Flat interface between amorphous ices and the role of MDA-like intermediate states in the LDA-HDA transformation

This study utilizes a flat LDA||HDA interface and SOAP-based neural network analysis to demonstrate that MDA-like intermediate states are not a distinct bulk phase but rather localize at the interface, exhibiting an elastic response with kinetic hysteresis during the pressure-induced polyamorphic transformation.

The Gist

Imagine water as a crowd of people at a party. Sometimes, they stand far apart in a relaxed, open circle (this is Low-Density Amorphous Ice, or LDA). Other times, they are squeezed tightly together in a dense, chaotic huddle (this is High-Density Amorphous Ice, or HDA).

For decades, scientists have been trying to figure out exactly how the crowd moves from the relaxed circle to the tight huddle when you squeeze the room (apply pressure). The big mystery was: What happens in the middle? Is there a distinct "middle group" that forms a new type of party, or is it just a messy transition zone?

This paper acts like a high-tech security camera with a smart AI, zooming in on the exact moment the crowd shifts. Here is what they found, explained simply:

1. The "Smart Glasses" (The AI Tool)

To see the tiny details of how water molecules arrange themselves, the researchers built a special pair of "smart glasses" using Neural Networks (a type of AI).

• The Trick: Previous tools mostly looked at where the big "Oxygen" atoms were standing. This new tool looks at both the Oxygen atoms and the smaller "Hydrogen" atoms (which act like the hands holding onto each other).
• The Discovery: The researchers found that looking at the "hands" (Hydrogen bonds) is crucial. It's like trying to understand a dance by only watching the dancers' feet; you miss the direction they are facing. By watching the hands, the AI could perfectly tell the difference between the relaxed crowd (LDA), the squeezed crowd (HDA), and the messy middle.

2. The "Border Guard" (The Interface)

The biggest surprise was about the boundary between the two crowds.

• Old Idea: Scientists thought there might be a whole new "Middle Crowd" (called MDA or Medium-Density Amorphous ice) that forms a separate layer or a new phase of matter in the middle of the room.
• New Reality: The paper shows that this "Middle Crowd" does not exist as a separate group. Instead, it only appears right at the border where the relaxed crowd meets the squeezed crowd.
• The Analogy: Imagine a wall separating a quiet library (LDA) from a loud concert (HDA). The "Middle Crowd" isn't a third room; it's just the people standing right against the wall, trying to be quiet but also getting ready to dance. They are the transition zone, not a new place.

3. The "Elastic Rubber Band" (How the Border Moves)

The researchers watched what happened when they squeezed the room (increased pressure) and then let go (decreased pressure).

• The Shift: When they squeezed, the border between the two crowds moved slightly toward the relaxed side, turning a few more people into the "squeezed" group.
• The Memory Effect: When they let go, the border didn't immediately snap back to its original spot. It stayed slightly shifted, like a rubber band that has been stretched and needs a little extra slack to return to its exact starting position. This is called hysteresis (or a "memory effect"). The border remembers it was squeezed.
• The Thickness: Interestingly, no matter how hard they squeezed, the width of that border zone stayed exactly the same (about 3 to 4 molecules thick). It didn't get wider or fuzzier; it just slid back and forth.

4. The "Shape-Shifter" (What is MDA?)

The paper confirms that the "Medium-Density" ice (MDA) that scientists recently discovered is real, but it's not a new, permanent type of ice.

• The Verdict: MDA is just the name we give to the molecules standing at the border. It's a "shape-shifter" that looks a bit like the relaxed crowd and a bit like the squeezed crowd, depending on where it stands in the transition zone. It is not a distinct, stable phase like the other two.

Summary

Think of the transformation of ice under pressure like a marching band changing formation.

• They don't stop to form a new, separate group in the middle.
• Instead, the front row (the interface) shifts forward.
• The people in the front row are the "middle" ones, holding hands in a way that is different from the back row and the front row.
• If you push them, the whole line moves, but the "front row" stays the same width. If you pull them back, they don't instantly return to where they started; they lag a little bit behind.

The paper proves that the "middle" of this transformation is just a thin, moving border, not a new world of its own.

Technical Summary

Technical Summary: Flat Interface between Amorphous Ices and the Role of MDA-like Intermediate States in the LDA–HDA Transformation

Problem Statement
The pressure-induced transformation between low-density amorphous ice (LDA) and high-density amorphous ice (HDA) is a prototypical polyamorphic transition. While extensively studied over the past 40 years, the microscopic mechanism of this transition remains debated, particularly regarding the nature of intermediate amorphous (IA) states. Previous research has largely focused on bulk properties or macroscopic kinetics, leaving the structure and role of the interface between coexisting amorphous phases largely unexplored. The recent discovery of medium-density amorphous ice (MDA) has further complicated the landscape, with ongoing debate as to whether MDA represents a distinct bulk phase, a shear-induced state, or an intermediate configuration within a continuum of amorphous structures. A critical methodological challenge in resolving these questions is the lack of a robust descriptor capable of distinguishing subtle structural differences between LDA, HDA, and MDA, particularly regarding the orientational order of the hydrogen bond network.

Methodology
The authors employ a combined approach of molecular dynamics (MD) simulations and machine learning (ML) to characterize the LDA–HDA transformation and the resulting interface.

• Simulations: MD simulations were performed using the TIP4P/Ice potential within the LAMMPS package. Systems containing 2,880 water molecules were prepared in LDA, HDA, and MDA configurations. Simulations included isothermal compression (0.1 GPa/ns at 77 K) and the construction of artificial flat LDA || HDA interfaces (15,339 molecules) to study equilibrium properties and response to pressure variations.
• Descriptors: To overcome the limitations of traditional low-dimensional order parameters, the study utilizes Smooth Overlap of Atomic Positions (SOAP) descriptors. Two versions were constructed: one considering only oxygen atoms (M3(O)) and a more comprehensive version including both oxygen and hydrogen atoms (M3(O,H)). The SOAP descriptors capture radial and angular correlations, including hydrogen bond orientations.
• Machine Learning: Neural network classifiers (M2 for binary LDA/HDA, M3 for ternary LDA/HDA/MDA) were trained on SOAP descriptors. The models output prediction probabilities for phase membership. A confidence threshold ($P_{threshold}$) was applied to classify molecules as LDA, HDA, or Intermediate Amorphous (IA) if the probability fell below the threshold.
• Analysis: Feature importance was quantified using permutation importance, and Principal Component Analysis (PCA) was used to map structural fingerprints. The spatial distribution of IA molecules was analyzed relative to growing HDA domains, and the properties of the flat interface (thickness, composition, and hysteresis) were characterized.

Key Contributions and Results

1. Importance of Hydrogen Bond Orientation:
   Feature importance analysis reveals that while oxygen-only models rely on local density and tetrahedral distortion, models including hydrogen atoms (M3(O,H)) identify hydrogen-bond orientation as the primary discriminator between LDA and HDA. Specifically, OH-type features with high angular moments (probing medium-range order, 4–6 Å) dominate the classification. Remarkably, a model trained only on hydrogen-related features (O-H and H-H correlations) achieves classification accuracy (90.92%) comparable to the full model, confirming that the hydrogen network encodes the critical structural differences between amorphous phases.

2. Nature of Intermediate Amorphous (IA) States:
   During the pressure-induced LDA $\to$ HDA transformation, molecules with low classification confidence (IA states) are not randomly distributed nor do they form a distinct bulk phase. Instead, they localize exclusively at the LDA–HDA interface.

   • Spatial Localization: IA molecules form a connected layer between LDA and HDA domains. Their population peaks when the interfacial area is maximal (during the growth of HDA clusters) and decreases as the system becomes fully HDA.
   • MDA-like Character: PCA projection of the interfacial layer environments onto the descriptor space of the ternary classifier (trained on LDA, HDA, and MDA) shows that the interface projects directly onto the region occupied by bulk MDA. This indicates that the interfacial layer is structurally similar to MDA, suggesting MDA is not a separate bulk phase but rather a configuration characteristic of the interface between LDA and HDA.

3. Characterization of the Flat Interface:
   The study characterizes a flat LDA || HDA interface at equilibrium (0.2 GPa, 77 K).

   • Thickness: The interface has a finite thickness of approximately 11–12 Å (3–4 molecular layers), calculated via the Gibbs method. This thickness remains constant under compression and decompression.
   • Structure: The transition from LDA to HDA across the interface is continuous. The M2(O,H) model reveals long tails in the probability profile, indicating that the hydrogen-bond network senses the phase boundary over a longer range than the oxygen sublattice, suggesting a two-stage relaxation process.
   • Elastic Response and Hysteresis: Under compression, the interface shifts reversibly toward the LDA region (by ~1–2 Å) without changing thickness. Upon decompression, the interface exhibits moderate kinetic hysteresis; it does not fully return to its original position until the pressure drops below the initial equilibrium pressure (0 GPa), indicating a memory effect.

4. Transformation Mechanism:
   The LDA $\to$ HDA transformation proceeds via a nucleation and growth mechanism. Critical HDA nuclei (5–10 molecules) form at low pressures. As pressure increases (0.6–0.7 GPa), these clusters grow, engulfing the LDA matrix. The growth is facilitated by the interfacial layer of IA molecules, which act as precursors. The interface thickness remains constant during growth, implying advancement occurs via interfacial migration rather than expansion.

Significance
The paper claims to resolve the long-standing debate regarding the nature of intermediate states in the LDA–HDA transition. By demonstrating that IA states (and by extension, MDA-like structures) localize at the phase boundary rather than forming a bulk intermediate phase, the study supports a continuum concept where the transition occurs through a spatially continuous evolution of local structure across a finite interface.

The work highlights the necessity of including hydrogen atoms in structural descriptors to accurately capture the physics of amorphous ices, as oxygen-only descriptions miss critical orientational information. Furthermore, the characterization of the interface's elastic response and kinetic hysteresis provides a new perspective on the thermodynamics and kinetics of polyamorphic transitions in disordered systems. The methodology offers a general framework for analyzing local structure in other polyamorphic materials where interface-mediated transitions may occur.