On the Boroxol Ring Fraction in Melt-Quenched B$_2$O$_3$ Glass

This study develops a DFT-accurate machine-learned potential and utilizes deep potential molecular dynamics with slow quench rates and extended geometry descriptors to successfully generate B$_2$O$_3$ glass models containing over 30% boroxol rings, revealing that the energy-minimized boroxol fraction of 75% closely matches experimental estimates.

The Gist

The Big Picture: The "Missing Puzzle Piece" in Glass

Imagine you are trying to build a perfect model of a specific type of glass called Boron Trioxide (B2O3). Scientists have known for a long time what this glass looks like under a microscope (or rather, through spectroscopy): it's made of tiny, flat, six-sided rings (like a honeycomb cell) made of Boron and Oxygen atoms. These are called Boroxol rings.

Experiments show that in real-world glass, about 75% of the atoms are part of these rings.

However, for decades, computer simulations have failed to build this glass correctly. No matter how hard they tried, the computer models only managed to get about 15% to 30% of the atoms into these rings. It was like trying to build a Lego castle where the instructions say you need 75% blue bricks, but your computer model keeps building a castle that is mostly red bricks.

The Problem: Why couldn't the computers get it right?

1. The Recipe was Wrong: The "force fields" (the rules the computer uses to tell atoms how to push and pull each other) were based on old, imperfect data.
2. The Cooling was Too Fast: In the real world, glass cools down slowly over hours or days. In computer simulations, to save time, they cool the "melt" (liquid glass) down in a split second. This is like trying to freeze a pond instantly; the water turns to ice before the molecules have time to arrange themselves into a perfect pattern.
3. The "Zoom" was Too Short: The computer models were only looking at atoms very close to each other (like looking at a single brick) and ignoring the bigger picture (the whole wall).

The Solution: A Smarter Brain and a Slower Freeze

The authors of this paper, Debendra Meher and his team, fixed these problems using three main tricks:

1. Teaching the Computer with a "Super-Teacher" (Machine Learning)

Instead of guessing the rules for how atoms interact, they trained a Machine Learning (ML) model. Think of this like training a dog.

• Old Way: You tell the dog, "Sit," and hope it understands.
• New Way: You show the dog thousands of pictures of people sitting, standing, and running, and you correct it every time it makes a mistake.
• The Result: They fed their AI model thousands of high-quality examples (calculated using very expensive, precise quantum physics) so the AI learned the exact rules of how Boron and Oxygen behave. This new AI model is called ML-31.

2. Slowing Down the Freeze (The Quench Rate)

They realized that cooling the glass too fast was the main culprit.

• The Analogy: Imagine a crowded dance floor. If you suddenly turn off the music and freeze everyone in place (fast cooling), people will be in awkward, random poses. If you slowly dim the lights and let the music fade out (slow cooling), people have time to find their partners and form nice, organized circles.
• The Action: They slowed the computer cooling process down to the slowest speed ever attempted for this type of glass. This gave the atoms enough time to find their way into those perfect six-sided rings.

3. Looking Further Away (The 9Å Range)

They discovered that the computer needed to "see" further.

• The Analogy: Imagine trying to understand a city's traffic flow. If you only look at the car right in front of you (a short range), you won't understand why the traffic jam happened three blocks away. You need to look at the whole neighborhood.
• The Action: They adjusted the AI to look at atoms up to 9 Angstroms away (about 9 times the width of an atom). They found that if the AI only looked at 6 Angstroms, it got the pressure wrong and couldn't form the rings correctly.

The Results: Getting Closer to Reality

With these improvements, the team finally got a computer model that looked like real glass:

• The Success: They managed to get 30% of the atoms into boroxol rings. This is a huge jump from the previous 15%, and it's the highest anyone has ever achieved in a simulation.
• The "Energy Sweet Spot": They also discovered something fascinating. They built many different versions of the glass with different amounts of rings (10%, 20%, 50%, 75%, etc.) and measured their energy.
  • They found that the glass is happiest (lowest energy) when it has about 75% rings.
  • This matches the real-world experiments perfectly! It suggests that nature "wants" the glass to have 75% rings because that is the most stable, comfortable state for the atoms.

Why This Matters

This paper is a breakthrough because it proves that if we give the computer the right tools (a smart AI) and enough time (slow cooling), it can finally simulate complex glass structures accurately.

Previously, scientists thought the problem was that glass was too messy to model. This paper says, "No, we just weren't looking closely enough or waiting long enough."

The Takeaway:
Think of this like baking a cake. For years, bakers (scientists) were trying to bake a Boron Trioxide cake, but it always came out flat and ugly. This team realized they were using the wrong flour (bad rules), baking it at the wrong temperature (cooling too fast), and not mixing it long enough. By fixing the recipe and the process, they finally baked a cake that looks and tastes exactly like the real thing.

This opens the door to designing new, stronger, or more heat-resistant glasses in the future, because now we can trust our computer models to tell us what will happen before we even build them in the lab.

Technical Summary

1. Problem Statement

The primary challenge in modeling Boron Trioxide ($B_2O_3$) glass via atomistic simulations is the accurate reproduction of the boroxol ring fraction.

• Experimental Reality: Spectroscopic evidence (Raman and $^{11}B$ NMR) indicates that approximately 75% of boron atoms in ambient-condition $B_2O_3$ glass exist within six-membered boroxol rings ($B_3O_6$).
• Simulation Failure: Previous Molecular Dynamics (MD) simulations, even those using ab initio methods or empirical force fields, have consistently failed to reproduce this high fraction. Typical simulations yield boroxol fractions of only 15–30%.
• Root Causes of Failure:
  1. Interatomic Potentials: Empirical potentials often lack the fidelity to capture specific intermediate-range order (IRO) motifs like boroxol rings.
  2. Quench Rates: Simulations typically use quench rates ($\sim 10^{10} - 10^{12}$ K/s) that are orders of magnitude faster than experimental rates, preventing the system from equilibrating into the thermodynamically preferred boroxol-rich state.
  3. Density Modeling: Previous models often quenched the melt at the glass density (1.834 g/cc) rather than the melt density (1.49 g/cc at 2000 K), artificially suppressing ring formation.
  4. Descriptor Range: Machine Learning Potentials (MLPs) often use short-range descriptors (e.g., 6 Å) that fail to capture the long-range structural correlations necessary for accurate pressure and ring statistics.

2. Methodology

The authors developed a high-fidelity simulation framework combining advanced Machine Learning Potentials (MLPs) with optimized quenching protocols.

• Machine Learning Potentials (MLPs):

  • Model: Developed ML-31, a DFT-accurate MLP based on the Deep Potential (DP) and MACE architectures.
  • Training Data: Unlike previous models (ML-26) trained on high-pressure data, ML-31 was trained on a diverse dataset (8,652 frames) specifically enriched with configurations containing high boroxol fractions and varying densities.
  • Descriptor Range Optimization: A critical investigation revealed that the geometry descriptor cut-off is vital. While 6 Å is standard, the authors found that a cut-off of at least 9 Å is required to accurately capture the pressure and structural motifs of $B_2O_3$. Shorter cut-offs (6–7 Å) led to significant errors in pressure and underestimation of boroxol fractions.
  • Validation: The model was validated against Density Functional Theory (DFT) using the revPBE functional, Grimme's D3 dispersion corrections, and high electron density cut-offs (1200 Ry).

• Quenching Protocols:

  • NVT-$\rho_{EXP}$: Instead of constant-volume quenching at the glass density, the authors implemented a step-wise quenching protocol that follows the experimental density-temperature curve. The system starts at 2000 K (1.49 g/cc) and is cooled to 300 K, with coordinates rescaled at each step to match experimental densities.
  • Ultra-Slow Quenching: Simulations were performed with quench rates as low as $10^9$ K/s, significantly slower than previous studies, to allow for better sampling of the viscous supercooled liquid state.
  • System Sizes: Simulations utilized system sizes of 480 and 1700 atoms to ensure statistical robustness.

• Analysis:

  • Boroxol ring identification was performed using Python/NetworkX with specific coordination criteria (B: 3, O: 2).
  • Vibrational Density of States (VDOS) was calculated via Velocity Auto-Correlation Functions (VACF) to compare with the experimental Raman peak at 808 cm$^{-1}$.
  • Energy minimization was performed on 500 amorphous configurations with varying boroxol fractions to map the energy landscape.

3. Key Results

• Impact of Descriptor Range:

  • The pressure of the melt and the boroxol fraction are highly sensitive to the descriptor cut-off.
  • A 6 Å cut-off yields incorrect pressures and lower boroxol fractions.
  • A 9 Å cut-off is necessary to converge pressure values and accurately model the intermediate-range order, leading to higher boroxol fractions.

• Quench Rate Dependence:

  • The boroxol fraction increases monotonically as the quench rate decreases.
  • At a quench rate of $1 \times 10^9$ K/s (using ML-31/R9), the simulation achieved a 30% boroxol fraction, a significant improvement over previous results (15–19%).
  • Extrapolation suggests that experimental quench rates (much slower) would yield the experimentally observed 75%.

• Density Protocol:

  • Following the experimental density-temperature curve (NVT-$\rho_{EXP}$) yielded higher boroxol fractions compared to constant-density quenching at the glass density.
  • Quenching at the melt density (1.49 g/cc) allows for the necessary structural relaxation to form rings before the system vitrifies.

• Energy Landscape and Stability:

  • The total energy of amorphous $B_2O_3$ decreases as the boroxol fraction increases, reaching a global minimum at approximately 75% boroxol fraction.
  • This energy minimum aligns perfectly with the experimental consensus (75–80%), suggesting that the 75% state is the thermodynamically preferred amorphous structure at ambient density.
  • Configurations with >75% boroxol rings show increased energy, likely due to strain required to maintain the bulk density constraint.

• Formation Mechanism:

  • Boroxol rings form via the reorganization of larger rings through a transient four-coordinated boron ($BO_4$) species.
  • This reorganization occurs on the picosecond timescale, but the nucleation barrier is high, requiring slow cooling to overcome.

• Spectroscopic Validation:

  • The VDOS of the simulated glass with high boroxol content shows a sharp peak at 790 cm$^{-1}$, closely matching the experimental Raman peak at 808 cm$^{-1}$.
  • The intensity of this peak scales linearly with the boroxol fraction, confirming the structural origin of the spectral feature.

4. Significance and Conclusion

• Resolution of a Long-Standing Discrepancy: This work bridges the gap between simulation and experiment for $B_2O_3$ glass, explaining why previous models failed (inadequate descriptor range, incorrect density protocols, and insufficient quench times).
• Methodological Advancement: The study establishes that for network glasses with intermediate-range order, long-range descriptors (9 Å) are non-negotiable in MLPs. It also highlights the critical importance of following experimental density-temperature trajectories in melt-quench simulations.
• Thermodynamic Insight: The discovery of an energy minimum at 75% boroxol fraction provides a thermodynamic justification for the experimental observation, suggesting that $B_2O_3$ glass is an "ideal glass former" that naturally seeks this specific structural arrangement.
• Future Outlook: While the current simulations (limited by computational cost) reached 30%, the trend strongly implies that the 75% experimental value is achievable with even slower quench rates or enhanced sampling techniques (e.g., swap Monte Carlo, diffusion models). The developed ML-31 potential serves as a robust tool for future investigations into ultrastable glasses and the crystallization anomaly of $B_2O_3$.

In summary, the paper demonstrates that by combining DFT-accurate MLPs with extended descriptor ranges and experimentally faithful quenching protocols, it is possible to model the complex intermediate-range order of $B_2O_3$ glass, successfully capturing the formation of boroxol rings and their thermodynamic stability.