A Local Structural Basis to Resolve Amorphous Ices

By applying a new probabilistic data-driven framework to molecular simulations of water, this study reveals that the distinction between low-density and high-density amorphous ices is encoded within the first coordination shell and that their pressure-induced transition occurs via a first-order-like redistribution of local environments without intermediate structures.

The Gist

The Big Picture: Sorting the "Messy" Ice

Imagine you have two piles of snow. One pile is fluffy and light (Low-Density Amorphous ice, or LDA), and the other is packed down tight and heavy (High-Density Amorphous ice, or HDA).

In a perfect crystal (like a snowflake), it's easy to tell them apart because they have a neat, repeating pattern. But these "amorphous" ices are messy; they look like random jumbles of water molecules. Scientists have long wondered: What is the specific, tiny difference between a molecule in the fluffy pile versus the heavy pile? And when you squeeze the fluffy ice to turn it into heavy ice, does it slowly morph, or does it snap into a new shape?

This paper acts like a high-tech detective that looks at the microscopic neighborhood of every single water molecule to solve these mysteries.

The Detective Tool: A "Smart Neighborhood Watch"

The researchers built a new computer program to act as a "Neighborhood Watch" for water molecules.

1. The Neighborhood: Instead of looking at the whole pile of ice, the program zooms in on one water molecule and looks at its 16 closest neighbors.
2. The ID Cards: It creates a "profile" for each neighborhood using two types of data:
   • Who is there? (Counting how many hydrogen and oxygen atoms are nearby).
   • How are they standing? (Measuring the angles and symmetry of the group).
3. The Filter: The program is smart enough to ignore the boring details and focus only on the clues that actually tell the difference between the "fluffy" and "heavy" ice.

Key Discovery 1: It's All About the "Extra Guests"

The biggest surprise was finding out what actually distinguishes the two types of ice.

• The Old Theory: Scientists thought you needed to look at the whole neighborhood (even the second or third ring of neighbors) to tell them apart.
• The New Finding: You only need to look at the immediate circle of neighbors (the first shell).
• The Metaphor: Imagine a party. In the "fluffy" ice (LDA), the guests are standing in a perfect, open circle with plenty of space. In the "heavy" ice (HDA), the party is still in the same room, but extra guests (water molecules) have squeezed into the gaps between the original guests.
• The Result: The most important clue isn't how the molecules are standing (their angles); it's simply how crowded the immediate area is. If there are extra "interstitial" guests jammed into the first circle, it's HDA. If the circle is open and orderly, it's LDA.

Key Discovery 2: The "Snap" Transformation

When you squeeze the fluffy ice to turn it into heavy ice, what happens?

• The Question: Does the ice slowly change shape, passing through a weird "middle stage" (like a half-fluffy, half-heavy mess)?
• The Answer: No. The paper found no middle ground.
• The Metaphor: Imagine a room full of people. When you squeeze the room, the people don't slowly shuffle into a new formation. Instead, the room suddenly splits: some people stay in their original "fluffy" spots, while others instantly jump into the "heavy" spots.
• The Result: The transformation is a redistribution. The ice doesn't turn into a new, weird type of ice in the middle. It just becomes a mix of "fluffy" molecules and "heavy" molecules. This proves the change is a sharp "snap" (like a first-order phase transition) rather than a slow, gradual slide.

Key Discovery 3: The Path Matters (Hysteresis)

The paper also looked at what happens when you squeeze the ice (compression) versus when you let it go (decompression).

• The Metaphor: Think of walking up a hill versus walking down the same hill.
  • Going Up (Compression): The molecules get squeezed, and the "extra guests" jam in. The structure collapses in a specific way.
  • Going Down (Decompression): When you release the pressure, the molecules don't just retrace their steps. They take a different path back to the fluffy state. They have to expand a lot before they can "un-jam" and return to their original open positions.
• The Result: The journey up is not the same as the journey down. This explains why the ice behaves differently depending on whether you are squeezing it or releasing it.

Key Discovery 4: Different "Recipes" Make Different Ice

The researchers tested two different computer models (simulations) of water. Even though both models were trying to simulate the same "fluffy" ice, they produced slightly different results.

• The Metaphor: Imagine two chefs making the same cake. One uses a slightly different flour, and the other uses a different sugar. Even though the cakes look the same from a distance, if you taste a single crumb, you can tell which chef made it.
• The Result: The computer program could tell the difference between the "fluffy ice" made by Chef A and the "fluffy ice" made by Chef B. This shows that the tiny details of how the water molecules pack together depend on the specific "recipe" (force field) used to simulate them.

Summary

This paper used a smart, data-driven detective to look at the microscopic neighborhoods of water molecules. It found that:

1. Crowding is key: The difference between light and heavy amorphous ice is simply how many extra water molecules are jammed into the immediate neighborhood.
2. No middle ground: When ice transforms, it doesn't become a weird hybrid; it just splits into a mix of "before" and "after" molecules.
3. Different paths: Squeezing the ice and releasing it follow different microscopic routes.

This helps scientists understand the fundamental rules of how water behaves when it's frozen in messy, glass-like states.

Technical Summary

Technical Summary: A Local Structural Basis to Resolve Amorphous Ices

Problem Statement
While crystalline phases are distinguished by unit cells and space groups, the structural basis differentiating amorphous phases remains elusive. Specifically, for water's low-density amorphous (LDA) and high-density amorphous (HDA) ices, it is unclear how their distinct thermodynamic properties emerge from microscopic configurations. Key open questions include: (1) Which specific local structural descriptors most effectively encode the distinction between LDA and HDA? (2) Over what length scales are these differences encoded (i.e., is information beyond the first coordination shell required)? and (3) Does the pressure-induced transformation between these phases proceed through structurally distinct intermediate states or via a redistribution of existing local motifs? Existing analysis strategies often struggle to answer these simultaneously; simple order parameters may be biased by construction, while high-dimensional machine learning models often lack interpretability and the ability to detect out-of-distribution (OOD) configurations.

Methodology
The authors present a systematic, interpretable, probabilistic machine learning framework applied to molecular dynamics (MD) simulation data of water. The methodology consists of four steps:

1. Descriptor Generation: Local atomic environments are represented using two complementary, symmetry-invariant descriptor families: Atom-Centered Symmetry Functions (ACSFs), which capture radially and angularly resolved correlations of interatomic distances, and Bond-Orientational Order (BOO) parameters, which quantify orientational symmetry via spherical harmonics.
2. Feature Selection: A two-stage procedure ranks descriptors by discriminative power using Mutual Information (MI) and removes redundant descriptors via a correlation filter (retaining features with $|r| < 0.8$). This is performed without a priori assumptions about relevant order parameters.
3. Probabilistic Classification: Class-conditional probability distributions for the selected descriptors are modeled using Gaussian Kernel Density Estimation (KDE). A Naïve Bayes assumption is used to compute joint likelihoods for phase assignment. Crucially, this framework includes an explicit mechanism to detect OOD configurations by identifying environments that fall below a probability threshold for all known classes.
4. Data Source: The framework is applied to 10,000 molecular environments from MD trajectories generated using two distinct machine-learned potentials: DP SCAN (trained on SCAN density functional data) and DP MBpol (trained on the many-body polarizable MBpol potential).

Key Contributions and Results

• Discriminative Descriptors: The analysis reveals that descriptors related to local hydrogen density dominate the discrimination between LDA and HDA. Specifically, the ACSF descriptor $G_1^H$ (a weighted count of hydrogen neighbors) exhibits the highest Mutual Information, surpassing orientational symmetry metrics. While BOO parameters (e.g., $q_3, q_4$) are also informative, the primary distinction is driven by density variations within the local environment.
• Local Encoding of Phase Identity: Contrary to prior suggestions that distinguishing amorphous water phases requires correlations extending beyond the first coordination shell, the authors find that phase identity is encoded within the first coordination shell. Classification accuracy reaches >95% using environments containing only 8–16 nearest neighbors and remains substantial (>86%) with as few as two neighbors. The distinction is primarily driven by the insertion of interstitial water molecules into the first shell in HDA, disrupting the tetrahedral network of LDA.
• Mechanism of Transformation: Analysis of compression and decompression trajectories at $T=80$ K reveals that the LDA–HDA transition proceeds entirely through a redistribution of discrete LDA-like and HDA-like environments. No evidence of structurally distinct intermediate states or novel motifs was found. This supports a first-order-like transition mechanism where the system shifts populations between two distinct structural basins on the potential energy landscape.
• Microscopic Hysteresis: While the transformation appears reversible when viewed solely through BOO parameters, the inclusion of local density descriptors ($G_1^H$) reveals pronounced microscopic hysteresis. Compression and decompression follow distinct pathways: compression involves an initial elastic densification followed by a collapse of tetrahedral order, whereas decompression involves a significant density decrease before the recovery of tetrahedral order.
• Force Field Sensitivity: The framework successfully distinguishes LDA environments generated by DP SCAN versus DP MBpol, achieving ~77.5% accuracy. The differences are primarily radial (packing density) rather than orientational, with DP SCAN producing slightly more compact local networks.

Significance
The paper establishes a well-defined, local structural basis for distinguishing LDA and HDA ices, resolving ambiguities inherent in analyses based solely on bulk observables. By demonstrating that the transformation occurs without intermediates, the work provides direct molecular-scale evidence for the first-order character of the glass–glass transition in water. The proposed probabilistic framework offers a general, interpretable recipe for characterizing systems where distinct amorphous or disordered phases exist but lack obvious distinguishing features. Its ability to detect OOD configurations ensures that the absence of intermediates is a physical result rather than an artifact of limited classification capabilities. Furthermore, the approach enables the comparison of molecular models at the environment level, revealing subtle structural differences between force fields that conventional metrics might obscure.