}\sum_{m=-l}^{l}|\bar{q}_{lm}(i)|^2}. \tag*{(12)}

The locally averaged descriptors qˉl\bar{q}_l provide enhanced discrimination between crystal structures compared to their unaveraged counterparts qlq_l, as the averaging reduces thermal fluctuations while preserving sensitivity to orientational order [52]. The corresponding locally averaged third-order descriptors wˉl\bar{w}_l are computed using Eq. (10) but with all qlm(i)q_{lm}(i) substituted by qˉlm(i)\bar{q}_{lm}(i).

Together with the ACSF descriptors, this comprehensive set of structural descriptors provides a detailed, physically interpretable representation of each oxygen-centered local environment.

Automatic Feature Selection

We extract the most informative descriptors from the full set using an automated ranking procedure (Figure 1, Step 2). We define local environments using the 16 nearest neighbors, consistent with prior work [17, 18, 20]. We assess the ability of each descriptor to discriminate between classes (LDA-like and HDA-like) using mutual information (MI) [53]. MI quantifies the reduction in uncertainty about the class label upon observing a descriptor value and thus provides a model-free measure of discriminative power. For a feature XX and class label YY, where xx and yy denote particular values of XX and YY respectively, the MI is defined as:

MI⁡(X;Y)=∑x∈X∑y∈Yp(x,y)log⁡p(x,y)p(x)p(y),(13)\operatorname{MI}(X;Y)=\sum_{x\in X}\sum_{y\in Y}p(x,y)\log\frac{p(x,y)}{p(x)p(y)}, \tag*{(13)}

where p(x,y)p(x,y) is the joint probability density of feature XX and class label YY, and p(x)p(x) and p(y)p(y) are the corresponding distributions. We rank features by MI and retain those with MI values of at least 10% of the maximum observed MI, such that all features are expected to possess useful information.

To ensure that selected descriptors contribute distinct structural information, we use a greedy high-correlation filter. Proceeding in order of decreasing MI, each descriptor is retained only if its Pearson correlation (rr) with all previously retained descriptors satisfies ∣r∣<0.8|r| < 0.8. The outcome is a set of informative and roughly uncorrelated descriptors, which we denote D={1,…,M}\mathcal{D} = \{1,\ldots,M\}. Because the procedure operates on arbitrary descriptor inputs, it enables data-driven identification of discriminative descriptors without requiring prior assumptions about which structural motifs are important. We retain D\mathcal{D} in its entirety for out-of-distribution detection (see Probabilistic Model, below), since each additional weakly correlated descriptor offers an independent opportunity to identify an anomalous environment and thus maximizes detection sensitivity. For classification, however, the number of descriptors retained from D\mathcal{D} is a user-specified parameter; here we select the m=5m = 5 most informative descriptors, denoted Dcls={1,…,m}⊂D\mathcal{D}_{\mathrm{cls}} = \{1,\ldots,m\} \subset\mathcal{D}, balancing classification accuracy with physical interpretability. The full set of ACSF and BOO descriptors considered as candidates spans all functional forms defined in Section S1, and the specific descriptors selected for LDA/HDA classification are shown in Figure 2E of the main text.

Probabilistic Model

The second component of our framework is a probabilistic classifier that assigns a label to each atomic environment as a calibrated likelihood of belonging to each structural class (Figure 1, Step 3). Rather than learning a decision boundary, we model the probability distribution of each selected descriptor for each class.

To avoid index ambiguity, we use a∈{1,…,m}a \in\{1,\ldots,m\} to index the selected mm features in Dcls\mathcal{D}_{\mathrm{cls}} used for classification and n∈{1,…,Ny}n \in\{1,\ldots,N_y\} to index training samples within class yy. For each feature aa and class yy, we estimate the corresponding one-dimensional probability density pa,y(x)p_{a,y}(x) of observing value xx using Gaussian kernel density estimation (KDE). In kk-fold cross-validation, the dataset is partitioned into kk equal subsets. The model is trained on k−1k - 1 subsets and evaluated on the held-out subset, with this procedure repeated kk times so that every sample serves as a validation point exactly once. Performance metrics are then averaged across all kk folds, providing a robust estimate of generalization performance with reduced sensitivity to any particular data split. Throughout this work, we use k=5k = 5 (five-fold cross-validation). Kernel bandwidths σa\sigma_a are selected via a grid search that maximizes the marginal log-likelihood estimated by five-fold cross-validation.

Each KDE is normalized such that ∫−∞∞pa,y(x) dx=1\int_{-\infty}^{\infty} p_{a,y}(x)\,\mathrm{d}x = 1, where

pa,y(x)=1Ny∑n=1Ny12πσa2exp⁡[−(x−xa,n)22σa2].(14)p_{a,y}(x) = \frac{1}{N_y}\sum_{n=1}^{N_y}\frac{1}{\sqrt{2\pi\sigma_a^2}}\exp\left[-\frac{(x-x_{a,n})^2}{2\sigma_a^2}\right]. \tag*{(14)}

Here, NyN_y is the number of samples for class yy and xa,nx_{a,n} are the observed values of feature aa within that class.

Given an observed feature value xa∗x_a^*, we define the probability within a kernel window of half-width 0.5σa0.5\sigma_a centered at xa∗x_a^* as

πa,y(xa∗)=∫xa∗−0.5σaxa∗+0.5σapa,y(x) dx.(15)\pi_{a,y}(x_a^*) = \int_{x_a^*-0.5\sigma_a}^{x_a^*+0.5\sigma_a} p_{a,y}(x)\,\mathrm{d}x. \tag*{(15)}

For an environment characterized by mm selected features, x⃗∗=(x1∗,x2∗,…,xm∗)\vec{x}^* = (x_1^*,x_2^*,\ldots,x_m^*), we compute a class-conditional likelihood score under a naïve Bayes assumption (Figure 1, Step 4):

Py(x⃗∗)=∏a=1mπa,y(xa∗).(16)\mathcal{P}_y(\vec{x}^*) = \prod_{a=1}^{m}\pi_{a,y}(x_a^*). \tag*{(16)}

Each environment is assigned the class with the largest likelihood score Py(x⃗∗)\mathcal{P}_y(\vec{x}^*).

A further advantage of this probabilistic formulation is explicit out-of-distribution (OOD) detection. To maximize sensitivity, OOD detection is evaluated over the complete descriptor set D={1,…,M}⊇Dcls\mathcal{D} = \{1,\ldots,M\} \supseteq\mathcal{D}_{\mathrm{cls}} (see Automatic Feature Selection, above) rather than the mm descriptors used for classification alone. For each descriptor d∈Dd \in\mathcal{D} and class yy, we construct pd,y(x)p_{d,y}(x) following the same KDE procedure described above, with bandwidth σd\sigma_d selected independently via the same five-fold cross-validated grid search. We define an environment as an outlier if pd,y(x)<Pcutp_{d,y}(x) < P_{\mathrm{cut}} for all classes yy, for any descriptor d∈Dd \in\mathcal{D}. This criterion ensures that the vast majority of in-distribution configurations are retained while enabling detection of novel or anomalous environments, a capability essential for analyzing structural transformations. In this work, we choose Pcut=5×10−4P_{\mathrm{cut}} = 5 \times10^{-4}. We find that our results are robust up to Pcut=1×103P_{\mathrm{cut}} = 1 \times10^{3}, beyond which outlier configurations are observed during compression/decompression trajectories. This threshold can be adjusted to make the classifier more or less conservative, depending on the application requirements.

Training Data and Simulation Details

All configurations were obtained from the Deep Potential (DP) MD trajectories reported in Ref. 24. Two machine-learned potentials were employed, providing complementary representations of water’s potential energy surface. The first, DP-SCAN, was trained on electronic-structure data generated with the SCAN density functional [38, 39], a meta-GGA functional known to reproduce many of water’s structural and thermodynamic properties. The second, DP-MBpol, was trained on the many-body polarizable (MBpol) potential [37, 46, 47, 48], which is parameterized against coupled-cluster reference data and is among the most accurate models of water available. All simulations were performed using LAMMPS (2 Aug 2023) [49] interfaced with DeePMD-kit (v3.0.0) [50, 51], with a timestep of 0.5 fs. Simulations employed a system of 512 water molecules in a cubic box with periodic boundary conditions and the temperature and pressure were controlled via a Nosé–Hoover thermostat and barostat, respectively.

LDA configurations were generated by isobaric quenching from equilibrated liquid water configurations at constant cooling rates of qc=1q_c = 1 K/ns and qc=10q_c = 10 K/ns, at 1 bar. Isothermal compression and decompression simulations were conducted starting from LDA configurations at constant rates of qp=300q_p = 300 bar/ns or qp=3000q_p = 3000 bar/ns. For each pressure and cooling rate, at least three independent simulations were performed to assess statistical variance.

The training and evaluation dataset comprises of 10,000 molecular environments, equally divided between LDA and HDA. Configurations were selected from thermodynamic conditions where phase identity is unambiguous, avoiding the transition region to ensure reliable labels. LDA environments were drawn from isobaric quenches at temperatures at least 30 K below the glass transition and from isothermal compression at pressures at least 3 kbar below the transformation pressure to HDA. HDA environments were taken from compression trajectories at least 3 kbar above the LDA-to-HDA transformation pressure and from decompression trajectories, the latter restricted to densities exceeding that of recovered HDA at 80 K and 1 bar to exclude partially transformed structures. As a benchmark for OOD detection, we curated 1,000 hexagonal ice (Ih) configurations. Ice Ih shares the tetrahedral hydrogen-bond network of LDA but exhibits long-range crystalline order absent in either amorphous phase, making it well-suited for evaluating OOD detection.

Benchmarking Model Performance for DP_SCAN Environments

Martelli and co-workers introduced a supervised neural network (BOO-NN) trained exclusively on bond-orientational order parameters, learning nonlinear mappings from BOO features to phase labels and adopting the traditional notion that orientational symmetry is the key structural discriminator between amorphous phases [17, 18]. DeFever et al. proposed a conceptually different strategy with PointNet [31], which bypasses predefined descriptors altogether by learning symmetry-invariant representations directly from atomic coordinates using a permutation-invariant architecture; although highly successful for classifying crystallization and melting pathways, PointNet has not previously been applied to amorphous ices. Boattini and co-workers developed an unsupervised autoencoder-based model [35] that compresses BOO descriptors into a low-dimensional latent space before identifying structural motifs via Gaussian mixture modeling. To enable a supervised comparison, we adapt this method (AE-GMM) by fitting class-conditional multivariate Gaussians to the learned embeddings, capturing nonlinear correlations among BOO features while introducing a latent representation that is less directly interpretable. Collectively, these baselines cover descriptor-based neural networks, coordinate-level representation learning, and latent-space probabilistic modeling, offering a comprehensive set of benchmarks for contextualizing the strengths of our probabilistic classifier.

Here, we provide precisions, recalls, and accuracies of BOO-NN, PointNet, AE-GMM, and the present work on LDA/HDA classification for local atomic environments generated by DP_MBpol (Table S1) and DP_SCAN (Table S2).

Performance metrics are lower for classifying environments generated by DP_SCAN than they are for DP_MBpol. Table S1 shows that accuracies for some models are greater than 0.99 for DP_MBpol environments, while here all accuracies are roughly between 0.96-0.97 for DP_SCAN environments. The relative performance of the models, however, is similar; BOO-NN and our work perform best, followed by PointNet and AE-GMM. These results agree with what is observed for DP_MBpol environments: our work maintains comparable accuracy to deep learning methods while adopting a physically interpretable classification scheme. Interestingly, BOO-NN, AE-GMM, and our work show higher HDA recall and LDA precision than HDA precision and LDA recall. This indicates that these models are liberal in their assignment of HDA but conservative in their assignment of LDA, suggesting a relationship between HDA and LDA structures for DP_SCAN not present in DP_MBpol.

Out-of-distribution Predictions for DP_SCAN and DP_MBpol Environments

In the main text, Section 2.1. we analyze the behavior of BOO-NN, PointNet, AE-GMM, and this work when making predictions on out-of-distribution atomic environments. Specifically, models are trained on HDA and LDA atomic environments, but are used to predict the labels of hexagonal ice (Ih) atomic environments. In the main text, we only perform this analysis for environments generated by DP_MBpol. Here, we report the same results for DP_SCAN configurations.

As observed in the main text, BOO-NN (Figure S1A) and PointNet (Figure S1B) struggle to separate LDA and Ih distributions, and these models confidently label Ih environments as LDA. Accordingly, the present work is capable of accurately distinguishing Ih environments as outliers from those on which the model was trained (i.e., LDA and HDA), as shown by the separation between the distributions in Figure S1D. Interestingly, while AE-GMM (Figure S1C) was capable of detecting outliers for DP_MBpol environments, this model struggles to do so for configurations generated by DP_SCAN. Such a result demonstrates the utility of considering all available features, as is done in our work, relative to density estimation on reduced feature spaces, as is done by AE-GMM.

Figure S1. Predicted LDA log probabilities from (A) BOO-NN, (B) PointNet, (C) AE-GMM, and (D) this work applied to DP_SCAN-generated environments of LDA and hexagonal ice (Ih). Distributions for true LDA configurations are shown in red, while distributions for Ih configurations are shown in purple. Each density is normalized so that its maximum value is 1.0.

Out-of-distribution Prediction for Alternative Phases of Ice

Our outlier detection method identifies local atomic environments that differ from the LDA/HDA training distribution along any structural dimension. To evaluate how the choice of features affects out-of-distribution (OOD) accuracy, we compare two approaches applied to an expanded dataset of ice polymorphs obtained from the datasets which were used in the training of the DP_SCAN and DP_MBpol models [37, 39].

The first approach uses only the five selected features most informative for LDA/HDA classification. As discussed in the main text, this results in a sparse, interpretable model whose features directly reflect what distinguishes the two amorphous states from one another. The second approach includes all available features after removing pairwise-correlated descriptors (Pearson ∣r∣>0.8|r| > 0.8), retaining a diverse, uncorrelated subset of the full 120-dimensional descriptor space. This is not equivalent to using all 120 descriptors simultaneously. Including strongly correlated features in a naive Bayes product model would double-count shared structural information and degrade OOD sensitivity.

Table S3 shows OOD detection rates for both approaches across 14 ice polymorphs. Using the full uncorrelated descriptor set outperforms the five-feature model in most cases, with several phases moving from near-zero to near-complete detection. This indicates that the combined ACSF and BOO descriptor set contains structural information sufficient to distinguish the majority of ice polymorphs from amorphous ice. A subset of high-pressure phases (e.g., Ice II, IV, VI, and XV) remain only partially distinguishable from LDA/HDA even when using the full descriptor set.

PhaseDP_MBpol n=5n = 5DP_MBpol n=120n = 120DP_SCAN n=5n = 5DP_SCAN n=120n = 120
Ice Ih1.0001.0001.0001.000
Ice Ic1.0001.0000.9980.998
Ice II0.0000.1220.0030.378
Ice III0.0090.9330.0011.000
Ice IV0.0120.0330.0480.071
Ice V0.0030.8720.0640.998
Ice VI0.0240.1340.0870.250
Ice VII1.0001.0001.0001.000
Ice VIII1.0001.0001.0001.000
Ice IX0.0050.9930.0001.000
Ice XI0.9930.9981.0001.000
Ice XII0.0620.9910.301.000
Ice XIII0.0060.7630.0520.980
Ice XIV0.0310.9870.3121.000
Ice XV0.0610.2140.2640.470

Table S3. Accuracy of classifiers that use different numbers of features for OOD classification. Models are used to classify structures of different ice phases generated by the DP MBpol and DP SCAN water models. Models can either use five features relevant to LDA/HDA classification (n=5n = 5) or use all uncorrelated provided structural descriptors (n=120n = 120).

This residual ambiguity is a consequence of the descriptor selection procedure used here, rather than the local environment representation itself. At the level of individual atomic neighborhoods, some high-pressure ice phases share geometric features with LDA/HDA that are not fully resolved by the descriptors identified as optimally distinguish LDA and HDA. This does not indicate that these crystalline phases are physically present within computational LDA or HDA. Independent molecular dynamics studies of LDA and HDA have found no evidence of ice-like domains in either amorphous phase [40]. Taken together, these results suggest that the observed classifier ambiguity for certain high-pressure ices arises from local geometric similarity between amorphous and crystalline packing motifs along the LDA/HDA classification axis, rather than from partially crystallized regions embedded within the amorphous structures. Distinguishing such phases would be possible if a classifier were directly trained to resolve the amorphous states from the crystalline ones.

Sensitivity of Structural Pathways to Selected Feature Spaces

In the main text, Section 2.6 analyzes how the evolution of average atomic environments proceeds during compressions and decompressions across a range of trajectories at different temperatures. We show these evolutions for DP_MBpol and DP_SCAN for two different feature spaces. Here, we show these evolutions for 25 feature spaces comprised of each pair of features selected by our model to be most informative for accurate LDA/HDA classification.

Results from Figure S2 and Figure S3 corroborate the results discussed in Section 2.6. Namely, we observe a hysteresis in environment pathways when considering feature spaces that consist of an ACSF and BOO, but this separation is not visible when visualizing pairs of BOOs. Therefore, it is necessary to consider both radial and spherical information to identify differences in the structural pathways taken by compressions and decompressions—an observation that would go unnoticed if we relied solely on the BOOs employed in previous studies.

Figure S2. Evolution of average local environment descriptors during compressions (blue) and decompressions (red) at temperatures from 60 K to 140 K for the DP_MBpol force field. Descriptors are plotted in feature spaces for every pair of features selected by our model for LDA/HDA classification. Darker red/blue shades correspond to lower temperatures.

Figure S3. Evolution of average local environment descriptors during compressions (blue) and decompressions (red) at temperatures from 60 K to 140 K for the DP SCAN force field. Descriptors are plotted in feature spaces for every pair of features selected by our model for LDA/HDA classification. Darker red/blue shades correspond to lower temperatures.

Differences between DP_SCAN- and DP_MBpol-generated HDA Environments

In the main text, Section 2.7 analyzes structural differences between LDA environments generated by DP_SCAN and DP_MBpol. Here, we consider structural differences between HDA environments generated by the same two force fields. Our classifier can distinguish between DP_SCAN and DP_MBpol HDA environments with an accuracy of 64.3±0.03%64.3 \pm0.03\%, indicating that there are structural differences (albeit less than LDA) present in these environments. We consider how distributions of DP_SCAN and DP_MBpol environments vary along a select set of structural descriptors with high mutual information for DP_SCAN/DP_MBpol classification, which we visualize in Figure S4.

Figure S4A and Figure S4B show distributions of DP_SCAN and DP_MBpol environments along the two most informative features identified by our classifier. One can clearly observe differences in the two distributions, suggesting that (similar to LDA environments) DP_SCAN and DP_MBpol disagree most in their reproduction of interatomic distances. This is most directly observed when considering the second descriptor, G2O(3.0,0.0)G_{2}^{\mathrm{O}}(3.0, 0.0), where higher values indicate reduced oxygen–oxygen pairwise distances. Figure S4C and Figure S4D show distributions of DP_SCAN and DP_MBpol environments along the two most informative BOOs. HDA environments appear to have similar distributions along these descriptors, indicating that DP_SCAN and DP_MBpol predict similar spherical symmetry in HDA environments.

References

  1. [2]P. G. Debenedetti and F. H. Stillinger, Supercooled liquids and the glass transition, Nature 410, 259 (2001).pubmed.ncbi.nlm.nih.gov/11258381
  2. [3]L. Berthier and G. Biroli, Theoretical perspective on the glass transition and amorphous materials, Rev. Mod. Phys. 83, 587 (2011).DOI
  3. [4]O. Mishima, L. Calvert, and E. Whalley, ‘melting ice’ at 77 K and 10 kbar: a new method of making amorphous solids, Nature 310, 393 (1984).DOI
  4. [5]O. Mishima, L. Calvert, and E. Whalley, An apparently first-order transition between two amorphous phases of ice induced by pressure, Nature 314, 76 (1985).DOI
  5. [6]P. H. Poole, F. Sciortino, U. Essmann, and H. E. Stanley, Phase behaviour of metastable water, Nature 360, 324 (1992).DOI
  6. [7]P. Gallo, K. Amann-Winkel, C. A. Angell, M. A. Anisimov, F. Caupin, C. Chakravarty, E. Lascaris, T. Loerting, A. Z. Panagiotopoulos, J. Russo, et al., Water: A tale of two liquids, Chem. Rev. 116, 7463 (2016).DOI
  7. [8]P. G. Debenedetti, F. Sciortino, and G. H. Zerze, Second critical point in two realistic models of water, Science 369, 289 (2020).DOI
  8. [9]K. Amann-Winkel, C. Gainaru, P. H. Handle, M. Seidl, H. Nelson, R. Böhmer, and T. Loerting, Water’s second glass transition, Proc. Natl. Acad. Sci. U.S.A. 110, 17720 (2013).
  9. [10]K. H. Kim, K. Amann-Winkel, N. Giovambattista, A. Späh, F. Perakis, H. Pathak, M. L. Parada, C. Yang, D. Mariedahl, T. Eklund, et al., Experimental observation of the liquid-liquid transition in bulk supercooled water under pressure, Science 370, 978 (2020).DOI
  10. [11]J. Finney, A. Hallbrucker, I. Kohl, A. Soper, and D. Bowron, Structures of high and low density amorphous ice by neutron diffraction, Phys. Rev. Lett. 88, 225503 (2002).pubmed.ncbi.nlm.nih.gov/12059427
  11. [12]T. Loerting, W. Schustereder, K. Winkel, C. G. Salzmann, I. Kohl, and E. Mayer, Amorphous ice: Stepwise formation of very-high-density amorphous ice from low-density amorphous ice at 125 K, Phys. Rev. Lett. 96, 025702 (2006).pubmed.ncbi.nlm.nih.gov/16486598
  12. [13]P.-L. Chau and A. Hardwick, A new order parameter for tetrahedral configurations, Mol. Phys. 93, 511 (1998).DOI
  13. [14]J. R. Errington and P. G. Debenedetti, Relationship between structural order and the anomalies of liquid water, Nature 409, 318 (2001).pubmed.ncbi.nlm.nih.gov/11201735
  14. [15]B. S. Jabes, M. Agarwal, and C. Chakravarty, Tetrahedral order, pair correlation entropy, and waterlike liquid state anomalies: Comparison of GeO2 with BeF2, SiO2, and H2O, J. Chem. Phys. 132, 234507 (2010).DOI
  15. [16]P. J. Steinhardt, D. R. Nelson, and M. Ronchetti, Bond-orientational order in liquids and glasses, Phys. Rev. B 28, 784 (1983).DOI
  16. [17]F. Martelli, F. Leoni, F. Sciortino, and J. Russo, Connection between liquid and non-crystalline solid phases in water, J. Chem. Phys. 153, 104503 (2020).DOI
  17. [18]Z. Faure Beaulieu, V. L. Deringer, and F. Martelli, High-dimensional order parameters and neural network classifiers applied to amorphous ices, J. Chem. Phys. 160, 081101 (2024).DOI
  18. [19]R. Foffi and F. Sciortino, Correlated fluctuations of structural indicators close to the liquid-liquid transition in supercooled water, J. Phys. Chem. B 127, 378 (2022).DOI
  19. [20]E. D. Donkor, A. Offei-Danso, A. Rodriguez, F. Sciortino, and A. Hassanali, Beyond local structures in critical supercooled water through unsupervised learning, J. Phys. Chem. Lett. 15, 3996 (2024).DOI
  20. [21]S. Klotz, T. Strässle, R. Nelmes, J. Loveday, G. Hamel, G. Rousse, B. Canny, . f. J. Chervin, and A. Saitta, Nature of the polymorphic transition in ice under pressure, Phys. Rev. Lett. 94, 025506 (2005).pubmed.ncbi.nlm.nih.gov/15698193
  21. [22]J. Chiu, F. W. Starr, and N. Giovambattista, Pressure-induced transformations in computer simulations of glassy water, J. Chem. Phys. 139, 184504 (2013).DOI
  22. [23]J. Wong, D. A. Jahn, and N. Giovambattista, Pressure-induced transformations in glassy water: A computer simulation study using the TIP4P/2005 model, J. Chem. Phys. 143, 074501 (2015).DOI
  23. [24]R. J. Szulako, N. Giovambattista, and P. G. Debenedetti, Computational investigation of water glasses using machine-learning potentials, Proc. Natl. Acad. Sci. U.S.A. 122, e2509609122 (2025).
  24. [25]Q.-S. Zeng, Y. Ding, W. L. Mao, W. Yang, S. V. Sinogeikin, J. Shu, H.-k. Mao, and J. Jiang, Origin of pressure-induced polymorphism in Ce 75 Al 25 metallic glass, Phys. Rev. Lett. 104, 105702 (2010).DOI
  25. [26]Q. Du, X.-J. Liu, Q. Zeng, H. Fan, H. Wang, Y. Wu, S.-W. Chen, and Z.-P. Lu, Polyamorphic transition in a transition metal based metallic glass under high pressure, Phys. Rev. B 99, 014208 (2019).
  26. [27]D. Dhabal and V. Molinero, Kinetics and mechanisms of pressure-induced ice amorphization and polyamorphic transitions in a machine-learned coarse-grained water model, J. Phys. Chem. B 127, 2847 (2023).DOI
  27. [28]G. Ramesh, V. Mahajan, D. Koner, and R. S. Singh, Microscopic pathways of transition from low-density to high-density amorphous phase of water, J. Chem. Phys. 160, 194501 (2024).arxiv.org/abs/2402.16761
  28. [29]J. Behler, Atom-centered symmetry functions for constructing high-dimensional neural network potentials, J. Chem. Phys. 134, 074106 (2011).DOI
  29. [30]A. P. Bartók, R. Kondor, and G. Csányi, On representing chemical environments, Phys. Rev. B 87, 184115 (2013).DOI
  30. [31]R. S. DeFever, C. Targonski, S. W. Hall, M. C. Smith, and S. Sarupria, A generalized deep learning approach for local structure identification in molecular simulations, Chem. 10, 7503 (2019).DOI
  31. [32]J. Schmidt, M. R. Marques, S. Botti, and M. A. Marques, Recent advances and applications of machine learning in solid-state materials science, npj Comput. Mat. 5, 83 (2019).DOI
  32. [33]M. Ceriotti, Unsupervised machine learning in atomistic simulations, between predictions and understanding, J. Chem. Phys. 150, 150901 (2019).DOI
  33. [34]F. Musil, A. Grisafi, A. P. Bartók, C. Ortner, G. Csányi, and M. Ceriotti, Physics-inspired structural representations for molecules and materials, Chem. 121, 9759 (2021).DOI
  34. [35]E. Boattini, M. Dijkstra, and L. Filion, Unsupervised learning for local structure detection in colloidal systems, J. Chem. Phys. 151, 154901 (2019).
  35. [36]C. Cai and T. Wang, Resolving chemical-motif similarity with enhanced atomic structure representations for accurately predicting descriptors at metallic interfaces, Nat. Comm. 16, 8761 (2025).DOI
  36. [37]S. L. Bore and F. Paesani, Realistic phase diagram of water from “first principles” data-driven quantum simulations, Nat. Comm. 14, 3349 (2023).DOI
  37. [38]J. Sun, A. Ruzsinszky, and J. P. Perdew, Strongly constrained and appropriately normed semilocal density functional, Phys. Rev. Lett. 115, 036402 (2015).arxiv.org/abs/1504.03028
  38. [39]L. Zhang, H. Wang, R. Car, and W. E, Phase diagram of a deep potential water model, Phys. Rev. Lett. 126, 236001 (2021).DOI
  39. [40]F. Martelli, N. Giovambattista, S. Torquato, and R. Car, Searching for crystal-ice domains in amorphous ices, Phys. Rev. Mater. 2, 075601 (2018).DOI
  40. [41]N. Giovambattista, H. E. Stanley, and F. Sciortino, Potential-energy landscape study of the amorphous-amorphous transformation in H₂O, Phys. Rev. Lett. 91, 115504 (2003).DOI
  41. [42]N. Giovambattista, F. Sciortino, F. W. Starr, and P. H. Poole, Potential energy landscape of the apparent first-order phase transition between low-density and high-density amorphous ice, J. Chem. Phys. 145, 224501 (2016).DOI
  42. [43]M. A. Anisimov, M. Duška, F. Caupin, L. E. Amrhein, A. Rosenbaum, and R. J. Sadus, Thermodynamics of fluid polyamorphism, Phys. Rev. X 8, 011004 (2018).
  43. [44]H. Tanaka, Thermodynamic anomaly and polyamorphism of water, EPL 50, 340 (2000).DOI
  44. [45]H. Tanaka, Liquid–liquid transition and polyamorphism, J. Chem. Phys. 153, 130901 (2020).DOI
  45. [46]V. Babin, C. Lefortestier, and F. Paesani, Development of a “first principles” water potential with flexible monomers: Dimer potential energy surface, VRT spectrum, and second virial coefficient, J. Chem. Theory Comput. 9, 5395 (2013).DOI
  46. [47]V. Babin, G. R. Medders, and F. Paesani, Development of a “first principles” water potential with flexible monomers. ii: Trimer potential energy surface, third virial coefficient, and small clusters, J. Chem. Theory Comput. 10, 1599 (2014).DOI
  47. [48]G. R. Medders, V. Babin, and F. Paesani, Development of a “first-principles” water potential with flexible monomers. iii. liquid phase properties, J. Chem. Theory Comput. 10, 2906 (2014).DOI
  48. [49]A. P. Thompson, H. M. Aktulga, R. Berger, D. S. Bolintineanu, W. M. Brown, P. S. Crozier, P. J. In’t Veld, A. Kohlmeyer, S. G. Moore, T. D. Nguyen, et al., LAMMPS-a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales, Comput. Phys. Commun. 271, 108171 (2022).DOI
  49. [50]L. Zhang, J. Han, H. Wang, R. Car, and W. E, Deep potential molecular dynamics: a scalable model with the accuracy of quantum mechanics, Phys. Rev. Lett. 120, 143001 (2018).DOI
  50. [51]H. Wang, L. Zhang, J. Han, and E. Weinan, DeePMD-kit: A deep learning package for many-body potential energy representation and molecular dynamics, Comput. Phys. Commun. 228, 178 (2018).DOI
  51. [52]W. Lechner and C. Dellago, Accurate determination of crystal structures based on averaged local bond order parameters, J. Chem. Phys. 129, 114707 (2008).DOI
  52. [53]C. E. Shannon, A mathematical theory of communication, The Bell System Technical Journal 27, 379 (1948).

Paper details

vlnc:W_d4x6r7e9
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