Machine Learning the order-disorder Jahn-Teller transition in LaMnO$_3$

This study employs machine-learning molecular dynamics to demonstrate that the Jahn-Teller structural phase transition in LaMnO$_3$ at approximately 750 K is an order-disorder process driven by the ordering of $Q_2$ distortions, while revealing the persistence of dynamic local distortions above the transition temperature and validating the method's ability to distinguish such mechanisms from displacive behaviors.

The Gist

Imagine a crowded dance floor where everyone is holding hands in a specific, rigid formation. This is how the atoms in a crystal called LaMnO₃ behave when it's cold. They are perfectly organized, stretching and squeezing in a synchronized way to keep their energy low. This specific "dance move" is called the Jahn-Teller effect.

Now, imagine turning up the heat on this dance floor. Usually, when things get hot, people just start dancing randomly and the formation breaks apart. But in this specific material, something more interesting happens around 750 Kelvin (about 477°C). The atoms don't just stop dancing; they stop agreeing on the dance steps, even though they are still moving vigorously.

Here is a breakdown of what the scientists did and what they found, using simple analogies:

1. The Problem: Too Hard to Watch

For a long time, scientists wanted to understand exactly how this "dance" changes from organized to chaotic as the temperature rises.

• The Old Way: Traditional computer simulations were like trying to watch a movie through a keyhole. They were either too small (only a few dancers) or too simple (ignoring the complex electronic rules that govern the dance). They couldn't capture the full picture.
• The New Tool: The researchers used Machine Learning (ML). Think of this as teaching a super-smart robot to watch the atoms. First, they showed the robot a few hours of high-definition, perfect physics simulations (the "training"). Once the robot learned the rules, they let it simulate millions of atoms dancing for a long time. This gave them a crystal-clear, wide-angle view of the transition.

2. The Discovery: Order vs. Disorder

The team was trying to figure out how the transition happens. There are two main theories for how crystals melt their structure:

• The "Displacement" Theory (The Slide): Imagine a row of dominoes. If you push the first one, they all slide over together in a smooth wave to a new position. The order changes, but it's a smooth, collective shift.
• The "Order-Disorder" Theory (The Shuffle): Imagine a group of people standing in a grid. Suddenly, everyone starts shuffling their feet randomly. They are still standing in the grid, but they are no longer facing the same direction. The pattern is gone, but the movement is still there.

What the paper found:
LaMnO₃ follows the Order-Disorder theory.

• Below 750K: The atoms are like a marching band. Every MnO₆ octahedron (a cluster of atoms) is stretched in the same direction. They are "ordered."
• Above 750K: The marching band breaks up. The atoms are still stretching and squashing (dynamically fluctuating), but they are doing it in random directions. One atom stretches left, its neighbor stretches right. The long-range "marching" order is gone, but the local "jiggling" remains.

3. The Evidence: The "Vibrating String"

How did they prove this? They looked at the "sound" of the atoms.

• Imagine a guitar string. If you pluck it gently, it rings out clearly at a specific note (a sharp peak in frequency). This is like a stable, ordered crystal.
• As the temperature rises, the researchers found that this "note" didn't just get quieter; it got fuzzy and broad. The sharp note turned into a muddy, sliding sound.
• This "muddiness" (anharmonicity) is the fingerprint of the order-disorder transition. It proves that the atoms aren't just sliding smoothly to a new spot; they are chaotically jiggling in a way that destroys the long-range pattern.

4. Why Does This Matter?

This isn't just about one specific crystal.

• The "Magic" Material: LaMnO₃ is the parent of a family of materials used in hard drives and sensors because they change their electrical resistance massively when you apply a magnetic field (Colossal Magnetoresistance).
• The Key to Control: To make better electronics, we need to understand exactly how these materials switch states. By proving that this transition is driven by "disorder" (randomness) rather than "displacement" (sliding), scientists can now design better materials.
• The Method: The biggest takeaway is the method itself. Combining Machine Learning with Molecular Dynamics is like giving scientists a time machine. It allows them to see the microscopic "dance" of atoms at high temperatures with an accuracy that was previously impossible.

Summary

Think of the material as a room full of people holding hands in a perfect circle.

• Cold: Everyone holds hands and faces the center. Perfect order.
• Hot: The circle breaks. People are still holding hands and moving, but they are facing random directions. The structure of the circle is gone, but the movement continues.

The paper used a "smart robot" to watch this happen, proving that the transition is a chaotic shuffle, not a smooth slide. This helps us understand how to build the next generation of super-efficient electronic devices.

Technical Summary

1. Problem Statement

The paper addresses the microscopic mechanism of the structural phase transition in LaMnO$_3$ (LMO) occurring at approximately 750 K. LMO is a prototypical correlated electron system where the Mn$^{3+}$ ion exhibits a Jahn-Teller (JT) effect due to the partial occupation of the $e_g$ orbitals.

• The Challenge: At low temperatures, LMO is an orthorhombic insulator with long-range orbital order and static JT distortions. Upon heating, it undergoes a transition to a metrically cubic (pseudo-cubic) phase.
• The Debate: While it is known that this transition involves the disordering of JT distortions, the precise nature of the transition (order-disorder vs. displacive) and the behavior of local distortions at finite temperatures have been difficult to resolve.
• Limitations of Previous Methods:
  • Standard DFT: Limited by small system sizes and the inability to simulate long-time dynamics at finite temperatures.
  • Classical MD: Lacks accurate electronic degrees of freedom (electron-electron correlations) and temperature-dependent electronic states, leading to inaccurate descriptions of complex electron-lattice transitions.

2. Methodology

The authors employ a hybrid approach combining Machine-Learning Force Fields (MLFF) with Ab Initio Molecular Dynamics (AIMD) to overcome the limitations of traditional methods.

• Force Field Training:
  • Software: VASP (Vienna Ab Initio Simulation Package) with on-the-fly MLFF training.
  • DFT Setup: Generalized Gradient Approximation (PBE) + Hubbard $U$ correction (Dudarev approach) to handle strong electron correlations.
  • Training Data: Generated via linear temperature ramps (600 K to 1100 K) on a $\sqrt{2}a \times \sqrt{2}b \times c$ supercell (8 formula units).
  • Parameters: Tested effective on-site interactions ($U_{eff}$) of 0, 2, and 3.5 eV. The final production runs utilized $U = 3.5$ eV to accurately capture the JT physics.
• Production Simulations:
  • System Size: A large supercell comprising 216 formula units (1080 atoms) to minimize finite-size effects.
  • Ensembles: NPT and NVT simulations using Langevin thermostats and Parrinello-Rahman barostats.
  • Temperature Range: Simulations performed at fixed temperatures (100 K to 900 K) and via heating ramps.
• Analysis Techniques:
  • JT Normal Coordinates: Calculation of $Q_2$ and $Q_3$ modes using the VanVleckCalculator to quantify octahedral distortions.
  • Correlation Functions: Analysis of site-site correlation functions ($C_{ij}$) of the $Q_2$ modes to determine long-range order.
  • Phonon Spectral Analysis: Calculation of the Velocity Autocorrelation Function (VACF) projected onto specific phonon eigenvectors (specifically the $A_g(1)$ mode) to derive temperature-dependent spectral functions.

3. Key Contributions

1. Validation of MLFF for Correlated Systems: Demonstrated that MLFFs trained on DFT+$U$ data can accurately reproduce the complex order-disorder transition in a correlated transition-metal oxide, bridging the gap between electronic structure accuracy and statistical mechanics sampling.
2. Identification of the Order Parameter: Confirmed that the $Q_2$ Jahn-Teller distortion acts as the primary order parameter for the transition.
3. Mechanism Clarification: Provided definitive evidence that the transition is order-disorder in nature, rather than displacive. This was achieved by distinguishing between the disappearance of long-range order and the persistence of local dynamic distortions.
4. Anharmonicity Characterization: Quantified strong anharmonic effects in the lattice dynamics, showing a breakdown of the phonon quasiparticle picture well below the transition temperature.

4. Key Results

A. Structural Evolution and Transition Temperature

• Role of Hubbard $U$: Simulations with $U=0$ failed to reproduce the transition, underestimating JT modes. Only with $U=3.5$ eV did the simulations correctly predict the progressive suppression of JT modes and the orthorhombic-to-pseudocubic transition.
• Transition Temperature: The MLFF predicted a transition temperature ($T_c^{JT}$) of ~600 K, slightly lower than the experimental 750 K. The authors attribute this discrepancy to the limitations of the single-parameter $U$ correction, noting that more advanced functionals (e.g., hybrid functionals) would likely improve quantitative agreement.
• Lattice Parameters: The simulations successfully reproduced the experimental convergence of lattice parameters toward a cubic symmetry in the 570–620 K range.

B. Nature of the Transition (Order-Disorder)

• Persistence of Local Distortions: Unlike a displacive transition where distortions vanish, the study found that dynamic JT distortions persist well above the transition temperature (up to 1150 K in experiments, and observed as non-zero fluctuations in simulations).
• Correlation Functions:
  • Below $T_c$: Strong long-range anti-correlation of $Q_2$ modes (checkerboard pattern).
  • Above $T_c$: Long-range order melts completely; correlations decay rapidly to the nearest neighbor, but local fluctuations remain significant.
  • Distribution: The probability distribution of $Q_2$ shifts from a double-peak structure (ordered) to a single broad peak centered at zero (disordered), confirming the loss of static order while retaining dynamic disorder.

C. Phonon Dynamics and Anharmonicity

• $A_g(1)$ Mode Analysis: The authors analyzed the in-phase anti-stretching mode ($A_g(1)$), which shares symmetry with the static distortion.
• Spectral Evolution:
  • At 300 K, the mode shows a double-peak structure, indicating that anharmonic effects are already strong enough to break the phonon quasiparticle picture.
  • As temperature increases, the mode softens and broadens significantly.
• Distinction from Displacive Transitions: In displacive transitions (e.g., in incipient ferroelectrics like SrTiO$_3$), the soft mode typically hardens with temperature. The observed softening and broadening in LMO is a distinct fingerprint of an order-disorder transition.

5. Significance

• Methodological Framework: The paper establishes a robust workflow combining Machine-Learning Interatomic Potentials (MLIPs) with VACF analysis as a general tool for investigating structural phase transitions in correlated materials. This approach allows access to length and time scales inaccessible to standard AIMD or classical MD.
• Physical Insight: It resolves the debate on the nature of the LMO transition, confirming it as a genuine order-disorder process driven by the thermal disordering of the $Q_2$ mode, while local lattice distortions remain dynamic.
• Future Applications: The ability to distinguish between order-disorder and displacive mechanisms via the temperature evolution of vibrational properties provides a predictive framework for studying other complex materials, including those with anharmonic or electronically driven transitions.

In summary, this work leverages advanced machine learning to simulate the thermodynamics of a strongly correlated material, successfully reproducing experimental trends and providing a microscopic explanation for the order-disorder nature of the Jahn-Teller transition in LaMnO$_3$.