WorksPhysics and Astronomy
Diverse polymorphism in Ruddlesden-Popper chalcogenides
Abstract
Ruddlesden-Popper (RP) chalcogenides are stable, non-toxic candidates for optoelectronic or thermoelectric applications. The structural diversity of RP oxides is already exploited to tune properties or achieve more advanced functionalities like multiferroicity, however, little is known about the structural evolution of RP chalcogenides. In this work, we develop a high-accuracy machine-learned interatomic potential to run large-scale molecular dynamics simulations on for to . We predict new polymorphs for each -value, calculate their corresponding phase transition temperatures, and validate our approach through comparison to published experimental results. We find that the phase exhibits negative thermal expansion, that and undergo unusual ascending symmetry breaking, and that phases with form layer-dependent tilt patterns previously unreported for inorganic RP materials. This unique behaviour results from competition between octahedral rotations and rumpling at the rocksalt interface, and suggests new strategies for accessing advanced functionalities.
Three-dimensional networks of corner-sharing octahedra constitute the fundamental structural units of perovskite materials. Layered perovskites, which consist of single or multiple perovskite layers separated by thin inorganic or organic spacers, have also been widely investigated. Perovskite oxides and their layered counterparts have received extensive attention due to their pronounced structural versatility, which can be exploited to probe fundamental physical phenomena and enable advanced functionalities such as superconductivity, ferroelectricity, and negative thermal expansion [1, 2, 3, 4]. A crucial step toward designing new perovskite-based functional materials is thus identifying the accessible polymorphs; however, this remains challenging because the structural distinctions between polymorphs are often subtle and difficult to resolve experimentally or theoretically [5].
Perovskite-based chalcogenides (X = S, Se) are an emerging class of tunable semiconductors for optoelectronic and thermoelectric applications [6, 7, 8, 9, 10, 11, 12]. Among them, is the most studied, exhibiting strong visible-light absorption and photoluminescence, defect tolerance, and chemical stability [13, 14, 15, 16]. The series is known to form in the RP structure, in which perovskite slabs are separated by rocksalt layers [17, 18, 19, 20, 21, 22] (Figure 1a). () is reported in the high-symmetry phase with untilted octahedra, () is reported in both the and phases, and () in the phase. While distortions to other phases are expected upon cooling or heating, as in other perovskite-like materials, these transitions remain largely unexplored in RP chalcogenides, leaving a critical gap in understanding their structural evolution and the potential for novel functional properties.
Figure 1. a) Ruddlesden-Popper (RP) crystal structures in the high-symmetry phase. The primitive cell is formed from two unique perovskite slabs with a half-cell offset. Each slab contains layers of octahedra; the limit is equivalent to . Common distortions in RP materials: b) in-phase () and c) out-of-phase () octahedral tilting; d) out-of-plane rumpling distortion; e) in-plane cation displacement. The rumpling amplitude is defined as positive when A-site cations at the rocksalt layer move towards the octahedra. In-plane refers to displacements along the - or -axes.
Here, we address this gap by combining a high-accuracy, machine-learned interatomic potential with large-scale molecular dynamics, revealing a rich polymorphic landscape and unexpected structural behaviour across –. This approach identifies the onset of 3D perovskite-like behaviour and shows how structural distortions evolve with temperature and -value.
| Slab I | Slab II | Space group | |
| 1 | |||
| 2 | |||
| 3 | |||
| 4 | |||
| 5 | |||
| 6 |
Table 1. Predicted ground state tilt patterns and resulting space groups for BaZrS materials, where is the number of perovskite layers within each slab. Aleksandrov notation is used to denote the out-of-phase () and in-phase () octahedral tilts along each axis, with an overbar to indicate that the tilt patterns between adjacent slabs are in opposite directions [44].
For each -value, we performed density functional theory (DFT) calculations of energies, forces, and stresses for the 33 unique tilt patterns enumerated by Aleksandrov and Bartolomé [23, 24, 25]. This data formed the initial training set for a neuroevolution potential (NEP) model, which enables efficient molecular dynamics sampling of the potential energy surface [26, 27] [28]. The final model, trained on 1375 RP and perovskite HSE06-evaluated structures, achieves a formation energy root mean squared error of 1.8 meV atom, providing a highly accurate representation of the energy landscape [29]. Additional computational details, including a comparison of DFT- and NEP-calculated phonon spectra, are provided in the Supplemental Material [30–43].
Table I shows the predicted ground state tilt patterns of the RP series, which can be divided into two groups. Low- phases (–3) feature a single out-of-phase tilt along the -axis in slab I and along the -axis in slab II. High- phases (–6) exhibit enhanced tilting in both slabs, with out-of-phase tilts along the - and -axes, and an in-phase tilt along the -axis. BaZrS adopts the same tilt pattern in its low-temperature structure, indicating that for , the BaZrS ground states converge to that of the parent perovskite. Across all -values, the ferroelectric phase is metastable but lies very close in energy, within 2 meV atom, to the paraelectric ground state. NEP-calculated energies for all 198 tilt patterns are provided in Tables S1 and S2.
We investigated the temperature evolution of RP structures via molecular dynamics (MD) heating simulations starting from the ground state polymorphs at 0 K (Figure 2). Lattice parameters cluster near 5.06 Å for low- phases and around 5.01 Å for high-, with the latter contracted due to the increased in-plane tilting in their ground state structures. Inspection of heat capacity peak positions and widths, along with octahedral tilt angles (Figure S18–Figure S23), indicates that high- compounds experience two transitions: a low-temperature, first-order transition, followed by a high-temperature, second-order transition. As increases, there is evolution from layered
Figure 2. Neuroevolution potential heating simulations for BaZrS: a) in-plane lattice parameters. Scatter points at 0 K correspond to the lattice parameters from HSE06 relaxations. Note the perovskite lattice parameter is for a conventional cubic cell; b) heat capacities. A sharper peak in the heat capacity indicates a first-order (discontinuous) phase transition, whilst a broader peak indicates more second-order (continuous) character.

Figure S18. Euler angles of the RP phase

Figure S23. Euler angles of the RP phase
to bulk-like behaviour, with transition temperatures progressively converging to the BaZrS perovskite limits at 650 K and 880 K.
BaZrS exhibits negative in-plane thermal expansion up to the second order phase transition at 420 K (Figure 2a), at which point the linear thermal expansion coefficient, , is K. In contrast, BaZrS and BaZrS display positive in-plane expansion with K at 420 K. These expansion coefficients are an order of magnitude smaller than those of the high- materials and BaZrS, with K. Negative thermal expansion has been previously reported for RP oxides with and , where the sign and magnitude of thermal expansion are highly sensitive to small changes in symmetry [4]. Recent quasi-harmonic calculations predict a negative volumetric thermal expansion in BaZrS at ambient pressure and below 10 K [45].
A summary of all predicted polymorphs and transition temperatures for the RP series and BaZrS is shown in Figure 3. For RP materials with –3, three distinct structural evolution pathways are observed. In contrast, all phases with follow an identical sequence of tilt patterns, matching that of the parent perovskite. By 830 K, all -values have transitioned into the untilted parent phase, which is equivalent to the hightemperature cubic structure of BaZrS perovskite [46].
Figure 3. Summary of space groups, tilt patterns, and phase transition temperatures for BaZrS. Bars are colour-coded by tilt pattern; see Table S3 for Aleksandrov tilt notation. Space groups, labelled in text, depend on both the tilt pattern and the parity of . Solid, dashed, and dotted lines mark first-order, second-order, and surface transitions, respectively. Gradient fade on indicates a possible low-temperature phase.
Comparison against a wider range of octahedral tilt patterns catalogued by Liu et al. reveals one additional low-temperature configuration in BaZrS, () [47]. This is marginally stabilized by 0.4 meV atom at 0 K, with harmonic phonon calculations confirming its stability up to 100 K. Kinetic barriers prevent its formation on cooling, and delay formation of the phase on heating. Therefore, all analyses and heating simulations are based on the ground state structure. The same high-temperature phases (, , and ) are recovered on heating, regardless of the starting configuration.
We computed X-ray diffraction (XRD) patterns for BaZrS and BaZrS to compare with available experimental data [21] [48] (Figure S30 and Figure S31). For BaZrS, the simulated pattern shows excellent agreement with experiment, confirming the formation of the phase at 300 K and validating our computational approach. In BaZrS, two phase transitions occur near room temperature. The to transition is second-order, leading to a gradual reduction of the soft mode XRD peaks, and complicating phase identification. We therefore tentatively assign the room-temperature structure as . Additional single-crystal XRD measurements, complemented by Raman spectroscopy, would help confirm this assignment more definitively.

Figure S30. X-ray diffraction patterns for different phases of Ba₃Zr₂S₇. Experimental results are from Ref. 24.

Figure S31. X-ray diffraction patterns for different phases of BaZrS. Experimental results are from Ref. 26 and Ref. 27.
Both the to transition in BaZrS and the to transition in BaZrS correspond to a reduction in symmetry with increasing temperature. Such ascending symmetry breaking in a displacive phase transition is unusual, as higher-symmetry phases typically stabilize at elevated temperatures [49], and generally indicate the presence of competing interactions with similar energy scales near the transition boundary [50]. In BaZrS, the symmetry lowering occurs via a first-order transition driven by enhanced in-plane octahedral tilting (Figure S24), coinciding with negative in-plane thermal expansion, whereas in BaZrS it notably occurs via a continuous transition.

Figure S24. Snapshots of Euler angles (, , ) of the octahedra in the a) , b) , c) and d) RP polymorphs at four temperature points. The octahedra are colour-coded according to the magnitude of their tilt angles. Solid lines separate the structures by temperature, dashed lines separate the structures by tilt axis. The structures are viewed along the x-axis for and , and along the y-axis for .
To probe the microscopic mechanism in BaZrS, we analyse the layer-resolved tilt angles of the ZrS octahedra along each axis at four different temperatures (Figure 4a). At 1 K, the space group with a tilt pattern . By 210 K, the system has transitioned to a lower-symmetry phase, , through the introduction of -axis tilting confined to the central layer of each slab. Similar interface-dependent tilt effects, where the tilt angle varies with distance from the rocksalt layer, have been reported in simulations of hybrid halide RP materials [42]; to our knowledge, this is the first observation of such behaviour in an inorganic RP system.
Figure 4. Snapshots of Euler angles for for a) and b) . Octahedra are colour-coded by the magnitude of their Euler tilt angles. Solid lines separate structures by temperature, and dashed lines distinguish tilt axis. Structures are viewed along the -axis for and , and along the -axis for .
We also analyse tilting in BaZrS as an example of large- behaviour (Figure 4b). At 1 K, the tilt configuration is (). By 250 K, tilting along the -axis at the perovskite-rocksalt interface is suppressed, although the space group remains unchanged (we refer to this as a “surface transition”). This analysis demonstrates that the perovskite-rocksalt interface impacts on RP structural behaviour in a number of ways. For high- RP materials it enables the surface transitions and layer-dependent tilt patterns, whilst for low- materials, it hinders octahedral tilting at low temperature, so that tilts within a slab are along a single axis only.
To better understand these interface effects, we examine cation displacements and their coupling to octahedral tilts. Out-of-plane Ba displacements in the interior layers, as well as Zr displacements from their high-symmetry positions, are strongly suppressed (Figure S25). In-plane Ba displacements are coupled to octahedral tilts and preserve the symmetry of the space groups (Figure S26–Figure S29).

Figure S25. Layer-by-layer rumpling amplitude in the to 6 RP phases

Figure S26. A-site displacements in the RP phase

Figure S29. A-site displacements in the RP phase
At the perovskite-rocksalt interface we observe out-of-plane Ba displacements, so that the Ba and S ions no longer lie in the same plane (Figure 1d). This distortion is termed “rumpling” in the literature. We predict ground state rumpling amplitudes consistent with other materials that have uncharged AX rocksalt layers (Figure 5a and Fig. 5b) [51]. We also find competition between out-of-plane octahedral rotations and rumpling, as has been previously reported for RP materials [51–53]. Figure 5b shows that, for large- RP materials only, the rumpling amplitude increases as the out-of-plane tilts at the perovskite-rocksalt interface decrease (Figure S25-Figure S27), reaching a maximum at the surface transition where out-of-plane tilts vanish. The increase in rumpling amplitude is also accompanied by negative out-of-plane thermal expansion, as shown in Figure 5c.
Figure 5. Rumpling amplitudes () in RP materials with a) to 3 and b) –6. Scatter points at 0 K correspond to the rumpling amplitudes from HSE06 relaxations. Surface transition temperatures are indicated by dashed vertical lines. c) Normalised out-of-plane lattice parameters for –6. One-dimensional potential energy surfaces for out-of-plane, in-phase octahedral tilts () in () are shown as a function of d) rumpling and e) in-plane, out-of-phase tilt amplitudes (). Distortions are illustrated in Figure 1.

Figure S27. A-site displacements in the RP phase
The inverse coupling between octahedral rotations and rumpling is further supported by analysis of the corresponding phonon modes: large rumpling amplitudes stabilise structures with zero out-of-plane tilting (), while decreasing rumpling amplitudes allows the tilt amplitude to grow (Figure 5d). The competition arises because both distortions adjust Ba–S distances; out-of-plane tilting shortens Ba–S bonds, leading to over-coordination of Ba when combined with rumpling (see Table S4). Rumpling and its competition with octahedral tilting thus provides an understanding of why surface transitions occur in these materials.
| Ground state | ||
| 1 | 3.56 | 3.26 |
| 2 | 3.58 | 3.24 |
| 3 | 3.59 | 3.24 |
| 4 | 3.59 | 3.21 |
| 5 | 3.59 | 3.21 |
| 6 | 3.59 | 3.21 |
Table S4. In-plane Ba-S distances in for the untilted structures and the ground-state tilted structure. Distances are given in Å. In both cases a rumpling distortion has been applied.
The unusual symmetry-lowering to phase transition in also arises from the delicate interplay between in-plane and out-of-plane octahedral tilts coupled with rumpling. In the ground state, large in-plane tilt amplitudes suppress the out-of-plane tilt (Figure 5e). As temperature increases, decreases, making the tilt energetically favourable and lowering the overall symmetry. The outermost perovskite layers remain untilted due to rumpling at the perovskite–rocksalt interface. In RP materials with , all perovskite layers are adjacent to rocksalt layers, suppressing tilts and associated symmetry-lowering transitions. For , tilts are already active at 0 K, preventing symmetry reduction. Thus the material represents a special intermediate case, featuring suppressed tilts at low temperature alongside an internal perovskite layer.
In this work, we have developed a high-accuracy machine-learned interatomic potential to systematically explore the temperature- and -dependent structural evolution of the RP chalcogenide series . Our simulations reveal a rich and diverse set of polymorphs exhibiting notable phenomena: negative thermal expansion, symmetry-lowering phase transitions, and complex, layer-dependent tilt patterns. Analysis shows that this structural diversity arises from the interplay between octahedral tilt modes and rumpling distortions at the perovskite-rocksalt interface. The methodology can be extended to investigate phase stabilization strategies, such as selective A-site doping at the perovskite-rocksalt interface, or applied to understand the dynamics of related materials with displacive structural phase transitions.
Acknowledgments
P.K. gratefully acknowledges funding through the Turing Scheme, which enabled a research visit to Chalmers University of Technology, and support from the UK Engineering and Physical Sciences Research Council (EPSRC) CDT in Renewable Energy Northeast Universities (ReNU) via Grant EP/S023836/1. This work also benefits from COST Action RenewPy CA21148, supported by COST (European Cooperation in Science and Technology), as well as funding from the Swedish Research Council (Nos. 2020-04935 and 2021-05072) and the Knut and Alice Wallenberg Foundation (No. 2024.0042). Computational resources were provided by the Oswald High-Performance Computing Facility at Northumbria University, the ARCHER2 UK National Supercomputing Service through the HEC Materials Chemistry Consortium (EPSRC EP/X035859), and the UK Materials and Molecular Modelling Hub (EPSRC EP/T022213, EP/W032260, EP/P020194). Additional support came from the National Academic Infrastructure for Supercomputing in Sweden (NAISS) at NSC, PDC, and C3SE, partially funded via the Swedish Research Council (Grant 2022-06725), and the Berzelius supercomputing resource, provided by the Knut and Alice Wallenberg Foundation at NSC.
We thank Kevin Ye and Shanyuan Niu for providing experimental X-ray diffraction data, and Kevin Ye, Shanyuan Niu, Emma McCabe, and Nick Bristowe for valuable discussions related to this work.
Data Availability Statement
The NEP models produced in this study are publicly accessible via Zenodo: https://doi.org/10.5281/zenodo.15829577. The DFT output data have been deposited in the NOMAD repository: https://doi.org/10.17172/NOMAD/2025.07.11-1. A paper-specific repository, containing Python code to reproduce the figures and analyses, is available at https://github.com/NU-CEM/2025_RP_phases.
Supplemental Material:
Methods
Neuroevolution potential method
An iterative strategy was employed to construct the neuroevolution potential outlined in Ref. 1. The GPUMD package [2, 3, 4] (version 3.9.5) was used to build the neuroevolution potential (NEP) model and run the molecular dynamics (MD) simulations. The ASE[5] and CALORINE[6] packages were used to prepare the training structures, set up MD simulations and post-process the results. Random displacements were generated using the HIPHIVE package[7].
The initial training set contains strained primitive structures and rattled supercells of the perovskite and the to 6 RP phases. The Ruddlesden-Popper training structures include ideal, distorted, and deformed unit cells and supercells for to , along with the 33 tilt pattern structures defined by Aleksandrov for to . Molecular dynamics (MD) structures were included for temperatures up to 1000 K. The different phases considered for the perovskite are described in Ref. 8. The final training set consists of 1375 structures.
The initial model was trained using density functional theory (DFT) data generated with the PBEsol exchange-correlation functional[9]. For each structure in the training set, formation energy (relative to the elemental phases), atomic forces and stress tensors were evaluated using DFT. This model was then used to run MD simulations in the NPT ensemble, over a temperature range of 1 K to 1000 K with varying supercell sizes, containing between 20 and 500 atoms. The model was retrained after adding snapshots from randomly selected structures from the MD runs.
On the final training set, we used higher accuracy HSE06 functional[10] to calculate energies, forces, and stress tensors. Stress tensors for structures containing more than 60 atoms were not evaluated as the memory requirements for these calculations were prohibitively large.
Density functional theory calculations
Formation energies, forces, and stress tensors used in the training set were calculated using the all-electron numeric atom-centered orbital code FHI-aims[11]. Due to large memory requirements, the stress tensors are not calculated for structures larger than 60 atoms in the unit cell. FHI-vibes[12] was used for pre and post-processing of DFT data. All DFT calculations used the light basis set and a Monkhorst-Pack -point mesh with a minimum -spacing of . For single-point calculations, the charge density was converged to an accuracy of , forces to , and stresses to . Geometry relaxations were carried out using the symmetry-constrained relaxation scheme as implemented in ASE[5], until the maximal force component was below . Harmonic phonon dispersions at 0 K using the PHONOPY package[13] with a supercell and a displacement distance. Forces for these phonon calculations were converged below .
Molecular dynamics simulations
To perform the heating runs, the MD simulations were carried out using the GPUMD software[14] with a timestep of 1 fs. The ASE[5] and calorine[6] packages were used to set up the MD simulations and post-process the results. Heating and cooling simulations were run in the NPT ensemble between 0 K to 1200 K for 50 ns using a supercell consisting of about 40 000 atoms. The potential energy and lattice parameters were recorded every 100 fs to discern phase transitions.
Additional cooling simulations were performed for both the Ruddlesden-Popper series and the perovskite phase, starting from the ideal and structures, respectively. For RP phases with and for the perovskite, we observe significant hysteresis during cooling, where the system becomes kinetically trapped in a metastable phase and does not undergo the expected first-order transition, persisting in its state down to 0 K.
The heat capacities across the different phase transitions were calculated by fitting the potential energy and the temperature data from MD to an analytical expression and then taking its derivative as done in Ref. 15. The analytical function for the energy is a polynomial in temperature for each phase, splined together with an S-like error-function, where the location and width of this splining are also fit parameters. This ensures that the analytical form can capture sharp discontinuous changes in the energy as well as broad continuous changes in the energy across phase transitions, and therefore allows a more reliable estimate of the heat capacity to be obtained compared to direct numerical differentiation.
To identify the finite-temperature polymorphs stabilized during heating, atomic displacements were projected onto the in-phase and out-of-phase octahedral tilt eigenmodes of the ideal phase (Figure S6-Figure S17). These mode projections were evaluated separately for the two inequivalent slabs within the Ruddlesden-Popper structure. For the perovskite phase, projections were instead based on the phonon eigenvectors of the structure, following the methodology in Refs. 16. The space groups of RP structures exhibiting suppressed surface-layer tilting were assigned by constructing idealized structural models with the same octahedral tilt as observed in MD, and evaluating them with spglib[17].
Figure S6. Mode projections in the heating run for the RP phase.

Figure S17. Mode projections in the cooling run for the RP phase.
Structural visualizations and analysis were performed using the OVITO package[18].
Model validation
The final NEP model achieved root mean squared errors of 1.8 meV/atom for formation energies, 68.4 meV/ for atomic forces, and 26.1 meV//atom for virial stresses. Loss curves, root mean square errors and parity plots are provided below.
Figure S1. Loss curves and root mean square errors (RMSE) of energies, forces, and virials for the full model.
Figure S2. Parity plots for the full model. Structures for which the stress tensors have not been evaluated with DFT are not included in the virial and stress tensor plots.
The 0 K harmonic phonon spectra for Ruddlesden-Popper materials in the high-temperature phase are displayed below. The solid blue lines correspond to phonons generated by evaluating forces from the NEP model, and the dashed black lines are phonons generated by evaluating forces from DFT. The largest discrepancies correspond to high-frequency modes associated with the sulfur species. [19]
Figure S3. Phonons and DOS of the RP in the phase using NEP (solid red) and DFT (dashed black)
Figure S4. Phonons and DOS of the RP in the phase using NEP (solid red) and DFT (dashed black)
Figure S5. Phonons and DOS of the RP in the phase using NEP (solid red) and DFT (dashed black)
Key sources of error in the predicted phase transition temperatures of halide perovskites, using the same NEP-based methodology, have been analysed in a previous publication [1]. This includes analysing the impact of heating and cooling rates, supercell size, model uncertainty and exchange-correlation functional. The error associated with simulation parameters can be minimised through careful convergence, as we have done here, however the largest source of error in most studies is likely the exchange-correlation functional, with phase transition temperatures varying by as much as 150 K (less for closely related functionals describing the same physics). In this work we train our NEP on a more accurate electronic structure method (the HSE06 functional[20]), thus reducing the largest source of error. Moving to a higher level of theory (e.g. RPA) may show another shift in transition temperature, however it is currently intractable to evaluate the required number of structures and system sizes with a method beyond hybrid DFT. We note that the difference between the predicted and experimentally measured orthorhombic-to-tetragonal phase transition temperature in , using the same HSE06-NEP methodology as outlined here, is only 50 K[8], which provides an estimate of the expected error bar in this work.
Ground state energies
NEP-calculated energies of the 33 unique octahedral tilt patterns, after geometry relaxation with constrained symmetry. Aleksandrov notation is used to denote the out-of-phase () and in-phase () octahedral tilts along each axis, with an overbar to indicate that the tilt patterns between adjacent slabs are in opposite directions[21]. Subscripts are used to differentiate between tilts of different magnitude, where e.g. . The energies are specified relative to the ground state (which is indicated in bold). Some tilt patterns have the same energy. In this case, the lower-symmetry phase is unstable and has relaxed to the higher symmetry phase. The ground state structures have been confirmed through DFT electronic relaxations with the HSE06 functional.
| space group | Slab I | Slab II | |||
| 000 | 000 | 0.00719 | 0.01314 | 0.01909 | |
| 000 | 0.00719 | 0.00516 | 0.00737 | ||
| 000 | 0.00719 | 0.01030 | 0.01451 | ||
| 0.00719 | 0.00516 | 0.00737 | |||
| 0.00719 | 0.00516 | 0.00467 | |||
| 000 | 0.00719 | 0.01030 | 0.01328 | ||
| 0.00719 | 0.00516 | 0.00467 | |||
| 0.00456 | 0.00245 | 0.00441 | |||
| 0.00486 | 0.00416 | 0.00538 | |||
| 0.00456 | 0.00245 | 0.00437 | |||
| 0.00000 | 0.00000 | 0.00287 | |||
| 0.00082 | 0.00194 | 0.00399 | |||
| 0.00000 | 0.00000 | 0.00284 | |||
| 0.00000 | 0.00000 | 0.00005 | |||
| 0.00000 | 0.00000 | 0.00286 | |||
| 0.00456 | 0.00245 | 0.00114 | |||
| 0.00456 | 0.00245 | 0.00492 | |||
| 0.00000 | 0.00000 | 0.00287 | |||
| 0.00000 | 0.00000 | 0.00287 | |||
| 0.00000 | 0.00245 | 0.00439 | |||
| 0.00000 | 0.00245 | 0.00439 | |||
| 0.00486 | 0.00191 | 0.00114 | |||
| 0.00486 | 0.00191 | 0.00466 | |||
| 0.00486 | 0.00192 | 0.00114 | |||
| 0.00486 | 0.00192 | 0.00467 | |||
| 0.00082 | 0.00013 | 0.00005 | |||
| 0.00082 | 0.00013 | 0.00371 | |||
| 0.00082 | 0.00005 | 0.00000 | |||
| 0.00082 | 0.00005 | 0.00367 | |||
| 0.00486 | 0.00192 | 0.00114 | |||
| 0.00082 | 0.00013 | 0.00005 | |||
| 0.00486 | 0.00191 | 0.00302 | |||
| 0.00082 | 0.00005 | 0.00195 |
Table S1. Space groups, tilt patterns, and NEP-calculated energies of RP structures with an odd number of layers within each slab. Energies are given in eV/atom and referenced to the ground state.
| space group | Slab I | Slab II | |||
| 000 | 000 | 0.01054 | 0.01661 | 0.02089 | |
| 0.01034 | 0.00921 | 0.00798 | |||
| 000 | 0.01054 | 0.01279 | 0.01575 | ||
| 0.01034 | 0.00921 | 0.00798 | |||
| 0.00826 | 0.00482 | 0.00457 | |||
| 000 | 0.01032 | 0.01202 | 0.01419 | ||
| 0.00826 | 0.00482 | 0.00457 | |||
| 0.00000 | 0.00166 | 0.00373 | |||
| 0.00200 | 0.00315 | 0.00457 | |||
| 0.00000 | 0.00166 | 0.00360 | |||
| 0.00340 | 0.00356 | 0.00502 | |||
| 0.00496 | 0.00487 | 0.00573 | |||
| 0.00340 | 0.00356 | 0.00488 | |||
| 0.00340 | 0.00141 | 0.00095 | |||
| 0.00340 | 0.00354 | 0.00523 | |||
| 0.00000 | 0.00006 | 0.00004 | |||
| 0.00000 | 0.00166 | 0.00431 | |||
| 0.00340 | 0.00356 | 0.00500 | |||
| 0.00340 | 0.00355 | 0.00520 | |||
| 0.00000 | 0.00166 | 0.00371 | |||
| 0.00000 | 0.00166 | 0.00408 | |||
| 0.00200 | 0.00006 | 0.00004 | |||
| 0.00200 | 0.00248 | 0.00401 | |||
| 0.00200 | 0.00000 | 0.00000 | |||
| 0.00200 | 0.00254 | 0.00405 | |||
| 0.00496 | 0.00141 | 0.00096 | |||
| 0.00496 | 0.00394 | 0.00479 | |||
| 0.00496 | 0.00141 | 0.00095 | |||
| 0.00496 | 0.00392 | 0.00479 | |||
| 0.00200 | 0.00006 | 0.00004 | |||
| 0.00496 | 0.00141 | 0.00095 | |||
| 0.00200 | 0.00248 | 0.00310 | |||
| 0.00496 | 0.00392 | 0.00399 |
Table S2. Space groups, tilt patterns, and NEP-calculated energies of RP structures with an even number of layers within each slab. Energies are given in eV/atom and referenced to the ground state.
Projection onto octahedral tilt modes
To identify the finite-temperature polymorphs, molecular dynamic trajectories were projected onto the (out-of-phase) and (in-phase) octahedral tilt modes in the heating and cooling runs of the to 6 RP phases. The tilt components are resolved along the , , and directions for each slab, which is identified by subscript index. The sign of the projection amplitudes are used to determine the relative tilt directions between slabs, with opposite tilt directions denoted with an overbar in Aleksandrov notation.
For the heating and cooling runs produce identical mode projection results, apart from sometimes an arbitrary flip of the sign of the mode coordinates. However, for and above in the cooling runs sometimes one does not achieve perfect long-range ordering between the slabs and thus the global mode coordinate (averaged over all slabs) can take on different values or sometimes be zero while the structure still is tilted and looks similar to the structure of the heating run.
Figure S7. Mode projections in the cooling run for the RP phase.

Figure S8. Mode projections in the heating run for the RP phase.

Figure S9. Mode projections in the cooling run for the RP phase.
Figure S11. Mode projections in the cooling run for the RP phase.

Figure S12. Mode projections in the heating run for the RP phase.

Figure S13. Mode projections in the cooling run for the RP phase.
Figure S14. Mode projections in the heating run for the RP phase.
Figure S15. Mode projections in the cooling run for the RP phase.

Figure S16. Mode projections in the heating run for the RP phase.
Tilt angle heatmaps
Layer-by-layer Euler angles (, , ) of the ZrS octahedra in heating runs of the to 6 RP phases.
In RP materials with , the out-of-plane tilts at the perovskite-rocksalt interface (layer 1 and layer ) decrease as the rumpling amplitude increases (5b in the main text). The tilts vanish at the maximum of the rumpling amplitude.

Figure S19. Euler angles of the RP phase

Figure S20. Euler angles of the RP phase

Figure S21. Euler angles of the RP phase

Figure S22. Euler angles of the RP phase
Finite-temperature polymorphs
Space group and tilt pattern for each of the polymorphs observed in the molecular dynamics simulations. Aleksandrov notation is used to denote the out-of-phase () and in-phase () octahedral tilts along each axis, with an overbar to indicate that the tilt patterns between adjacent slabs are in opposite directions[21]. The P/c phase corresponds to a more complex, layer-dependent tilt pattern (Figure S24) in which there is a superposition of both and tilts (Figure S10). This superposition is denoted with .
Figure S10. Mode projections in the heating run for the RP phase.
| Space group | Aleksandrov | |
Table S3. Space group and tilt pattern for each of the polymorphs observed in this study.
Octahedral tilt angles for to RP phases
A-site displacements
Rumpling refers to the displacement of A-site cations along the -direction. Here, we present the layer-resolved rumpling amplitudes for RP phases with to 6. Our analysis reveals that rumpling is most pronounced at the outermost layers, at the perovskite-rocksalt interfaces. The rumpling amplitudes are symmetric across the layers of the slab and remain the same between the two slabs in each RP phase.
As has been observed in previous publications, the competition between rumpling and out-of-plane octahedral tilting results from both distortions reducing the cation-anion distance (in this case between Ba and S) [22]. This is demonstrated in Table S4, where the distance between Ba and S atoms at the perovskite-rocksalt interface are provided for the untilted structures and the ground-state tilted structure. In both cases a rumpling distortion has also been applied.
Next, we look at A-site displacements in the -plane. Here, the displacements are computed with the high-symmetry structure as reference. For the A-site displacements alternate within a layer (normal along -axis) - see Figure S26, Figure S27, Figure S28. Additionally, one can clearly see how the A-site displacements in the and directions follows that of the out of phase tilting around the and axes. For example, in at 0 K we have a tilting pattern of , and A-sites in the first slab are only displaced in the x-direction whereas in second slab they are displaced in the y-direction. As the system transitions to we see both slabs exhibiting A-site displacements along both and direction. Finally, as the system transitions to untilted phase the A-site displacements disappear.

Figure S28. A-site displacements in the RP phase
In (and above) the in-plane A-site displacements within each layer are activated in and directions. This follows from the tilt pattern . We note that within slab I there is a net displacement of A-sites, however this cancels out as the net displacements in slab II is in the opposite direction. This is expected for the structure as it does not exhibit the improper ferroelectric behaviour seen in the structure.
X-ray diffraction
X-ray diffraction (XRD) patterns were obtained with the VESTA software using experimental values for the lattice parameters [23].
In Figure S30, we show excellent agreement with the experimentally reported phase at room temperature.
In Figure S31, we show the simulated XRD for three different phases of the RP material – the low temperature polymorph, the intermediate , and the high temperature polymorph. We also plot the experimentally reported data from Ref. 24 for comparison (where it is assigned to ).
We find that, as expected for such closely-related structures, the XRD patterns of all phases are highly similar. At room temperature our heating simulations predict the phase to be stable. This assignment is supported by peak shoulders at 25° and 36° in the experimental XRD. However, there are also a number of peaks missing in the experimental XRD at 18°, 32°, and 50°. XRD patterns can have broad and unresolved peaks for polycrystalline samples of BaZrS₃ [25]. In addition, we note that the peaks associated with the will gradually decrease in intensity with increasing temperature due to its predicted second-order transition to the phase at 400 K. The phase transition temperatures are sensitive to the small energy differences between polymorphs, and so although we can tentatively assign the experimental XRD pattern to the phase, we cannot rule out formation of the or phases at room temperature. Further XRD measurements on a single crystal sample, combined with complementary techniques such as Raman spectroscopy, could aid in a more definite determination of the crystal structure.
Supplemental References
[1] Erik Fransson, Julia Wiktor, and Paul Erhart. Phase transitions in inorganic halide perovskites from machine-learned potentials. J. Phys. Chem. C, 127:13773–13781, 2023. doi: 10.1021/acs.jpcc.3c02623.
[2] Zheyong Fan, Zezhu Zeng, Cunzhi Zhang, Yanzhou Wang, Keke Song, Haikuan Dong, Yue Chen, and Tapio Ala-Nissila. Neuroevolution machine learning potentials: Combining high accuracy and low cost in atomistic simulations and application to heat transport. Phys. Rev. B, 104:104309, 2021. doi: 10.1103/PhysRevB.104.104309.
[3] Zheyong Fan. Improving the accuracy of the neuroevolution machine learning potential for multicomponent systems. J. Phys.: Condens. Matter, 34:125902, 2022. doi: 10.1088/1361-648x/ac462b.
[4] Zheyong Fan, Yanzhou Wang, Penghua Ying, Keke Song, Junjie Wang, Yong Wang, Zezhu Zeng, Ke Xu, Eric Lindgren, J. Magnus Rahm, Alexander J. Gabourie, Jiahui Liu, Haikuan Dong, Jianyang Wu, Yue Chen, Zheng Zhong, Jian Sun, Paul Erhart, Yanjing Su, and Tapio Ala-Nissila. GPUMD: A package for constructing accurate machine-learned potentials and performing highly efficient atomistic simulations. J. Chem. Phys., 157:114801, 2022. doi: 10.1063/5.0106617.
[5] Ask Hjorth Larsen, Jens Jørgen Mortensen, Jakob Blomqvist, Ivano E. Castelli, Rune Christensen, Marcin Dułak, Jesper Friis, Michael N. Groves, Bjørk Hammer, Cory Hargus, Eric D. Hermes, Paul C. Jennings, Peter Bjerre Jensen, James Kermode, John R. Kitchin, Esben Leonhard Kolsbjerg, Joseph Kubal, Kristen Kaasbjerg, Steen Lysgaard, Jón Bergmann Maronsson, Tristan Maxson, Thomas Olsen, Lars Pastewka, Andrew Peterson, Carsten Rostgaard, Jakob Schiøtz, Ole Schütt, Mikkel Strange, Kristian S. Thygesen, Tejs Vegge, Lasse Vilhelmsen, Michael Walter, Zhenhua Zeng, and Karsten W. Jacobsen. The atomic simulation environment—a Python library for working with atoms. J. Phys.: Condens. Matter, 29:273002, 2017. doi: 10.1088/1361-648X/aa680e.
[6] Eric Lindgren, Magnus Rahm, Erik Fransson, Fredrik Eriksson, Nicklas Österbacka, Zheyong Fan, and Paul Erhart. calorine: A Python package for constructing and sampling neuroevolution potential models. J. Open Source Softw., 9:6264, 2024. doi: 10.21105/joss.06264.
[7] Fredrik Eriksson, Erik Fransson, and Paul Erhart. The Hiphive package for the extraction of high-order force constants by machine learning. Adv. Theory Simul., 2:1800184, 2019. doi: 10.1002/adts.201800184.
[8] Prakriti Kayastha, Erik Fransson, Paul Erhart, and Lucy Whalley. Octahedral tilt-driven phase transitions in BaZrS₃ chalcogenide perovskite. J. Phys. Chem. Lett., 16:2064–2071, 2025. doi: 10.1021/acs.jpclett.5c00040.
[9] John P. Perdew, Adrienn Ruzsinszky, Gábor I. Csonka, Oleg A. Vydrov, Gustavo E. Scuseria, Lucian A. Constantin, Xiaolan Zhou, and Kieron Burke. Restoring the density-gradient expansion for exchange in solids and surfaces. Phys. Rev. Lett., 100:136406, 2008. doi: 10.1103/PhysRevLett.100.136406.
[10] Aliaksandr V. Krukau, Oleg A. Vydrov, Artur F. Izmaylov, and Gustavo E. Scuseria. Influence of the exchange screening parameter on the performance of screened hybrid functionals. J. Chem. Phys., 125:224106, 2006. doi: 10.1063/1.2404663.
[11] Volker Blum, Ralf Gehrke, Felix Hanke, Paula Havu, Ville Havu, Xinguo Ren, Karsten Reuter, and Matthias Scheffler. Ab initio molecular simulations with numeric atom-centered orbitals. Comput. Phys. Commun., 180:2175–2196, 2009. doi: 10.1016/j.cpc.2009.06.022.
[12] Florian Knoop, Thomas Purcell, Matthias Scheffler, and Christian Carbogno. FHI-Vibes: Ab initio vibrational simulations. J. Open Source Softw., 5:2671, 2020. doi: 10.21105/joss.02671.
[13] Atsushi Togo. First-principles phonon calculations with phonopy and phono3py. J. Phys. Soc. Jpn., 92:012001, 2023. doi: 10.7566/JPSJ.92.012001.
References
- [1]Erik Fransson, Julia Wiktor, and Paul Erhart. Phase transitions in inorganic halide perovskites from machine-learned potentials. J. Phys. Chem. C, 127:13773–13781, 2023. doi: 10.1021/acs.jpcc.3c02623.
- [2]Zheyong Fan, Zezhu Zeng, Cunzhi Zhang, Yanzhou Wang, Keke Song, Haikuan Dong, Yue Chen, and Tapio Ala-Nissila. Neuroevolution machine learning potentials: Combining high accuracy and low cost in atomistic simulations and application to heat transport. Phys. Rev. B, 104:104309, 2021. doi: 10.1103/PhysRevB.104.104309.DOI
- [3]Zheyong Fan. Improving the accuracy of the neuroevolution machine learning potential for multi-component systems. J. Phys.: Condens. Matter, 34:125902, 2022. doi: 10.1088/1361-648x/ac462b.DOI
- [4]Zheyong Fan, Yanzhou Wang, Penghua Ying, Keke Song, Junjie Wang, Yong Wang, Zezhu Zeng, Ke Xu, Eric Lindgren, J. Magnus Rahm, Alexander J. Gabourie, Jiahui Liu, Haikuan Dong, Jianyang Wu, Yue Chen, Zheng Zhong, Jian Sun, Paul Erhart, Yanjing Su, and Tapio Ala-Nissila. GPUMD: A package for constructing accurate machine-learned potentials and performing highly efficient atomistic simulations. J. Chem. Phys., 157:114801, 2022. doi: 10.1063/5.01066617.DOI
- [5]Ask Hjorth Larsen, Jens Jørgen Mortensen, Jakob Blomqvist, Ivano E. Castelli, Rune Christensen, Marcin Dułak, Jesper Friis, Michael N. Groves, Bjørk Hammer, Cory Hargus, Eric D. Hermes, Paul C. Jennings, Peter Bjerre Jensen, James Kermode, John R. Kitchin, Esben Leonhard Kolsbjerg, Joseph Kubal, Kristen Kaasbjerg, Steen Lysgaard, Jón Bergmann Maronsson, Tristan Maxson, Thomas Olsen, Lars Pastewka, Andrew Peterson, Carsten Rostgaard, Jakob Schiøtz, Ole Schütt, Mikkel Strange, Kristian S. Thygesen, Tejs Vegge, Lasse Vilhelmsen, Michael Walter, Zhenhua Zeng, and Karsten W. Jacobsen. The atomic simulation environment—a Python library for working with atoms. J. Phys.: Condens. Matter, 29:273002, 2017. doi: 10.1088/1361-648X/aa680e.
- [6]Eric Lindgren, Magnus Rahm, Erik Fransson, Fredrik Eriksson, Nicklas Österbacka, Zheyong Fan, and Paul Erhart. calorine: A Python package for constructing and sampling neuroevolution potential models. J. Open Source Softw., 9:6264, 2024. doi: 10.21105/joss.06264.DOI
- [7]Fredrik Eriksson, Erik Fransson, and Paul Erhart. The Hiphive package for the extraction of high-order force constants by machine learning. Adv. Theory Simul., 2:1800184, 2019. doi: 10.1002/adts.201800184.
- [8]Prakriti Kayastha, Erik Fransson, Paul Erhart, and Lucy Whalley. Octahedral tilt-driven phase transitions in BaZrS₃ chalcogenide perovskite. J. Phys. Chem. Lett., 16:2064–2071, 2025. doi: 10.1021/acs.jpclett.5c00040.
- [9]John P. Perdew, Adrienn Ruzsinszky, Gábor I. Csonka, Oleg A. Vydrov, Gustavo E. Scuseria, Lucian A. Constantin, Xiaolan Zhou, and Kieron Burke. Restoring the density-gradient expansion for exchange in solids and surfaces. Phys. Rev. Lett., 100:136406, 2008. doi: 10.1103/PhysRevLett.100.136406.
- [10]Aliaksandr V. Krukau, Oleg A. Vydrov, Artur F. Izmaylov, and Gustavo E. Scuseria. Influence of the exchange screening parameter on the performance of screened hybrid functionals. J. Chem. Phys., 125:224106, 2006. doi: 10.1063/1.2404663.
- [11]Volker Blum, Ralf Gehrke, Felix Hanke, Paula Havu, Ville Havu, Xinguo Ren, Karsten Reuter, and Matthias Scheffler. Ab initio molecular simulations with numeric atom-centered orbitals. Comput. Phys. Commun., 180:2175–2196, 2009. doi: 10.1016/j.cpc.2009.06.022.
- [12]Florian Knoop, Thomas Purcell, Matthias Scheffler, and Christian Carbogno. FHI-Vibes: Ab initio vibrational simulations. J. Open Source Softw., 5:2671, 2020. doi: 10.21105/joss.02671.DOI
- [13]Atsushi Togo. First-principles phonon calculations with phonopy and phono3py. J. Phys. Soc. Jpn., 92:012001, 2023. doi: 10.7566/JPSJ.92.012001.
- [14]Zheyong Fan, Wei Chen, Ville Vierimaa, and Ari Harju. Efficient molecular dynamics simulations with many-body potentials on graphics processing units. Comput. Phys. Commun., 218:10–16, 2017. doi: 10.1016/j.cpc.2017.05.003.DOI
- [15]Erik Fransson, Julia Wiktor, and Paul Erhart. Impact of organic spacers and dimensionality on templating of halide perovskites. ACS Energy Lett., 9:3947–3954, 2024. doi: 10.1021/acsenergylett.4c01469.DOI
- [16]Erik Fransson, Petter Rosander, Fredrik Eriksson, J. Magnus Rahm, Terumasa Tadano, and Paul Erhart. Limits of the phonon quasi-particle picture at the cubic-to-tetragonal phase transition in halide perovskites. Commun. Phys., 6:173, 2023. doi: 10.1038/s42005-023-01297-8.DOI
- [17]Atsushi Togo, Kohei Shinohara, and Isao Tanaka. Spglib: a software library for crystal symmetry search. Sci. Technol. Adv. Mater. Methods, 4:2384822, 2024. doi: 10.1080/27660400.2024.2384822.DOI
- [18]Alexander Stukowski. Visualization and analysis of atomistic simulation data with OVITO—the open visualization tool. Modell. Simul. Mater. Sci. Eng., 18:015012, 2010. doi: 10.1088/0965-0393/18/1/015012.DOI
- [19]Yu Wu, Ying Chen, Zhenxing Fang, Yimin Ding, Qiaoqiao Li, Kui Xue, Hezhu Shao, Hao Zhang, and Liujiang Zhou. Ultralow lattice thermal transport and considerable wave-like phonon tunneling in chalcogenide perovskite BaZrS₃. J. Phys. Chem. Lett., 14:11465–11473, 2023. doi: 10.1021/acs.jpclett.3c03041.
- [20]Jochen Heyd, Gustavo E. Scuseria, and Matthias Ernzerhof. Hybrid functionals based on a screened Coulomb potential. J. Chem. Phys., 118:8207–8215, 2003. doi: 10.1063/1.1564060.DOI
- [21]K. S. Aleksandrov, B. V. Beznosikov, and S. V. Misyul. Successive phase transitions in crystals of K₂MgF₄-type structure. Phys. Status Solidi A, 104:529–543, 1987. doi: 10.1002/pssa.2211040204.DOI
- [22]Yajun Zhang, Jie Wang, and Philippe Ghosez. Unraveling the suppression of oxygen octahedra rotations in A₃B₂O₇ Ruddlesden–Popper compounds: Engineering multiferroicity and beyond. Phys. Rev. Lett., 125:157601, 2020. doi: 10.1103/PhysRevLett.125.157601.DOI
- [23]Koichi Momma and Fujio Izumi. VESTA 3 for three-dimensional visualization of crystal, volumetric and morphology data. J. Appl. Crystallogr., 44:1272–1276, 2011. doi: 10.1107/S0021889811038970.DOI
- [24]Shanyuan Niu, Boyang Zhao, Kevin Ye, Elisabeth Bianco, Jieyang Zhou, Michael E. McConney, Charles Setten, Ralf Haiges, Rafael Jaramillo, and Jayakanth Ravichandran. Crystal growth and structural analysis of perovskite chalcogenide BaZrS₃ and Ruddlesden–Popper phase Ba₃Zr₂S₇. J. Mater. Res., 34:3819–3826, 2019. doi: 10.1557/jmr.2019.328.
- [25]Laura-Bianca Pasca, Yuanbin Liu, Andy S. Anker, Ludmilla Steier, and Volker L. Deringer. Machine-learning-driven modelling of amorphous and polycrystalline BaZrS₃. J. Mater. Chem. A, 13:35447–35454, 2025. doi: 10.1039/D5TA04536C.DOI
- [26]Masanobu Saeki, Yoshiyuki Yajima, and Mitsuko Onoda. Preparation and crystal structures of new barium zirconium sulfides, Ba₂ZrS₄ and Ba₃Zr₂S₇. J. Solid State Chem., 92:286–294, 1991. doi: 10.1016/0022-4596(91)90337-I.DOI
- [27]Shanyuan Niu, Debarghya Sarkar, Kristopher Williams, Yucheng Zhou, Yuwei Li, Elisabeth Bianco, Huaixun Huyan, Stephen B. Cronin, Michael E. McConney, Ralf Haiges, Rafael Jaramillo, Douglas J. Law, William M. Kline, Jayakanth Ravichandran, Rafael Jaramillo, and Ralf Haiges. Optimal bandgap in a 2D Ruddlesden–Popper perovskite chalcogenide for single-junction solar cells. Chem. Mater., 30:4882–4886, 2018. doi: 10.1021/acs.chemmater.8b01656.