Melting Behavior and Phase Stability of CaO from Neural Network Potentials: a Molecular Dynamics Study

This study employs a machine learning interatomic potential to perform large-scale molecular dynamics simulations that determine the melting temperature, enthalpy of fusion, and high-pressure melting curve of calcium oxide, revealing a pressure-dependent overheating ratio and establishing MLIPs as a robust framework for investigating ionic oxide phase stability.

The Gist

Imagine trying to figure out exactly when a block of rock-hard cheese (in this case, Calcium Oxide, or CaO) turns into a gooey liquid. Scientists have been arguing about this temperature for a long time. Some say it's around 2,800 degrees, others say it's over 3,200 degrees. The problem is, CaO is so hot and reactive that it's like trying to melt a piece of metal while it's also trying to eat the container it's sitting in. It's hard to get a clean measurement in a real lab.

To solve this, the researchers in this paper built a digital twin of CaO. Instead of melting real rocks, they created a "smart computer brain" (called a Machine Learning Interatomic Potential) that knows exactly how every single atom in CaO behaves. Think of this brain as a super-accurate rulebook that tells the computer how atoms push and pull on each other, but it runs a million times faster than the old, slow physics simulations used before.

Here is how they used this digital brain to find the answers:

1. The Two Ways to Melt a Digital Rock

To find the exact melting point, they tried two different "games" in their simulation:

• The "Hole in the Wall" Method (Void-Nucleated Melting):
  Imagine a perfect brick wall. If you heat it up, it might stay solid way past its melting point because there are no cracks to start the collapse. To fix this, the researchers punched a hole in the middle of their digital wall. This hole acts like a weak spot. As they heated the wall, the liquid started forming right around the hole. By making the hole bigger and bigger, they found the temperature where the wall always collapses.

  • The Result: They found the melting point to be 3,055 Kelvin (about 2,782°C). This matched the best recent real-world experiments.

• The "Half-and-Half" Method (Two-Phase Coexistence):
  Imagine a long train car where the front half is frozen ice and the back half is boiling water. They put this train car in the simulation and watched the boundary between the ice and water. If the ice melts, the whole thing is too hot. If the water freezes, it's too cold. They adjusted the temperature until the ice and water stayed perfectly balanced.

  • The Result: This method gave a lower number, 2,847 Kelvin. The paper notes this method is known to sometimes underestimate the temperature, but it's still a useful check.

2. Checking the "Heat Bill" (Enthalpy)

Melting isn't just about temperature; it's also about how much energy you have to pour into the system to break the solid structure. The researchers calculated this "energy bill" (Enthalpy of Fusion).

• They found that their digital brain predicted an energy cost of about 73 kJ/mol.
• This number lined up perfectly with the best estimates from real-world chemistry tables and other high-level physics calculations. It proved their digital brain was telling the truth.

3. The "Squeeze" Test (High Pressure)

Finally, they asked: "What happens if we crush this rock?" They squeezed their digital CaO up to 20 Gigapascals (that's like the pressure at the bottom of the ocean, but multiplied by a thousand).

• The Old Assumption: Scientists used to think that as you squeeze a material, the "overheating" (the extra heat needed to melt a perfect crystal) stays the same percentage.
• The New Discovery: The researchers found this assumption was wrong. As they squeezed the CaO harder, the "overheating" gap actually grew. At normal pressure, a perfect crystal needed about 17% extra heat to melt. At high pressure, it needed 24% extra heat.
• Why? Think of it like a crowded dance floor. When the room is empty (low pressure), it's easy for a few dancers to start moving (melting). But when the room is packed tight (high pressure), it takes a massive amount of energy to get the crowd to break formation and start dancing, especially if there are no "weak spots" (defects) to help them start.

The Bottom Line

This paper didn't just guess the melting point of Calcium Oxide; they built a highly accurate, fast computer model to prove it. They confirmed that CaO melts around 3,055 K at normal pressure and showed that the rules for how it melts change when you squeeze it. Their new "digital brain" is now a reliable tool for scientists to study other extreme materials without needing to melt them in a real lab.

Technical Summary

Technical Summary: Melting Behavior and Phase Stability of CaO from Neural Network Potentials

Problem Statement
Determining the precise melting temperature ($T_m$) and high-pressure phase stability of calcium oxide (CaO) remains a significant challenge due to experimental limitations, including high vapor pressure and reactivity. Experimental values for $T_m$ vary widely (2843 K to 3223 K), and the high-pressure melting curve is largely unexplored. Previous computational studies faced a trade-off: empirical potentials (e.g., Born-Mayer-Huggins) lacked the necessary quantum accuracy and transferability for extreme conditions, while ab initio molecular dynamics (AIMD) offered precision but at a computational cost prohibitive for the large system sizes and long simulation times required to accurately map melting curves and avoid superheating artifacts.

Methodology
To bridge this gap, the authors developed a Machine Learning Interatomic Potential (MLIP) for CaO using the PANNA 2.0 framework and the LATTE descriptor.

• Potential Architecture: The model utilizes a multilayer perceptron (MLP) with a 256:124:1 hidden layer structure. It employs the LATTE descriptor, based on Cartesian tensor contractions, to encode local atomic environments invariant to translation, rotation, and permutation. The descriptor includes 2-body, 3-body, and 4-body terms, totaling 900 features.
• Training Dataset: The potential was trained on $\sim$12,000 configurations derived from AIMD simulations (using PBEsol functional and Quantum ESPRESSO). The dataset is diverse, comprising:
  • Solid and liquid bulk structures (300–6000 K).
  • Solid-liquid interface structures (for two-phase coexistence).
  • "Voided" solid structures with central holes (1–40% volume) for void-nucleated melting.
  • Compressed and expanded structures (up to 20 GPa).
  • The data was split into training (75%), validation (15%), and test (10%) sets.
• Training Performance: The final potential achieved a Mean Absolute Error (MAE) of $<5$ meV/atom for energy and $<45$ meV/Å for forces.
• Simulation Techniques: Large-scale Molecular Dynamics (MD) simulations were performed using LAMMPS with the trained MLIP in the isobaric (NPT) ensemble. Two methods were employed to determine $T_m$:
  1. Void-Nucleated Melting (VNM): Introducing a central void in a 11$\times$11$\times$11 supercell to induce heterogeneous nucleation and eliminate superheating.
  2. Two-Phase Coexistence (TPC): Simulating a system with coexisting solid and liquid phases in a 6$\times$6$\times$60 supercell to find the equilibrium temperature.

Key Results

• Melting Temperature at Ambient Pressure:
  • The VNM method yielded $T_m = 3055 \pm 11$ K. This value aligns closely with recent experimental data and previous MLIP studies, suggesting it is the most accurate estimate among the methods tested.
  • The TPC method yielded $T_m = 2847 \pm 15$ K. The authors note this underestimation is a known limitation of the TPC method for CaO and other materials.
• Thermodynamic Properties:
  • Enthalpy of Fusion ($\Delta H_f$): Calculated as $\sim$73 kJ/mol (at 3055 K) and $\sim$71.6 kJ/mol (at 2847 K). These values agree with thermodynamic assessments and AIMD calculations.
  • Entropy of Fusion ($\Delta S_f$): Calculated values ($\sim$24.05–25.14 J·mol$^{-1}$·K$^{-1}$) are consistent with NIST-JANAF tables and estimates derived from MgO and BeO.
  • Volume Expansion: The simulation predicts a volume increase of $\sim$29% at $T_m$, consistent with density profiles showing the liquid phase is less dense than the solid phase.
  • Thermal Expansion: The MLIP reproduces thermal expansion trends consistent with AIMD and experimental data, correcting the steeper slope observed in previous empirical BMH potential simulations.
• High-Pressure Melting Curve (up to 20 GPa):
  • Using the TPC method, the authors calculated the melting curve up to 20 GPa.
  • The study confirms that the overheating ratio ($\eta = T_s/T_m - 1$) is not constant under pressure. It increases from 17.3% at 0 GPa to 24.0% at 20 GPa.
  • Empirical fits ($T = aP^b + c$) show that while $T_m$ and the thermal instability temperature ($T_s$) follow similar scaling laws, $T_s$ increases more steeply with pressure due to the lack of nucleation sites in perfect crystals under high pressure.

Significance and Claims
The paper establishes MLIP-based simulations as a robust and efficient framework for investigating phase stability in ionic oxides. By achieving ab initio accuracy at the computational cost of empirical potentials, the study resolves thermodynamic inconsistencies found in previous empirical models (specifically regarding thermal expansion slopes and overheating ratios). The results provide one of the few computational determinations of the CaO high-pressure melting curve up to 20 GPa. Crucially, the work validates that assuming a fixed overheating ratio for estimating $T_m$ under pressure is inaccurate, as the ratio between the thermal instability limit and the true melting point varies significantly with pressure. The study confirms a melting temperature for CaO above 3000 K at ambient pressure.