The ground state of CuInP$_2$S$_6$ thin films: A study of the deep potential method

By combining first-principles calculations with the deep potential method, this study resolves the discrepancy between experimental observations and DFT predictions for CuInP$_2$S$_6$ thin films by demonstrating that vibrational entropy stabilizes a ferrielectric ground state with intralayer ferroelectric ordering at finite temperatures.

The Gist

Imagine you have a stack of very thin, magical sheets of paper. These aren't ordinary paper; they are made of a special material called CuInP2S6 (or CIPS for short). These sheets have a superpower: they can act like tiny magnets, but instead of North and South poles, they have "up" and "down" electrical charges. This is called ferroelectricity, and it's the secret sauce behind super-fast, super-small computer memory.

For a long time, scientists have been arguing about how these sheets behave when you stack them up.

The Great Disagreement: The "Map" vs. The "Reality"

Think of this like a debate between a Cartographer (a map-maker) and a Hiker (someone actually walking the terrain).

• The Cartographer (The Computer Model): For years, scientists used powerful computer simulations (called DFT) to predict how these sheets should stack. The computer said, "If you stack these sheets, the charges should cancel each other out perfectly, like a team of people pulling a rope in opposite directions. The net result is zero. We call this an Antiferroelectric (AFE) state." The computer was very confident; it said this was the most stable, lowest-energy state.
• The Hiker (The Experiment): But when real scientists actually made these sheets in a lab and measured them, they found something different. The sheets didn't cancel out. They still had a net electrical charge, acting like a strong magnet. They were behaving like Ferroelectrics (FE).

The map said "No net charge," but the hiker said "There is a charge!" This contradiction was a big problem. If the map is wrong, we can't build reliable devices.

The New Detective: The "Deep Potential" Method

The authors of this paper decided to solve the mystery. They realized the computer map was missing a crucial piece of the puzzle: vibrations.

Imagine the atoms in these sheets aren't frozen statues. They are like a crowd of people at a concert, constantly dancing, jiggling, and vibrating.

• The old computer models only looked at the static energy (how much energy it takes to hold the atoms in a specific pose).
• The new method, called Deep Potential (DP), is like a high-tech camera that can film the atoms dancing. It calculates not just the pose, but the heat and motion (entropy) of the atoms.

The Discovery: The "Ferroelectric" Dance Wins

When the researchers used this new "dance-capturing" method, they found the truth:

1. The Static View (Old Way): If you freeze the atoms in place, the computer is right. The "Antiferroelectric" state (where charges cancel out) looks like the cheapest, most stable option.
2. The Dynamic View (New Way): But atoms aren't frozen! When you let them dance and vibrate (which happens at room temperature), the rules change.
   • The "Ferroelectric" dance (where charges point the same way) is actually easier and more comfortable for the atoms to perform. It costs less "energy" to keep them vibrating in this pattern.
   • The "Antiferroelectric" dance is stiff and awkward when the atoms start moving.

The Result: When you add up the cost of the "dance" (vibrational energy), the Ferroelectric state becomes the winner. It is the true "ground state" (the most natural, stable state) for these materials at room temperature.

The "Goldilocks" Solution: The Ferrielectric State

There was one more twist. The researchers found a middle-ground state called Ferrielectric (FiE).

• Imagine a stack of sheets. The top and bottom sheets are dancing one way, but the middle sheets are dancing slightly differently, creating a mix.
• This state is almost as cheap as the "Antiferroelectric" state in terms of static energy, but once the atoms start dancing, it becomes the most stable state of all.

This explains why experiments see a net charge (because the FiE state has a net charge) while old computer models predicted zero charge (because they ignored the dance).

Why Does This Matter?

This is a huge deal for technology.

• Fixing the Map: It reconciles the difference between what computers predicted and what we see in the lab. We now know the "map" was just missing the "weather" (temperature and vibration).
• Better Devices: Because we now understand that these materials naturally want to be in a charged, active state, we can design better, more reliable memory chips and sensors that use less power.
• The Lesson: Sometimes, to understand how a system works, you can't just look at it sitting still. You have to watch it move, dance, and vibrate.

In short: The paper solved a decades-old mystery by realizing that atoms are energetic dancers, not frozen statues. Once you account for their dance moves, the material behaves exactly as the experiments showed, not as the old computer models predicted.

Technical Summary

1. Problem Statement

The two-dimensional ferroelectric (FE) material CuInP2S6 (CIPS) is a promising candidate for non-volatile memory and nanodevices due to its room-temperature out-of-plane ferroelectricity. However, a significant discrepancy exists between experimental observations and theoretical predictions regarding its ground state in thin films:

• Experimental Evidence: Experiments indicate that CIPS thin films (down to ~4 nm) exhibit a state with net polarization (ferroelectric or ferrielectric behavior).
• Theoretical Predictions (DFT): Standard Density Functional Theory (DFT) calculations predict that the antiferroelectric (AFE) state is the lowest-energy ground state. Specifically, DFT suggests that intralayer AFE ordering is energetically favorable over intralayer FE ordering for monolayers, and interlayer AFE coupling is favored for bilayers and thicker films.

This contradiction hinders the fundamental understanding of CIPS polarization mechanisms and its application in devices. The core issue is that standard DFT considers only electronic energy ($E_{ele}$) at 0 K, neglecting lattice dynamics (phonons) and vibrational entropy, which are critical at finite temperatures.

2. Methodology

To resolve this discrepancy, the authors combined first-principles calculations with a machine learning approach to accurately model the lattice dynamics of large-scale CIPS systems.

• Deep Potential (DP) Method:

  • The authors developed a high-precision Deep Potential (DP) model using the DeepMD-kit software.
  • Training Data Generation: They employed an active learning strategy using the DP-GEN platform. This involved an iterative cycle of training, exploration (molecular dynamics), and labeling (DFT calculations) to generate a dataset of ~18,000 configurations covering bulk, monolayer, and multilayer (up to 4 layers) CIPS with various polarization states.
  • Validation: The DP model was rigorously benchmarked against DFT results for energy, forces, and phonon dispersions. It achieved a root mean square error (RMSE) of $\approx 4.57 \times 10^{-4}$ eV/atom for energy and 0.023 eV/Å for forces. It also correctly predicted the Curie temperature ($T_c \approx 325$ K), matching experimental values.

• Phonon Free Energy Calculation:

  • Unlike standard DFT studies limited to small unit cells, the DP model enabled the calculation of phonon free energy ($E_{ph}$) for large supercells (up to 40 layers, ~26 nm thick).
  • The total free energy at finite temperature $T$ was calculated as: $E_{total}(T) = E_{ele} + E_{ph}(T)$.

• Molecular Dynamics (MD):

  • Long-timescale MD simulations (up to 1000 ns) were performed using the DP model to observe dynamic phase transitions (e.g., FE to FiE) at finite temperatures.

3. Key Contributions

• Development of a Scalable DP Model: Created a highly accurate machine learning potential capable of simulating CIPS thin films with hundreds of atoms, overcoming the computational infeasibility of first-principles phonon calculations for such large systems.
• Resolution of the Theory-Experiment Discrepancy: Demonstrated that the inclusion of phonon free energy (vibrational entropy) reverses the energetic stability of polarization states, reconciling DFT predictions with experimental observations.
• Identification of the True Ground State: Established that the ferrielectric (FiE) state, rather than the AFE or uniform FE state, is the thermodynamically stable ground state for multilayer CIPS at finite temperatures.

4. Key Results

A. Electronic Energy ($E_{ele}$) vs. Total Free Energy

• DFT Results (0 K): Consistent with previous studies, the AFE state (intralayer AFE ordering with surface layers having FE ordering) has the lowest electronic energy. The uniform FE state is significantly higher in energy (e.g., ~189 meV higher for bilayers).
• Phonon Contribution: The intralayer FE ordering possesses a significantly lower phonon free energy than intralayer AFE ordering. This is attributed to the lower frequency phonon modes (higher density of states at low energy) in the FE configuration, which increases vibrational entropy.
• Total Energy ($E_{total}$):
  • For monolayers, the intralayer FE state becomes more stable than AFE when phonon contributions are included.
  • For multilayers (3L and above), a Ferrielectric (FiE) state emerges as the most stable. This state features:
    • Surface Layers: Intralayer FE ordering (aligned with the surface polarization).
    • Inner Layers: Uniform Intralayer FE ordering.
    • Interlayer Coupling: Antiferroelectric coupling between the surface and the sub-surface layer (similar to the AFE surface structure) but with a net polarization due to the inner layers.
  • Temperature Dependence: For 3L-CIPS, the FiE state becomes energetically favorable over the AFE state above 220 K. For thicker films (6L, 20L, 40L), the FiE state remains the most stable across the entire temperature range studied (up to 500 K).

B. Mechanism of Stabilization

• Phonon Density of States (DOS): Analysis reveals that the AFE state has a "blue-shift" (higher frequencies) in the 5–20 THz region compared to FE and FiE states. This results in a higher phonon free energy for AFE.
• Robustness: The preference for intralayer FE ordering is robust against interlayer coupling. While the interlayer coupling determines the specific arrangement (AFE vs. FiE), the intralayer FE ordering drives the overall stability via vibrational entropy.

C. Dynamic Verification

• MD simulations starting from an FE state for 3L-CIPS at 250 K showed a spontaneous transition to the FiE state after 1000 ns, confirming that the FiE state is the thermodynamic attractor at finite temperatures.

5. Significance

• Theoretical Reconciliation: This study provides a definitive explanation for the long-standing conflict between DFT predictions (which favored AFE) and experimental observations (which showed net polarization). It proves that vibrational entropy is the decisive factor stabilizing the ferroelectric/ferrielectric order in CIPS at room temperature.
• Device Implications: Understanding that CIPS naturally adopts a ferrielectric state with net polarization in multilayers validates its suitability for non-volatile memory and logic devices. It suggests that the "ferroelectric" behavior observed in experiments is intrinsic and thermodynamically stable, not an artifact of measurement or surface effects.
• Methodological Impact: The work demonstrates the power of combining Deep Potential machine learning with thermodynamic integration to solve complex phase stability problems in 2D materials where traditional DFT fails due to system size and temperature limitations.

In conclusion, the paper establishes that CuInP2S6 thin films possess a ferrielectric ground state at finite temperatures, driven by the lower phonon free energy of intralayer ferroelectric ordering, thereby resolving the theoretical-experimental gap and guiding future device engineering.