Long-range interaction effects on the phase transition, mechanical effect, and electric field response of BaTiO3 by machine learning potentials

This study demonstrates that while a long-range MACELES model significantly improves the quantitative accuracy of predicting transition temperatures, elastic constants, and dielectric constants for BaTiO3 compared to a local-only model, both approaches successfully reproduce the material's key qualitative ferroelectric behaviors such as phase transitions and polarization switching.

The Gist

Imagine you are trying to predict how a complex Lego castle (a crystal of Barium Titanate, or BaTiO₃) will behave when you heat it up, squeeze it, or run an electric current through it.

In the world of computer science, scientists use "Machine Learning Potentials" (MLPs) as a super-fast calculator to simulate this behavior. Think of these ML models as super-smart students who have memorized the rules of how Lego bricks snap together.

The Problem: The "Local" Student vs. The "Global" Student

For a long time, these "students" (ML models) had a blind spot. They were trained to only look at the immediate neighbors of a Lego brick.

• The Short-Range Model (MACE): Imagine a student who only looks at the 5 bricks touching a specific brick. They are great at seeing how the local structure holds together, but they miss the fact that the whole castle is a giant, charged magnet. They ignore the "long-distance whispers" (long-range electrostatic forces) that travel across the entire structure.
• The Long-Range Model (MACELES): This is the upgraded student. They still look at the neighbors, but they also have a "sixth sense" (Latent Ewald Summation) that lets them feel the electric pull and push from bricks far away, even across the room.

The big question the authors asked was: Does ignoring those "long-distance whispers" actually matter? Or is the local view good enough?

The Experiment: Putting the Models to the Test

The researchers put both models through a series of stress tests on a BaTiO₃ crystal to see how they differed.

1. The Vibration Test (Phonons)

• The Analogy: Imagine ringing a bell. The sound waves travel through the metal. In a charged crystal, the "sound" (vibrations) interacts with the electric field, creating a split in the sound frequencies (called LO-TO splitting).
• The Result: The Short-Range student couldn't hear this split; they thought the bell sounded like a dull thud. The Long-Range student heard the split perfectly, just like a real bell.
• Takeaway: If you want to know exactly how the material vibrates, you need the long-range model.

2. The Temperature Test (Phase Transitions)

• The Analogy: Imagine heating a block of ice. It melts into water, then steam. The "phase transition" is the moment it changes state.
• The Result: Both students agreed on the order of events: The crystal changes shape from a sphere to a cube, then to a box, then melts into chaos. They both got the "story" right.
• The Difference: The Long-Range student predicted the melting would happen at a slightly higher temperature. Why? Because the long-range electric forces make the crystal slightly "fluffier" (larger volume), making it a bit harder to shake apart.
• Takeaway: Both models get the qualitative story (what happens), but the long-range model gets the quantitative details (exactly when it happens) slightly better.

3. The Squeeze Test (Mechanical Stress)

• The Analogy: Imagine squeezing a sponge. At a certain point, the sponge snaps and flips its internal structure.
• The Result: Both models said, "Squeeze it to 120 MPa, and it will flip!" They predicted the exact same force needed to break the symmetry.
• Takeaway: For figuring out how much force is needed to switch the material's state, the simple local model is actually good enough.

4. The Electric Field Test (Hysteresis)

• The Analogy: Imagine a compass needle. You spin a magnet around it, and the needle flips back and forth, tracing a loop.
• The Result: Both models drew the same loop. They both showed the needle flipping and staying put.
• The Difference: The Long-Range model predicted the needle was slightly more "wiggly" (higher dielectric constant) because the crystal was slightly softer.
• Takeaway: The overall shape of the electric response is the same, but the long-range model gives a more precise measurement of how "squishy" the material is.

The Big Conclusion: Topology vs. Curvature

The authors found a beautiful pattern that can be explained with a Mountain Analogy:

• The Landscape (Topology): Imagine a map of mountains and valleys. The "Short-Range" model sees the map perfectly. It knows there is a valley here (a stable phase) and a mountain there (an unstable phase). It knows the path from one valley to another. This is why it gets the "story" right.
• The Slope (Curvature): The "Long-Range" model looks at the same map but measures the steepness of the slopes and the exact depth of the valleys.
  • Because the long-range electric forces are like a gentle wind blowing on the mountains, they don't change where the valleys are (the phases stay the same).
  • But they do change how steep the sides are and how deep the bottom is. This affects the temperature at which you slide down, the stiffness of the ground, and the electricity the ground holds.

Why This Matters

This paper tells us that we don't always need the most complex, expensive model.

• If you just want to know what will happen (e.g., "Will this material switch polarization?"), the simple, fast Short-Range model is perfect.
• If you need to know exactly how much (e.g., "What is the exact melting point?" or "How stiff is it?"), you must use the Long-Range model to get the numbers right.

In short: Short-range models tell you the plot of the movie; Long-range models give you the high-definition special effects and the precise budget.

Technical Summary

1. Problem Statement

Machine Learning Potentials (MLPs) have revolutionized atomic-scale molecular dynamics (MD) by offering near-Density Functional Theory (DFT) accuracy at a fraction of the computational cost. However, most standard MLP frameworks (such as the widely used MACE model) rely on local atomic environment descriptors with a finite cutoff radius. This locality assumption inherently neglects long-range electrostatic interactions, which are rigorously treated in DFT via Ewald summation.

In ferroelectric materials like Barium Titanate ($BaTiO_3$), long-range dipole-dipole interactions and macroscopic electric fields are critical. While previous studies suggested that neglecting these interactions might not drastically alter qualitative phase behavior in $BaTiO_3$, the systematic energetic errors and their impact on quantitative properties (e.g., transition temperatures, elastic constants, dielectric response) remained unexplored. The authors aim to determine whether the omission of long-range interactions in MLPs leads to significant inaccuracies in predicting material properties.

2. Methodology

Model Development: MACELES

The authors developed a new potential, MACELES, by integrating the Latent Ewald Summation (LES) framework into the MACE architecture.

• Mechanism: Unlike traditional models that require predefined atomic charges, LES infers latent atomic charges directly from structural information. These charges are then incorporated into an Ewald summation scheme to capture long-range electrostatics.
• Training Data: The model was trained on the same dataset as a previously reported short-range MACE model: 4,045 $BaTiO_3$ configurations generated via Ab Initio Molecular Dynamics (AIMD) across four structural phases (rhombohedral, orthorhombic, tetragonal, cubic) and a temperature range of 1–4000 K.
• Validation: The MACELES model achieved Mean Absolute Errors (MAE) of 1.01 meV/atom for energy and 11.6 meV/Å for forces, comparable to the short-range MACE model (0.85 meV/atom and 17.5 meV/Å), confirming that the addition of LES did not degrade the fundamental predictive capability.

Comparative Simulations

The study performed a systematic comparison between the short-range MACE and the long-range MACELES models across four key areas:

1. Phonon Dispersion: Calculated using the finite displacement method on supercells (3×3×3 to 6×6×6) to observe Longitudinal Optical–Transverse Optical (LO-TO) splitting.
2. Phase Transition Behavior: NPT ensemble MD simulations heating from 1 to 350 K to determine transition sequences and temperatures.
3. Mechanical Response: Calculation of elastic constants and stress-induced polarization switching (coercive stress).
4. Ferroelectric Response: P-E hysteresis loop simulations using a triangular electric field, coupled with an Equivar model to calculate Born Effective Charges (BEC) dynamically.

3. Key Contributions

• Development of MACELES: Successfully implemented a data-driven, long-range electrostatic model for $BaTiO_3$ without requiring predefined charges, bridging the gap between local MLPs and global electrostatic physics.
• Systematic Benchmarking: Provided the first direct, quantitative comparison of short-range vs. long-range MLPs specifically for ferroelectric properties, isolating the effects of electrostatics from other variables.
• Theoretical Insight: Established a clear distinction between the topological structure of the Potential Energy Surface (PES) and its local curvature, explaining why qualitative behaviors remain robust while quantitative metrics are sensitive to long-range interactions.

4. Key Results

A. Phonon Dispersion (Validation of Long-Range Physics)

• LO-TO Splitting: The short-range MACE model failed to reproduce the LO-TO splitting near the $\Gamma$ point, regardless of supercell size. In contrast, MACELES successfully captured the gradual shift of the LO branch toward the $\Gamma$ point as supercell size increased, converging toward DFT results with non-analytic corrections (NAC).
• Conclusion: MACELES correctly captures the long-range dipole-dipole interactions essential for polar materials, evidenced by the emergence of LO-TO splitting.

B. Phase Transition Behavior

• Qualitative Consistency: Both models reproduced the correct phase transition sequence: Rhombohedral $\to$ Orthorhombic $\to$ Tetragonal $\to$ Cubic (R-O-T-C).
• Quantitative Shift: MACELES predicted slightly higher transition temperatures (R$\to$O: 180 K vs. 150 K; O$\to$T: 245 K vs. 225 K; T$\to$C: 297 K vs. 290 K) compared to MACE.
• Origin: This shift is attributed to a slightly larger equilibrium unit-cell volume in MACELES (due to finite-size effects and polarization rotation in larger supercells), which softens the lattice and stabilizes phases at higher temperatures.

C. Mechanical Properties

• Elastic Constants: MACELES predicted slightly lower elastic constants (softer lattice) than MACE, bringing the values closer to GGA-PBEsol DFT calculations and experimental data.
• Coercive Stress: Both models predicted a similar coercive stress of approximately 120 MPa for stress-induced polarization switching, indicating that long-range interactions have a negligible effect on the critical stress required for switching.

D. Ferroelectric and Dielectric Response

• Hysteresis Loops: Both models produced nearly identical P-E hysteresis loops with comparable remnant polarization and coercive fields. The qualitative switching mechanism (two-step switching) was preserved.
• Dielectric Constants: The inclusion of long-range interactions significantly improved the in-plane dielectric constant ($\varepsilon_a$).
  • MACE: $\varepsilon_a \approx 1070$
  • MACELES: $\varepsilon_a \approx 1440$ (a ~35% increase).
  • Reasoning: The softer lattice in MACELES reduced the tetragonality ($c/a$ ratio), weakening polarization anisotropy and allowing larger in-plane fluctuations, thus enhancing $\varepsilon_a$.

5. Significance and Conclusion

The study concludes that the role of long-range interactions in MLPs for ferroelectrics is property-dependent:

1. Qualitative Topology: Short-range MLPs are sufficient for capturing the topological structure of the PES. They correctly predict which phases are stable, the sequence of transitions, and the mechanism of polarization switching.
2. Quantitative Curvature: Long-range interactions are essential for accurately predicting the local curvature of the PES. This directly impacts quantitative metrics such as transition temperatures, elastic constants, and dielectric constants.

Practical Implication:
Researchers can use computationally cheaper short-range MLPs for exploring phase diagrams and switching mechanisms. However, for applications requiring high-precision quantitative data (e.g., designing capacitors where dielectric constants are critical, or precise thermodynamic modeling), incorporating long-range electrostatics via frameworks like MACELES is necessary. The paper provides a clear guideline for selecting the appropriate level of MLP complexity based on the target material property.