Neutron and X-ray Diffraction Reveal the Limits of Long-Range Machine Learning Potentials for Medium-Range Order in Silica Glass

By combining neutron and X-ray diffraction with large-scale molecular dynamics simulations, this study demonstrates that while explicit long-range interactions improve liquid structure predictions, they remain insufficient for accurately modeling the medium-range order of silica glass, as both short-range and long-range machine learning potentials fail to capture the necessary network flexibility and ring statistics during the liquid-to-glass transition.

The Gist

The Big Picture: Building a Glass House with AI

Imagine you are trying to build a perfect, invisible house out of tiny Lego bricks (atoms) to make Silica Glass (the stuff in windows and fiber optics).

Scientists have been using Machine Learning (ML) to teach computers how to build this house. The goal is to get the computer to arrange the bricks exactly like nature does.

The problem? The computer is great at building the immediate neighborhood of each brick (the short-range order), but it keeps messing up the layout of the whole neighborhood (the medium-range order). This layout is crucial because it determines how light passes through the glass and how strong it is.

This paper asks a simple question: If we give the computer "long-distance vision" (so it can see bricks far away), will it finally build the perfect glass house?

The Experiment: Two Architects

The researchers set up a test with two different "AI Architects" to see how they build the glass:

1. Architect A (The Short-Range Model): This AI only looks at the 5 bricks immediately next to it. It has a very narrow field of view.
2. Architect B (The Long-Range Model): This AI has a super-power. It can see and feel the influence of bricks much further away, like having a long-range radio or a telescope.

Both architects were trained on the same "blueprint" (data from quantum physics) and asked to build the glass in two stages:

• Stage 1: The Liquid Soup. First, they build the glass while it's still hot and melting (like a pot of soup).
• Stage 2: The Glass. Then, they cool it down rapidly (quenching) to turn it into solid glass.

What They Found

1. The "Liquid Soup" Phase

When the glass was still hot and liquid:

• Architect A (Short-Range) got too excited. It started arranging the bricks in a very rigid, overly organized pattern. It was like a crowd of people in a hot room suddenly holding hands in a perfect circle, even though they should be moving freely.
• Architect B (Long-Range) did much better. By seeing the "big picture," it realized the bricks shouldn't be so rigid. It loosened up the structure, making it look much more like the real liquid silica scientists see in experiments.

The Lesson: Giving the AI "long-range vision" helps it understand how the liquid flows.

2. The "Solid Glass" Phase (The Big Surprise)

Here is where things got tricky. They cooled the liquid down to make solid glass.

• Architect A built a glass that was too perfect and rigid. It had too many "six-sided rings" (like a honeycomb), which made the structure too stiff and unnatural.
• Architect B did a better job than A, but it still failed to match reality. Even though it could see far away, the glass it built was still slightly wrong. It didn't have the right "wiggles" and "bumps" that real glass has.

The Analogy: Imagine trying to learn how to dance by watching a video.

• Architect A only watches your feet. You learn the steps but look stiff.
• Architect B watches your whole body and the music. You move better, but you still don't have the soul of the dance. You are still dancing like a robot, not a human.

Why Did Architect B Fail?

The researchers realized that the problem wasn't just about how far the AI could see. The problem was how the AI learned to dance.

When you cool glass down (quenching), the atoms get "frozen" in place very quickly. If the AI was trained mostly on the "perfect" liquid state, it gets stuck in that memory. When it tries to freeze the glass, it freezes it in a way that is kinetically trapped.

Think of it like this:

• The AI learned to build the liquid perfectly.
• But when it tried to turn that liquid into solid glass, it didn't know how to handle the "panic" of the atoms trying to freeze.
• It kept the "memory" of the liquid structure, which was slightly wrong for the solid state.

The Conclusion: What's Next?

The paper concludes that giving AI "long-range vision" is necessary, but not enough.

To build the perfect glass, we need two things:

1. Long-range vision: So the AI understands how atoms talk to each other over distances.
2. Better Training: We need to teach the AI specifically about the chaos of the transition from liquid to solid. We need to show it more examples of the "freezing" process, not just the "melting" process.

In short: You can't just give a student a better telescope (long-range interactions) and expect them to ace the exam. You also have to teach them how to handle the specific test conditions (the liquid-to-glass transition). Without both, the glass they build will always be a little bit "off."

Technical Summary

1. Problem Statement

Glassy silica is a foundational material in optics and electronics, yet accurately predicting its Medium-Range Order (MRO) remains a significant challenge for Machine Learning Interatomic Potentials (MLIPs).

• The Gap: While existing MLIPs successfully reproduce short-range order (the local SiO₄ tetrahedral network), they often fail to accurately predict the First Sharp Diffraction Peak (FSDP). The FSDP is the primary experimental signature of MRO, arising from correlations over 5–20 Å (specifically the periodicity of adjacent cages and ring structures).
• The Hypothesis: It is unclear whether the failure to capture MRO is due to the inherent locality of most MLIPs (finite cutoff radii) which miss long-range electrostatics and extended network correlations, or if other factors (such as training data or sampling strategies) are the limiting factors.

2. Methodology

The authors employed a comparative approach combining experimental validation with large-scale molecular dynamics (MD) simulations using two distinct models based on the MACE (Message Passing Atomic Cluster Expansion) framework.

• Experimental Validation:

  • Techniques: High-energy X-ray diffraction (at APS) and Neutron diffraction (at Spallation Neutron Source).
  • Samples: Both liquid silica (at high temperatures) and amorphous silica (quenched glass).
  • Metrics: Structure factors $S(Q)$, bond-angle distributions (O-Si-O and Si-O-Si), and ring size statistics.

• Computational Models:

  • Short-Range (SR) Model: A standard MACE model with a finite cutoff radius ($r_{cut} = 5$ Å). It captures local many-body interactions but excludes explicit long-range physics.
  • Long-Range (LR) Model: An extension of the SR backbone augmented with Reciprocal-Space Gated Attention (RSGA). This incorporates long-range electrostatics and correlations via a reciprocal-space channel using Ewald-inspired weights and gated linear attention.
  • Training Data: Both models were trained on the same Density Functional Theory (DFT) dataset (SCAN functional) containing crystalline, liquid, and amorphous silica configurations. This ensures that differences in results are attributed solely to the interaction range, not training data bias.

• Simulation Protocol:

  • System: ~3,000 atoms in a periodic box.
  • Protocol: A multi-stage melt-quench process. The system was equilibrated at 4000 K (liquid), cooled to 2273 K, and then quenched to 300 K at a rate of 1 K/ps ($10^{12}$ K/s).
  • Analysis: Structural descriptors included bond-angle distributions, Guttman ring statistics, and Persistent Homology (PH) to analyze topological loop structures and their persistence (lifetime).

3. Key Results

A. Liquid State Analysis

• Local Structure: Both SR and LR models accurately reproduced the local tetrahedral geometry (O-Si-O angles 107°) and the broad Si-O-Si angle distribution (138°).
• Medium-Range Order (FSDP):
  • SR Model: Systematically over-structured the network, producing an FSDP intensity significantly higher than experimental data.
  • LR Model: Explicit long-range interactions reduced the excess ordering, shifting the liquid structure factor closer to experimental values. This confirms that long-range physics is essential for describing the liquid state correctly.

B. Amorphous (Glass) State Analysis

• Failure to Recover MRO: Despite improvements in the liquid state, neither model fully recovered the experimental MRO after quenching.
  • The SR model produced an artificially intense FSDP, indicating an overly constrained network.
  • The LR model reduced the FSDP intensity but still underestimated it compared to experiment. The experimental FSDP intensity lies between the SR and LR predictions.
• Ring Statistics:
  • SR Model: Exhibited a narrow ring-size distribution dominated almost exclusively by six-membered rings, indicating a rigid, over-ordered topology.
  • LR Model: Showed a broader distribution with more small ($\leq 5$) and large ($\geq 7$) rings, reflecting increased topological diversity. However, this diversity was still insufficient to match the experimental balance.
• Bond Angles: Both models showed constrained Si-O-Si angle distributions (mean 140°) compared to the experimental value (146°), suggesting limited network flexibility.
• Persistent Homology (PH):
  • PH analysis revealed that the SR model produced distorted, irregular high-persistence loops with significant geometric strain, indicative of kinetic trapping (non-equilibrium states).
  • The LR model showed more relaxed, physically plausible loop structures but still retained "structural memory" of the parent liquid that prevented full equilibration during vitrification.

4. Key Contributions

1. Isolation of Interaction Range: By using two models with identical short-range backbones and training data, the study definitively isolates the effect of long-range interactions. It proves that while long-range interactions are necessary to correct liquid structure, they are not sufficient to predict the final amorphous MRO.
2. Identification of Kinetic Trapping: The study demonstrates that the failure to reproduce MRO is partly due to the models retaining excessive "memory" of the parent liquid network, leading to kinetically trapped, nonphysical configurations during the quench.
3. Topological Analysis: The application of Persistent Homology provided new insights into the geometric strain and topological rigidity of MLIP-generated glass structures, distinguishing between physically meaningful motifs and artifacts of rapid quenching.

5. Significance and Conclusion

The paper concludes that accurate predictive modeling of disordered silica requires a dual approach:

1. Physical Correctness: Explicit inclusion of long-range interactions is mandatory to capture intermediate-range correlations.
2. Sampling and Data Strategy: Interaction range alone cannot fix the problem. The training data and sampling strategies must adequately represent the liquid-to-glass transition. Current datasets may lack sufficient diversity in the structural rearrangements that occur during vitrification, leading models to get stuck in local energy minima that do not reflect the true experimental glass structure.

Implication: Future MLIP development for glasses must move beyond simply extending interaction cutoffs; it requires active learning strategies that specifically target the complex configurational pathways of the vitrification process to break the "structural memory" of the liquid state.