WorksPhysics and Astronomy
Halide diffusion in mixed-halide perovskites and heterojunctions
Abstract
Migration of halide defects guides ion transport in metal halide perovskites and controls the kinetics of halide mixing and phase separation. We study the diffusion of halide vacancies and interstitials in \ce{CsPb(I_{x}Br_{1-x})_{3}} and \ce{CsPbI_{3}}/\ce{CsPbBr_{3}} heterojunctions by molecular dynamics simulations using neural network potentials trained on density functional theory calculations. We observe enhanced diffusion of both vacancies and interstitials in the mixed halide compounds compared to the single halide ones, as well as a difference in mobility between Br and I ions in the mixed compound. Diffusion across heterojunctions is governed by the interface structure, where a Br-rich interface blocks migration of vacancies in particular, but an I-rich interface is permeable.
main.pdf
Introduction
Metal halide perovskites have gained traction in optoelectronic applications such as photovoltaics and light-emitting diodes.[1, 2, 3] One of their strong points is chemical flexibility, where the perovskite structure allows for mixing different organic or inorganic cations (A), metal cations (B), and halide anions (X). Through compositional alloying, the physical properties such as band gaps and band offsets can be tuned, and at the same time this stabilizes the perovskite structure.[4, 5, 6] For instance, mixing Br into lowers the temperature at which it transitions to the (black) phase, and is less prone to degrade into the photo-inactive (yellow) phase,[7] thereby improving the phase stability of this material. The currently most stable metal halide perovskites comprise a mix of X halide anions, as well as organic/inorganic A cations.[8, 9]
Although these alloys are stable under dark conditions, continuous light exposure commonly leads to demixing of I and Br ions[10, 11, 12] and consequently the formation of low band gap I-rich domains and high band gap Br-rich domains.[13] In previous work we presented a thermodynamic theory for light-induced halide segregation, based on the free energy gain of photocarriers assembled in domains with a halide composition that has a lower band gap than the starting mixed phase.[14, 15] The growth of such domains is then stimulated by this thermodynamic driving force. While the origin of halide phase segregation is thermodynamic, the kinetics of this process are poorly understood, and a fundamental strategy for blocking or slowing down the kinetics is missing.
Microscopically, the most likely candidate for mass transport in these materials is the migration of halide defects.[16] Halide perovskites are relatively soft materials and have a substantial concentration of defects.[17, 18, 19] Halide point defects are generally considered the dominant mobile species.[20] The relative importance of vacancies versus interstitials seems to depend somewhat on the particular operational conditions and techniques used in different experimental studies, and may also reflect the different conditions under which the materials are synthesized.21–25
From a modeling perspective, the most straightforward calculations make use of density functional theory (DFT) calculations on static structures along predicted migration paths to obtain diffusion barriers. 21,26–28 Such migration paths may however be difficult to find. In addition, they may also not be sufficiently representative, as such modeling relies on one or a small number of transition states being adequate for describing diffusion. Irrespective of that point, halide perovskites display different phases as a function of temperature, so DFT calculations on K structures may not describe diffusion at room or higher temperatures very well.
Molecular dynamics (MD) is a technique to simulate mass transport at the atomistic level directly, which also takes into account finite-temperature effects. Until recently, MD simulations have yielded limited information on diffusion in halide perovskites, as empirical model force fields tend to be too crude, 29 and ab initio MD simulations are computationally too expensive to reach the required time and length scales for ion migration. Machine learned force fields (MLFFs) have advanced the capabilities of MD simulations, by combining a near ab initio accuracy with a computational cost only an order of magnitude above empirical models. 30–32 Moreover, by carefully selecting the ab initio reference data for training, MLFFs can describe the dynamics of a material at various electronic states. 33
In recent work we utilized atomic descriptor-based machine learned force fields 34–37 to study the migration of different charged states of iodide interstitials and vacancies in . 38 We found that negative iodide interstitials and positive iodide vacancies, the most stable charge states for their respective defect type, migrate at similar rates at room temperature. In contrast, oppositely charged defects, i.e. positive iodide interstitials and negative iodide vacancies, migrate at much slower rates.
In the current work, we focus on the ion migration in mixed compounds, where diffusion of point defects may involve both I and Br interstitials, as well as vacancies, and can depend on the local I versus Br environment. 27 We perform ns-long molecular dynamics (MD) simulations using MLFFs to quantify the migration behavior of halide defects in mixed . In particular, we focus on halide diffusion through interfaces between and layers and study the initial phases of halide mixing. To conclude, we examine the stability of a nanodomain embedded in a matrix and vice versa with respect to halide mixing at the interfaces induced by defect migration. 27
Methods
MLFFs based on Gaussian approximation potentials (GAPs) 35,36 with smooth overlap of atomic position (SOAP) descriptors 34,37 can accurately capture the diffusion of point defects, as we demonstrated in previous work on . 38 Likewise, they are suited to describe defect migration in , as shown in SI note 1. However, they suffer from a poor scaling with the number of different chemical species, such as occurs in mixed compounds.
Although, message passing neural network (MPNN) potentials 39,40 overcome this shortcoming, global message passing gives them a poor computational scaling with system size. The Allegro neural network architecture 41 keeps the message passing local and is -equivariant (e3nn), 42 which restrains the scaling with size and complexity of the system, see the SI note 2. As a result of this, an efficient MLFF for long-time scale MD simulations on large systems. Here, we train an Allegro model, as implemented in the NequIP package, 40 to generate a neural network potential (NNP) for halide interstitial and vacancy defects in .
The datasets for training and validating the network are generated using the on-the-fly module as implemented in VASP, 35,43 where structures are sampled from 100 ps long MD runs. Energies, forces, and stress tensors are obtained from density functional theory (DFT) calculations with the PBE+D3-BJ exchange-correlation functional, 44,45 for a set of structures selected through Bayesian inference, to ensure sufficient variance in the local environments. The training and validation sets are generated using cubic supercells of mixed with I and Br positioned randomly on halide sites, and one halide point defect for a negatively charged halide interstitial (), or a positively charged halide vacancy (). Details of the training procedure are explained in the SI note 3.
The accuracy of the NNP energies can be explored, for instance, by comparing migration barriers along paths generated using the climbing image nudged elastic band (CI-NEB) technique46 with those calculated by DFT along the same paths. The computed energies for vacancies and interstitials in mixed systems with I and Br positioned randomly on halide sites are given in Figure 1a-Figure 1d. The difference between the migration barriers calculated using DFT and the NNP is less than 0.08 eV in all cases, indicating the NNP accurately probes the energy landscape for defect migration. The details of the CI-NEB calculations and calculated migration barriers are given in the SI note 4.
Figure 1. Energies along CI-NEB migration paths for I-mediated interstitial (a) and vacancy (b), and for Br-mediated interstitial (c) and vacancy (d) in from DFT and NNP calculations. Zero is set at the energy minima.
The NNP model is used to perform ns-long MD simulations in LAMMPS47 on cubic supercells of bulk with one halide point defect, interstitial or vacancy.
To test the accuracy of the forces generated by the NNP, we sample structures from MD simulations at 600 K, and calculate the forces by DFT. Comparing these to the NNP forces for the system for which the NNP was developed, i.e., mixed CsPb(IBr) with I and Br positioned randomly on halide sites, gives and a mean absolute error (MAE) meV/Å for forces acting on all atoms, and and MAE meV/Å for forces acting on atoms in the defect environment, see the SI note 4. Using the same NNP in MD simulations on the pure compounds, CsPbI or CsPbBr, gives a very similar accuracy in the forces, see SI figure S11, which is a strong indication of the transferability of the NNP among CsPb(IBr) compounds for .
Besides studying the diffusion of halide defects in CsPb(IBr), the NNP force field is also used to investigate migration of halide defects across interfaces between mono-halide layers, as well as halide mixing starting from mono-halide cubic nanodomains. The supercell size, temperature range, number of runs, and simulation time per run for all these systems are given in Table 1. The volume is kept constant during these runs, where the lattice constants for different halide concentrations in CsPb(IBr) are extracted from MD runs at constant temperature and pressure, see the SI note 5.
| Model | Supercell | Temperature (K) | time (ns) | number of runs |
| Mixed halide | 2 | 5 | ||
| Interfaces | 600 | 2 | 5 | |
| Nanodomains | 600 | 10 | 1 |
Table 1. The cubic supercell size, temperature range, simulation time per MD run, and number of runs used for the different systems in this study.
Results
Diffusion in CsPb(IBr)
Migration behavior is studied in mixed CsPb(IBr) with I and Br positioned randomly on halide sites with one halide point defect, vacancy or interstitial, added. We choose vacancy V or interstitial I, X = I or Br, in their most stable charge states, as they occur under intrinsic or moderately doped conditions.[18] To ensure sufficient migration events for statistics, MD runs are performed at elevated temperatures in the range 500-600 K. The details of these production runs are given in the SI note 6. To compare with the pure systems, we also perform MD simulations for the pure systems and . For the data obtained for , see Ref. 38. The training and production run details for and the diffusion behavior of other charge states of halide point defects are reported in SI note 1.
In characterizing halide defect diffusion in these systems it makes no sense to trace a specific type of halide ion, as migration typically proceeds via kick-out processes. Therefore, we use both halide species to extract the diffusion coefficient using the Einstein relation, which links the slope of the mean square displacement to time
where is the number of atoms of a particular atomic species in the simulation box, and are the atomic positions.
In the following we will label diffusion coefficients by interstitial or vacancy, according to the starting situation of each MD run. In the temperature and time ranges studied one observes multiple sequential kick-out events, so the actual species that is migrating can change over the trajectory. However, when starting with a single vacancy (interstitial), then one vacancy (interstitial) remains present; no additional long-lived defects are created.
Diffusion coefficients are given in Figure 2. They are extracted by applying Eq. (1) to the runs for the mixed-halide compound and similar systems created for the single-halide compounds. Comparing and one observes that diffusion coefficients of Br and I interstitial, and , in the temperature range 500-600 K are very similar, i.e., within 30% or so, with that of the I interstitial higher. The difference between the diffusion coefficients of Br and I vacancies, and , is slightly larger, approximately a factor of two, with that of the I vacancy being the larger of the two. In the mixed system the two halide species can change positions dynamically. As a first analysis we look at the migration of the halide species X without distinguishing between Br and I. The diffusion coefficients for X interstitials and vacancies are remarkably close to those found in the pure systems, but slightly closer to the values found in , in particular for the vacancy.
Figure 2. Temperature-dependent diffusion coefficients of halide interstitials () and vacancies () in pure (I), mixed with I and Br positioned randomly on halide sites (X), and pure (Br). The filled square and open circle symbols represent halide interstitials and vacancies, respectively. The dashed lines represent fits to an Arrhenius expression, and the error bars represent the standard error in mean at each point.
The diffusion coefficients can be fitted with an Arrhenius expression
The fits are given in Figure 2, and the extracted activation energies () and pre-exponential factors () are given in Table 2. The error bars reflect the quality of the linear regression fit.
| System | Defect | (eV) | () | () |
Table 2. Activation energies () and pre-exponential factors () extracted from the Arrhenius fits, and extrapolated diffusion constants () at room temperature for halide interstitial () and vacancy () defects in pure (I), mixed with I and Br positioned randomly on halide sites (X), and pure (Br).
The activation energies are generally quite low and close to one another across all compositions, lying between and . While I ions are larger than the Br ions, the structures of and scale accordingly, making the diffusion of I interstitials in similar to that of Br interstitials in , and the same holds for halide vacancy diffusion in both materials. The barriers for interstitial diffusion are somewhat higher than those for vacancy diffusion, but they are actually remarkably close. The diffusion barriers in the mixed halide are somewhat lower than in the pure compounds. In addition, there is a slight difference in the migration behavior of I and Br defects in the mixed halide. We will come to this point below.
Extrapolating the Arrhenius expression to room temperature, , one obtains values for the diffusion coefficients in the range –, see Table 2. The extrapolated diffusion coefficients of vacancies and interstitials in the mixed halide are approximately a factor of two larger than those in the pure compounds. In a similar fashion, the diffusion coefficients of vacancies are approximately a factor of two larger than those of interstitials.
In principle, Eq. (1) also allows for decomposing the migration behavior of halide interstitials and vacancies in the down to halide species, i.e., distinguishing between Br and I. Diffusion coefficients for each species are given in Figure 3a. As it turns out, diffusion does not depend on the species, Br or I, initially chosen to represent interstitial of vacancy. As diffusion proceeds via kick-out processes, after a few of those events, the dominant species takes over. Regarding interstitials the Br species migrates faster than the I species, irrespective of the starting species of interstitial. Remarkably, for vacancies it is the other way around, i.e., I vacancies migrate faster than Br vacancies.
Figure 3. a) Halide species decomposed temperature-dependent diffusion coefficients of halide interstitials and vacancies in CsPb(IBr). Blue is for I, and red is for Br. The filled square symbols represent halide interstitials, and the open circle symbols represent halide vacancies. The dashed lines represent the fits to an Arrhenius expression, and the error bars represent the standard error in mean at each point. Schematic representation of the diffusion paths for (b) halide interstitial and (c) halide vacancy defects in CsPb(IBr). The arrows represent the migration directions of the ions.
Interstitials and vacancies have opposite charge, so in an electric field they move in opposite directions. As the negatively charged interstitials will be driven to the anode, it will lead to a surplus of Br there. At the same time, the positively charged vacancies migrate to the cathode, which promotes a net flux of I in the direction of the anode. In an electric field the two processes, interstitial- and vacancy-mediated diffusion, thus tend to compensate one another as a driving force for halide separation.
Arrhenius fits, Eq. (2), are shown in Figure 3a, with the parameters and the extrapolation to room temperature given in Table 3. These results confirm the trends discussed in the previous paragraph. Halide interstitials migrate an order of magnitude faster through Br than through I. Meanwhile, halide vacancies migrate faster through I than through Br.
| System | (eV) | ( cms) | ( cms) |
| (I) | 0.22 0.06 | 0.43 0.51 | 0.87 |
| (Br) | 0.15 0.04 | 0.29 0.26 | 8.8 |
| (I) | 0.11 0.02 | 0.08 0.05 | 11.4 |
| (Br) | 0.18 0.04 | 0.24 0.19 | 2.3 |
Table 3. Activation energies () and pre-exponential factors () extracted from the Arrhenius fits to diffusion coefficients decomposed according to halide species in CsPb(IBr) and diffusion constants () extrapolated to room temperature.
Focusing on the dominant migration mechanisms, migration of Br interstitials and I vacancies is roughly equally fast.
To rationalize halide diffusion in mixed CsPb(IBr), we examine the defect geometries along the migration paths. The most stable configuration of a halide interstitial is the bridge structure shown schematically in Figure 3b. [17] [48] In CsPb(IBr), a Pb–BrBr–Pb bridge configuration is more frequently observed than a Pb–IBr–Pb bridge configuration, both of which are observed significantly more than Pb–II–Pb bridge configuration, see the SI note 7. We conclude that the more prevalent a configuration is, the more stable it is. On the basis of occupation numbers, SI Figure S16, and assuming that these are determined by Boltzmann factors, we estimate an energy difference between the Pb–IBr–Pb bridge and the Pb–BrBr–Pb bridge of 34.5 meV, and between the Pb–II–Pb bridge and the Pb–BrBr–Pb bridge of 143.9 meV at 600 K. As I is much less likely to be found in an interstitial position than Br, the latter is the main species that causes diffusion of interstitial defects, as also shown in Figure 3.
By contrast, the halide vacancy has a preference for diffusion by I. The Pb–Br bond is stronger than a Pb–I bond, suggesting that the latter can be more easily broken to let an I fill the vacancy, Figure 3c. Consistent with that notion, the I ions fluctuate more around their mean position, indicating a flatter energy landscape, than Br ions (SI note 8). Therefore, I ions would jump more easily to fill a vacancy, and hence form the dominant species for diffusion via vacancies.
A result regularly observed in simple materials is that compressive strain hinders interstitial diffusion, and tensile strain promotes it, whereas, in contrast, compressive strain promotes vacancy diffusion, and tensile strain hinders it. 49,50 The general reasoning is that compression stabilizes the vacancy and destabilizes the interstitial, decreasing the diffusion barriers of the former and increasing those of the latter, whereas tension has the opposite effects. There are many exceptions to be found, but diffusion in follows this rule. The lattice parameter of is in between that of the pure compounds and . From the perspective of Br, is under tensile strain, whereas from the perspective of I it is under compressive strain. Hence, both the diffusion of Br interstitials and that of I vacancies in should be promoted, which indeed it does considering Table 2 and Table 3.
Our results are in qualitative agreement with those obtained in experiments by McGovern et al. on mixed perovskites. 51 Compared to the pure I or Br perovskites, halide migration in mixed-halide perovskites has a lower activation energy and a larger diffusion coefficient, cf. Table 2. In addition, in the mixed-halide perovskite one finds two sets of diffusion coefficients and activation energies, describing the migration of the Br and I ions respectively, cf. Table 3. Full quantitative agreement with experiments on cannot be expected, as the A cation (MA methylammonium versus Cs) is likely to have some influence on the ion migration behavior.
Diffusion across interfaces
Besides where the Br and I are positioned randomly on halide sites on an atomic scale, it is interesting to look at halide migration in materials comprising larger domains of pure and . This occurs, for instance, when a material in a phase-separated state is kept in a dark environment. [14] The simplest model geometry of such a state consists of layers of the pure compounds, as illustrated in Figure 4a, with planar interfaces along the direction. At the interfaces between the pure I- and Br-perovskites, Pb atoms are coordinated partially by I and partially by Br atoms.
Figure 4. (a) Bulk interface and layers away from the interface in cubic supercell of , with 2 Br-rich interfaces along the periodic boundaries. Defect occupancy at different layers () of (b) halide vacancy for I-rich interfaces, (c) halide vacancies for Br-rich interfaces, (d) halide interstitials for I-rich interfaces, and (e) halide interstitials for Br-rich interfaces.
We consider the two extreme cases, an I-rich interface where the Pb atoms are coordinated by five I atoms and one Br atom, and a Br-rich interface with Pb atoms coordinated by five Br atoms and one I atom (Figure 4a). The calculated difference between the static formation energies of these interfaces is very small, 2.5 meV per Pb atom at the interface, see the SI note 9. As this is much smaller than at room temperature, we conclude that both these interfaces are equally likely to form under equilibrium conditions.
Our model system consists of alternating layers of and , each layer eight sublayers deep, where the interfaces are either all I-rich or all Br-rich, Figure 4a. We use an in-plane supercell, and start with a defect, vacancy or interstitial, positioned at a specific interface. The supercell dimensions are obtained by the procedure described in the SI note 9. As we have two different interfaces and two different defects, this gives us four different cases in total. For each case we run a 2 ns MD simulation at K, and repeat this five times for each system, while monitoring the position of the defect. As a result, Figure 4b-Figure 4e shows the defect occupancies of the individual lead halide sublayers accumulated over all MD runs.
Starting with the vacancy, it clearly has a preference for residing in the layer, Figure 4b and Figure 4c. One could map occupancies onto an energy landscape , and the sublayer index from a Boltzmann expression , K, between the two sublayers near the center of the and layers one then finds an energy difference meV.
Comparing I-rich and Br-rich interfaces, there is a remarkable difference between of finding a vacancy there. Whereas the I-rich interface easily incorporates a vacancy, leading to an occupancy comparable to that of the layer, the Br-rich interface actually repels vacancies. Again, mapping onto an energy gives an interface barrier meV. The consequence is that I-rich interfaces are very permeable for the diffusion of vacancies, whereas Br-rich interfaces hinder the diffusion.
The interstitial gives a more complex picture. As interstitials diffuse by means of kick-out
processes, they can change character, i.e., a Br interstitial can become an I interstitial in the next diffusion step and vice versa. The effects of these kick-out processes are visible in Figure 4d and Figure 4e, where the occupancies of Pb–II–Pb and Pb–BrBr–Pb bridges are high in the and layers, respectively, but there is no crossover of these bridges into the other layer.
For the Br-rich interface, mixed Pb–IBr–Pb bridges are found at or very near the interface, with negligible occupancy further from the interface. Interestingly, also in this case the I-rich interface behaves differently. It is more permeable for I interstitials, allowing for diffusion of I into the layer, as documented by a substantial occupancy of Pb–IBr–Pb bridges there. The general conclusion then is that Br-rich interfaces hinder inter diffusion of halide species, whereas I-rich interfaces do not.
To further demonstrate the influence of defects on halide phase mixing in , we build models consisting of a cubic nanodomain of inside a matrix (Figure 5a,Figure 5b), and vice versa (Figure 5c,Figure 5d). These model structures have six I-rich interfaces for I cubic nanodomains, and six Br-rich interfaces for Br cubic nanodomains. Starting with one halide point defect in the nanodomain, we perform 10 ns long MD runs at K. The details of these MD runs are given in SI note 10. The final frames from the MD runs are analyzed for the number of halide atoms of the cubic nanodomain halide intermixed with one in the rest of the lattice.

Figure 5. Final frames from the MD simulations along all three axes for (a) halide interstitial and (b) vacancy in Br cubic domain, and (c) halide interstitial and (d) vacancy in I cubic domain. The non-domain halide species atoms are omitted for the sake of clarity.
Starting with a vacancy in the nanodomain (Figure 5b), no mixing events are observed in the 10 ns time frame. This is consistent with the interface studies discussed above, where the vacancy prefers to reside in the phase, and Br-rich interfaces hinder the diffusion. Starting with an interstitial in a nanodomain (Figure 5a), 11 mixing events are observed, which is compatible with the fact that Br-rich interfaces are more permeable for interstitials than for vacancies.
In contrast, the nanodomain is much more fragile, with 36 mixing events for the halide vacancy in the 10 ns time frame (Figure 5d), which is in line with the vacancy
preferring the phase, and I-rich interfaces being very permeable. For the halide interstitial 29 mixing events are observed (Figure 5b), stressing that the fragility of the nanodomain does not depend too much on whether diffusion is executed by halide vacancies or interstitials.
Summary and Conclusions
In summary, we study the diffusion of halide vacancies and interstitials in and heterojunctions. We trained a neural network potential for negatively charged halide interstitial and positively charged halide vacancy defects in using DFT calculations. Using this force field, we performed ns-long MD simulations to study the migration behavior of these defects and their influence on halide mixing through interfaces and domains.
We observe enhanced diffusion of both vacancies and interstitials by a factor 2-3 at room temperature in the mixed halide compounds as compared to the single halide ones, where vacancies diffuse faster than interstitials by a factor of two. Our findings furthermore suggest that the preferred migrating halide species in is different for interstitial and vacancy. At room temperature, bromide migrates an order of magnitude faster than iodide for interstitials, while iodide migrates a factor of five faster than bromide for vacancies.
These results are in qualitative agreement with experimental results on mixed , which displays enhanced halide migration compared to and . 51 In addition, in the mixed-halide perovskite two diffusion coefficients are found, which we suggest could describe the migration of the Br and I ions respectively.
Diffusion across heterojunctions is governed by the interface structure, where a Br-rich interface blocks migration of vacancies in particular, and to a lesser extent for interstitials. An I-rich interface is much more permeable, both for interstitials, as well as for vacancies, and therefore facilitates diffusion and intermixing of halides. The latter is demonstrated by MD simulations on nanodomains in a matrix and vice versa.
Acknowledgement
The authors thank Guus Bertens for his help with the compilation of LAMMPS with the pair_allegro patch. V.T. and S.T. acknowledge funding from Vidi (project no. VI.Vid.213.091) from the Dutch Research Council (NWO).
SI.pdf
Supporting Information:
Defect migration in CsPbBr
Model training
We start by training different MLFFs for all three negative, neutral, and positive charge states of halide interstitial and halide vacancy defects in CsPbBr. Variants of the GAP-SOAP model[1, 2] are trained as implemented in VASP, [3, 4, 5] where the training structures are sampled from three MD runs. These runs are performed at different temperatures for each system to ensure that the force fields can describe diverse atomic environments during the production runs. Temperatures for these runs are given in Table S1. These runs are performed using cubic supercells ( units of CsPbBr) with one halide point defect for ps with MD time-steps of fs. Most of these runs are performed in ensembles at Pa pressure. To maintain constant temperature and pressure, Parinello-Rahman dynamics is used.[6, 7] For , unphysically large volume expansions are observed during the first training run ( K), possibly due to excess electronic charge in a relatively small training structure. Hence, the volume is kept constant during this run, while the shape of the cell is allowed to change.
| System | Step 1 | Step 2 | Step 3 |
| 700 K | 750 K | 600 K | |
| 700 K | 750 K | 600 K | |
| 700 K | 750 K | 600 K | |
| 700 K* | 750 K | 750 K | |
| 750 K | 750 K | 800 K | |
| 700 K | 750 K | 700 K |
Table S1. Temperatures at which training was performed for each defect system. Here * corresponds to the runs for which the cell volume was kept constant.
During these training runs, the frames are evaluated using density functional theory (DFT) calculations whenever the estimation error in forces exceeds a threshold value. For these DFT calculations, the projector-augmented wave (PAW)[8] technique is used to model the electron-ion interaction, with the outermost electrons of Br (), Cs (), and Pb () treated as valence electrons. The electronic interactions are modeled using the PBE exchange-correlation functional within the generalized gradient approximation (GGA) supplemented with D3-BJ Vanderwaals contributions.[9, 10] Energy convergence criterion of eV is used for the Self-consistent field cycles. The calculations are performed using Monkhorst-Pack k-grid[11] and a kinetic energy cutoff of eV.
To describe the local chemical environments, a combination of two descriptors similar to the smooth overlap of atomic positions (SOAP) descriptor is used.[1] The first of these is a descriptor closely resembling a radial distribution function that can also be regarded as a two-body descriptor, for which a cutoff of is used. The second descriptor is purely angular, containing no two-body components, for which a cutoff of is used.[12]
The atomic positions are expanded using Gaussian distributions of width . The radial descriptor is expanded using 12 radial basis functions, whereas the angular descriptor is expanded using 8 radial basis functions and spherical harmonics with a maximum angular momentum quantum number of 6. As the structures are sampled, a set of local reference configurations is constructed. The maximum size of this set is 4500 for all systems. The local potential energy of an atom in a structure from the training set is expressed as a linear combination of Gaussian kernels that measure the similarity between the local reference configuration from the training set and the basis set. A polynomial power of 4 is used for these kernels, and the weight of the radial descriptor in these kernels is set to 0.1, setting the weight of the angular descriptor to 0.9.
The forces and energies predicted by the current force field can be compared with forces and energies calculated using DFT during training. The root mean squared difference between these forces over training time is given in Figure S1. These errors plateau over time for all systems, signifying that the force field accurately captures the potential energy surface of defective perovskite systems, making it suitable for studying their dynamics. Furthermore, the energies were calculated for all training structures using the final force fields and compared with DFT-calculated energies. These comparisons are given in Figure S2. The root mean squared errors from these comparisons () are in the order 1 meV/atom, as the energies of the training structures are of the order of 1 eV/atom. These values indicate that the force fields are accurately learning the energies during training.
Figure S1. Root mean squared error in forces () calculated by the force field compared to DFT calculated forces during training of all charge states of bromide interstitial and vacancy defects in CsPbBr.
Figure S2. Comparison between energies calculated using DFT with energies calculated using the final MLFF for all structures in the training set of all defect systems of , along with the corresponding root mean squared errors.
Finally, all the force fields are refitted onto faster descriptors without any error estimate to speed up their evaluations for use in production runs. The number of structures in the training sets () and the number of local reference configurations for the different elements in the MLFFs () are given in Table S2.
| System | ||||
| 1328 | 1723 | 1764 | 4129 | |
| 2057 | 1449 | 1903 | 4002 | |
| 2370 | 1561 | 1841 | 4500 | |
| 3317 | 787 | 2986 | 4500 | |
| 1965 | 2039 | 2167 | 4500 | |
| 2124 | 1615 | 2202 | 4500 |
Table S2. Size of the training set and number of local reference configurations in the MLFFs for each defect system.
Model validation
To test the accuracy of these models, we compare the energies predicted by the force fields with those calculated using DFT for migration paths as they would enter transition state theory (TST). These paths are constructed using climbing image nudged elastic band (CI-NEB) calculations on orthorhombic supercells (16 units of ) with one point defect. Three intermediate images, connected by springs with spring constant are optimized using kinetic energy cut-off of 300 eV and Monkhorst-Pack -point grid. The computed energies are given in Figure S3, and the migration barriers are given in Table S3. The DFT calculated trends in migration barriers are well captured by the MLFFs (Table S3). Further, the difference between the migration barriers calculated using the MLFFs and DFT are within 0.11 eV for all defects except , where the difference is 0.26 eV (Figure S3). This could be due to only a few migration events being sampled while training this MLFF.
Figure S3. Energies along the NEB migration paths for different charge states of the (a-c) Br interstitial and (d-f) Br vacancy defects in , calculated using DFT and the MLFFs. The energy minima are set to 0.
| System | (eV) | (eV) |
| 0.21 | 0.19 | |
| 0.20 | 0.16 | |
| 0.38 | 0.37 | |
| 1.32 | 1.06 | |
| 0.58 | 0.47 | |
| 0.38 | 0.38 |
Table S3. Defect migration barriers () calculated using DFT/TST/CI-NEB migration paths.
To test the accuracy of the MLFFs during the MD production runs, we sample 20 structures from 0.5 ns MD runs at 600 K performed using cubic supercells (216 units of CsPbBr with one bromide point defect, and compare forces calculated using DFT with forces calculated using the force fields for these structures. We compare forces acting on all atoms (Figure S4) and also, in particular, the forces of atoms close to the defect (Figure S5) to check the accuracy with which the force fields could describe the defect environments. The procedure for identifying atoms in the defect environment is explained in the SI note 7 of our previous work. [13] From these comparisons, we note that the and MAE meV/\AA{} for forces acting on all atoms, and and MAE meV/\AA{} for forces acting on atoms in the defect environment. As the forces acting on the atoms are of the order of 1 eV/\AA{}, all force fields are transferable to bigger system sizes and accurate in capturing the defect environments.
Figure S4. Comparison between forces acting on all atoms calculated by the MLFFs with forces calculated using DFT for different charge states of halide interstitial (a-c) and halide vacancy (d-f) defects in supercells of CsPbBr, along with corresponding values and mean absolute errors (MAE).
Figure S5. Comparison between forces acting on atoms in the defect environment calculated by the MLFFs with forces calculated using DFT for different charge states of halide interstitial (a-c) and halide vacancy (d-f) defects in supercells of , along with corresponding values and mean absolute errors (MAE).
Phase transition model
To avoid the influence of volume expansion on diffusion, all equilibration and production runs are performed at constant volume. The cell volumes at different temperatures are deduced from constant temperature MD runs using a phase transition model. The model is trained using (8 units of ) pseudo-cubic cells in all three perovskite phases of . We start the training in the cubic phase at 500 K, followed by the tetragonal phase at 370 K, and finally in the orthorhombic phase at 370 K and 150 K. All these runs are performed at Pa pressure for 100 ps with MD time-steps of 10 fs. For the DFT calculations, we use the same functional, PAW potentials, and kinetic energy cutoff as in § 7.1.1, energy convergence criteria of eV, and a Monkhorst-Pack -point grid. To construct the MLFF, cutoff radii of 6 Å and 5 Å are used for the radial and the angular descriptors, respectively. The atomic positions are expanded using Gaussian distributions of width 0.5 Å. The radial descriptor is expanded using 6 radial basis functions, and the angular descriptor is expanded using 6 radial basis functions and spherical harmonics with a maximum angular momentum quantum number of 9. Gaussian kernels of polynomial power 4 are used, where the radial and angular descriptors have weights 0.7 and 0.3, respectively.
The accuracy of this model is validated by performing a heating MD run from 100 K to 500 K for 0.8 ns with an MD time step of 10 fs using a (72 units of ) orthorhombic supercell. The window averaged pseudo-cubic lattice vectors with the window length 0.16 ns from this run are given in Figure S6, and the overall phase diagram closely resembles that observed in experiments.[14]

Figure S6. Pseudo-cubic unit cell lattice vectors of CsPbBr as function of temperature in the heating MD run.
Using this force field, we perform 100 ps long constant temperature MD runs using cubic (216 units of ) supercells in the ensemble at various temperatures between 400 K and 500 K. The variation in the cube root of the final volume with temperature from these runs is extrapolated to get lattice constants at various temperatures between 500 K and 600 K.
Production runs and diffusion coefficients
Following training and validation, the force fields are used to perform multiple MD runs on cubic supercells ( units of ) with one bromide point defect at temperatures between K and K. The volume is kept constant during each run with lattice parameters at different temperatures extracted from the constant temperature MD runs on pristine . The lattice constants at different temperatures are given in Table S4.
| Temperature (K) | Lattice constants (Å) |
| 500 | 5.939 |
| 525 | 5.943 |
| 550 | 5.947 |
| 575 | 5.950 |
| 600 | 5.954 |
Table S4. Lattice constants of unit cell at different temperatures in cubic phase.
The structures are first equilibrated at their target temperature for 100 ps with timesteps of 2 fs in an ensemble using Langevin thermostat with friction coefficients set to for all atomic species. The atomic positions and velocities of the final frames from the equilibration runs are used as the starting point for 2 ns long production runs in an ensemble using Nose-Hoover thermostat with Nose mass of 0.08 ps. To ensure the proper sampling of diffusion coefficients, we perform at least 5 production runs at each temperature.
Multiple migration events are observed for all defects, except . To quantify the migration behavior of the defects, the mean squared displacement (MSD) was plotted over time for each atomic species using the MDAnalysis Python library; [15, 16] one such plot is given in Figure S7. The MSD is calculated using
Figure S7. Mean Squared displacement (MSD) curves decomposed to the atomic species over simulation time for I defect in CsPbBr at 550 K.
where is the number of atoms of a particular atomic species, and are their coordinates in dimensions (3 for our systems). The Diffusion coefficient is calculated using
which is proportional to the slope of the MSD curve, where the factor ensures that the diffusion coefficient is defect concentration independent.
The diffusion behavior for all five mobile species can be fitted by an Arrhenius relation
where is the Boltzmann constant, the activation energy, and the pre-exponential factor. The fits are given in Figure S8.
Figure S8. Temperature-dependent diffusion coefficients of halide point defects obtained from the MD simulations with the MLFFs for halide point defects in . The filled symbols represent halide interstitials, and the open symbols represent halide vacancies. The dashed lines represent the fits to an Arrhenius expression, and the error bars represent the standard error in mean at each point.
The , , and diffusion coefficients at room temperature (), extrapolated from
the Arrhenius relation (Eq. (5)), are given in Table S5. At this temperature, interstitials and vacancies in their most stable charge states, and , have similar diffusion coefficients ( vs ). The neutral defects, with capturing a hole or capturing an electron, behave differently. The bromide interstitial migrates at a similar rate to , while the vacancy migrates at an order of magnitude slower than . Finally, capturing two holes makes the interstitial much less mobile, whereas capturing two electrons makes the vacancy immobile in the present simulations.
| System | (eV) | () | () |
Table S5. Activation energies () and pre-exponential factors () extracted from the Arrhenius fits, and extrapolated diffusion constants () at room temperature.
Model efficiency with chemical species
In addition, we compare the efficiency of the Allegro model with the GAP-SOAP model as implemented in VASP, these comparisons are given in Figure S9. Here, the MD runs using the GAP-SOAP model are performed in VASP using 8 cores on AMD EPYC 9654P CPUs, and MD using the Allegro model is performed in LAMMPS using a single NVIDIA Tesla v100 GPU.
Figure S9. Number of MD steps per second of simulation time for the and phase transition models in the GAP-SOAP and Allegro architecture.
For the GAP-SOAP model, the efficiency drops significantly from 53.66 steps/s to 8.96 steps/s when the number of chemical species in the system is increased from three () to four () (Figure S9). On the other hand, the efficiency of the Allegro model does not scale with the number of chemical species in the system, staying around 29 steps/s for both and (Figure S9).
defect migration model training
Training structures are sampled from three constant temperature MD runs using the on-the-fly learning procedure as implemented in VASP, for both negatively charged halide interstitials and positively charged halide vacancies. These runs are performed for 100 ps with an MD timestep of 2 fs, using cubic supercells of (8 units) with one halide point defect in them. These runs are performed at 700 K, 600 K, and 750 K in ensembles at Pa pressure. To maintain constant temperature and pressure, Parinello-Rahman dynamics is used with friction coefficients set to for all atomic species and lattice degrees of freedom.
A total of 4231 and 4740 structures are sampled for the halide interstitial and halide vacancy, respectively. We run DFT single point calculations on all these structures using the same functional, PAW potentials for Br,Cs, and Pb, and kinetic energy cutoff as in Section 1.1 of SI note 1, () PAW potential for I, energy convergence criteria of eV, and Monkhorst-Pack -grid to generate the final training and validation sets.
The Allegro model[17, 18] is trained using a radial cutoff of 6.5 Å with interatomic distances projected onto a radial basis using trainable Bessel functions, and 2 tensor product layers; 32 ordered-pair tensor features are used, expanded using spherical harmonics with a maximum angular quantum number () of 2, while preserving full symmetry. The 2-body latent multilayer perceptron (MLP) had the dimensions [128, 256, 512, 1024] and SiLU nonlinearity, and the latent MLP had the dimensions [1024, 1024, 1024] also with SiLU nonlinearity. To predict the pair energies, the final MLP has dimensions [128] without nonlinearity.
The training and validation sets consist of 7177 and 1794 structures, respectively, a 80% and 20% split of the total number of structures. These structures are shuffled at each epoch during training. The training is performed using the total energy of the system, forces acting on the atoms, and stress tensors. We use the per-atom MSE loss function, with the weight set to 1 for both energy and forces. The Adam optimizer in PyTorch is used with the default parameters , , and . A learning rate of 0.001 and a batch size of 5 are used, and training ran for 592 epochs.
The RMSE in energy and forces during model training and model validation are given in Figure S10. The final model has the training errors of 11.2 meV/atom in energy and 2.88 meV/Å in forces, and validation errors of 14.6 meV/atom in energy and 12.90 meV/Å in forces.
Figure S10. Root mean squared errors in energy () and forces () during model training and model validation of the CsPb(IBr) defect migration model.
CsPb(IBr) defect migration model validation
To check the accuracy of the model, we sample 20 structures from 0.5 ns MD runs at 600 K performed using supercells with one halide point defect for pure CsPbI, CsPb(IBr) with I and Br positioned randomly on halide sites, and pure CsPbBr, and compare forces calculated using DFT with forces calculated using the NNP for these structures. The forces categorized according to the atomic species acting on all atoms, with the corresponding values and mean absolute errors (MAE) are given in Figure S11. From these comparisons, we note that the and MAE meV/Å for all systems. Considering that the forces acting on these atoms are of the order of 1 eV/Å, the NNP is highly accurate in calculating forces acting on all atoms. On top of that, the NNP is also highly transferable to different halide compositions, with the values of MAE being very close for all systems, between 48.33 meV/Å and 51.55 meV/Å.
Figure S11. Comparison between forces acting on all atoms calculated by the NNP with forces calculated using DFT for supercells with one halide interstitial (a-c), or one halide vacancy (d-f) for CsPbI, CsPb(IBr), and CsPbBr, along with corresponding values and mean absolute errors (MAE).
We also compare the forces acting on atoms close to the defect environment, these comparisons are given in Figure S12. From these comparisons, we conclude that the NNP also calculates forces acting on atoms in the defect environment with high accuracy, with the and MAE meV/Å for all systems.
Figure S12. Comparison between forces acting on atoms in the defect environment calculated by the NNP with forces calculated using DFT for supercells with one halide interstitial (a-c), or one halide vacancy (d-f) for CsPbI, CsPb(IBr), and CsPbBr, along with corresponding values and mean absolute errors (MAE).
Along with the forces, we also compare the energies predicted by the NNP with those calculated using DFT along defect migration paths. We take 4 different migration paths, one for each halide species, for both vacancy and interstitial. These paths are constructed using climbing image nudged elastic band (CI-NEB) calculations on orthorhombic supercells of CsPb(IBr) (16 units) with I and Br placed randomly on halide sites with one halide point defect for all systems except interstitial migrating through Br (I (Br)), for which convergence is not achieved using a orthorhombic supercell, hence a cubic supercell (8 units of CsPb(IBr)) is used. 3 intermediate images, connected by springs with spring constant 5 eV/Å, are optimized, using a plane wave cut-off energy of 300 eV, and and Monkhorst-Pack -point grid for the orthorhombic and cubic supercells, respectively. The computed migration barriers are given in Table S6.
| System | (eV) | (eV) |
| (Br) | 0.429 | 0.358 |
| (I) | 0.186 | 0.113 |
| (Br) | 0.362 | 0.389 |
| (I) | 0.294 | 0.277 |
Table S6. Defect migration barriers () calculated using DFT and the NNP for CI-NEB migration paths.
The difference between the migration barriers calculated using DFT and the NNP is less than 0.08 eV for all systems, indicating the high accuracy and transferability of the NNP in calculating the migration barriers.
Finally, the NNP optimized defect structures are compared with those optimized using DFT to check the accuracy of the model in capturing subtle structural differences between defect configurations. orthorhombic supercells of CsPb(IBr) (16 units) with I and Br positioned randomly on halide sites, with one halide point defect, interstitial, in either one of the three configurations, or vacancy, are used. The DFT optimization is performed using the same functional, PAW potentials, and kinetic energy cutoff as in Section 3, energy convergence criteria of eV between electronic steps, force convergence criteria of eV/Å between ionic steps, and Monkhorst-Pack -grid. The NNP optimization is performed using force convergence criteria of eV/Å. The average Pb–I and Pb–Br bond distances are given in S7, the NNP calculated distances are in excellent agreement with the DFT calculated ones.
| Defect | (Å) DFT | (Å) NNP | (Å) DFT | (Å) NNP |
| 3.20 | 3.20 | 3.02 | 3.01 | |
| 3.21 | 3.20 | 3.02 | 3.01 | |
| 3.22 | 3.22 | 3.00 | 3.00 | |
| 3.19 | 3.19 | 3.00 | 3.00 |
Table S7. Average Pb–I () and Pb–Br () bond distance in CsPb(IBr) structures with one halide point defect, interstitial, in either one of the three configurations, or vacancy, optimized using DFT and the NNP.
Further, the distance between the halide atoms forming the interstitial, and the Pb atoms adjacent to the vacancy are given in Table S8. Distances for all the NNP optimized structures are also in very good agreement with the DFT optimized structures.
| Defect | (Å) | (Å) |
| 3.85 | 3.79 | |
| 3.65 | 3.65 | |
| 3.57 | 3.56 | |
| 6.50 | 6.50 |
Table S8. Distance between halide atoms forming the interstitial in the \mathrm{Pb\mathchar`-II\mathchar`-Pb} (), \mathrm{Pb\mathchar`-IBr\mathchar`-Pb} (), and the \mathrm{Pb\mathchar`-BrBr\mathchar`-Pb} () configurations, and the distance between the Pb atoms adjacent to the halide vacancy () optimized using DFT and the NNP.
CsPb() phase transition
Training structures are sampled from six constant temperature MD runs using the on-the-fly learning procedure as implemented in VASP. These runs are performed for 50 ps with timesteps of 10 fs, using pseudo-cubic supercells of CsPb() (8 units) in all three perovskite phases. The starting structures are constructed such that Pb atoms are coordinated to different numbers of I and Br atoms. This ensures that diverse halide environments are captured for Pb atoms. We start training in the cubic phase at 700 K and 500 K, following that we train in the tetragonal phase at 500 K and 370 K, and finally in the orthorhombic phase at 350 K and 150 K. All these runs are performed in an ensemble with Parinello-Rahman dynamics with friction coefficients set to for all atomic species and lattice degrees of freedom.
A total of 1062 structures are sampled, and we run DFT single-point calculations on all these structures using the same DFT parameters as in SI note 3 to generate the final training and validation sets.
An Allegro model is trained using the same model parameters as in SI note 3, with the training and validation sets consisting of 850 and 212 structures, respectively, and the training ran for 581 epochs. The RMSE in energy and forces during model training and model validation are given in Figure S13. The final model has the training errors of 14.2 meV/atom in energy and 4.08 meV/ in forces, and validation errors of 24.1 meV/atom in energy and 24.8 meV/ in forces.
Figure S13. Root mean squared errors in energy () and forces () during model training and model validation of the CsPb(IBr) phase transition model.
This NNP is used to perform constant temperature MD runs in an ensemble to obtain equilibrated volumes at different temperatures and different halide mixing ratios for CsPb(IBr).
The accuracy and transferability of this model are validated by performing heating runs from 100 K to 500 K for 0.4 ns using a time step of 2 fs for CsPbBr, and from 100 K to 700 K for 0.6 ns using a time step of 2 fs for CsPbI. These runs are performed using orthorhombic supercells (72 units of CsPbX). The window-averaged pseudo-cubic lattice vectors with the window lengths 0.12 ns for CsPbI and 0.08 ns for CsPbBr from these runs are given in Figure S14 and b, respectively. The overall phase diagram closely resembles that observed in experiments.[14, 19]
Figure S14. Pseudo-cubic unit cell lattice vectors of a) CsPbI and b) CsPbBr as functions of temperature in the heating MD runs performed using the CsPb(IBr) phase transition NNP.
Halide diffusion in
2 ns long MD runs are performed at different temperatures between 500 K and 600 K using supercells (216 units of ) with one halide point defect in LAMMPS.[20] The lattice constants are obtained using the procedure described in SI note 5, and are given in Table S9.
| Temperature (K) | Lattice constants (Å) |
| 500 | 6.143 |
| 525 | 6.148 |
| 550 | 6.154 |
| 575 | 6.159 |
| 600 | 6.164 |
Table S9. Lattice constants of unit cell at different temperatures in cubic phase.
The structures are first equilibrated to the target temperature for 100 ps with timesteps of 2 fs in an ensemble using a Langevin thermostat with the relaxation time set to 0.33 ps. Following equilibration, 2 ns long production runs with timesteps of 2 fs are performed in an ensemble using a Nose-Hoover thermostat with the relaxation time set to 0.2 ps. To ensure proper sampling of diffusion coefficients, 5 runs are performed at each temperature. The MSD is plotted over time for each atomic species, one such plot is given in Figure S15. The diffusion coefficients for halide species are calculated using the procedure mentioned in § 7.1.4 for the SI note 1, and are given in Figure 2 in the main text.
Figure S15. Mean squared displacement (MSD) curves decomposed to the atomic species over simulation time for halide interstitial in CsPb(IBr) at 550 K.
Halide bridge configuration
To check which halide bridge configuration is the most energetically favorable, we do a population analysis of the three different configurations from the MD simulations, which are given in Figure S16. As evident from the figure, the Pb–BrBr–Pb is the most observed configuration, followed by the Pb–IBr–Pb configuration, both of which are observed significantly more than the Pb–II–Pb configuration at all simulated temperatures.
Figure S16. Population () of all three interstitial bridge configurations at different temperatures.
Halide fluctuations
To check which halide species fluctuates more around their mean position, we perform a 5 ns long MD run using a pristine cubic supercell of at 600 K. The structure is first equilibrated to the target temperature for 100 ps, and both equilibration and production runs are performed in an ensemble using the same thermostat settings as described in SI note 6. The MSD is plotted over time for each atomic species, and the plot is given in Figure S17. The MSD value is Å for I and Å for Br, indicating that I fluctuate more around their mean positions than Br.
Figure S17. Mean squared displacement (MSD) curves decomposed to the atomic species over simulation time for pristine at 600 K.
Bulk interfaces
To study the influence of halide migration on halide phase mixing in , migration behavior of halide point defects along bulk interfaces is studied. We define two bulk interfaces, the Br-rich interface, where the Pb atoms on the interface are coordinated to five Br and one I atoms (Figure S18), and the I-rich interface, where the Pb atoms on the interface are coordinated to five I and one Br atoms (Figure S18).
Figure S18. a) Br-rich and b) I-rich bulk interfaces, the interface layer is surrounded by dashed pink lines.
We start by calculating the formation energies of these interfaces. First, the geometry of a cubic supercell of with one I-rich interface and one Br-rich interface is optimized using the same functional, PAW potentials, kinetic energy cutoff, and convergence criteria as in section 1.1 of the SI note 1, and a Monkhorst-Pack -grid.
Using the volume of this optimized cell as the basis, structures with 2 interfaces of each kind are created, and positions of the atoms are optimized using the DFT parameters mentioned above. Finally, we calculate the formation energies of these interfaces () as
where is the energy of the interface system, is the ratio of number of I atoms to the total number of halide atoms in the system, and are the energies of bulk CsPbI and CsPbBr, respectively, and is the number of unit interfaces in the system. The difference between the formation energies of both interfaces, i.e., is very small, specifically 2.5 meV or 0.1 at room temperature, therefore, we conclude that both of these interfaces are equally likely to form under equilibrium conditions.
Next, we perform five 2 ns long MD runs at 600 K using cubic supercells of CsPb(IBr) with 2 interfaces of each kind and one halide point defect, such system sizes are chosen to minimize the interactions between the interfaces. The lattice constants are obtained using the procedure described in SI note 5. The structures are first equilibrated to the target temperature for 100 ps, and both equilibration and production runs are performed in an ensemble using the same thermostat settings as described in SI note 6.
Cubic domains
To demonstrate the influence of defect occupancy at bulk interfaces on halide phase mixing, we perform 10 ns long MD runs at 600 K using cubic supercells of with I cubic domain in the center of the Br lattice and vice versa, with one halide point defect. These model structures consist of six I-rich interfaces for I cubic domain and six Br-rich interfaces for Br cubic domain. The lattice constants are obtained using the procedure described in SI note 5. The structures are first equilibrated to the target temperature for 100 ps, and both equilibration and production runs are performed in an ensemble using the same thermostat settings as described in SI note 6. The final frames from these simulations are given in Figure 5 of the main text.
References
- [1]Bartók, A. P.; Kondor, R.; Csányi, G. On representing chemical environments. Phys. Rev. B 2013, 87, 184115.DOI
- [2]Bartók, A. P.; Payne, M. C.; Kondor, R.; Csányi, G. Gaussian Approximation Potentials: The Accuracy of Quantum Mechanics, without the Electrons. Phys. Rev. Lett. 2010, 104, 136403.pubmed.ncbi.nlm.nih.gov/20481899
- [3]Kresse, G.; Furthmüller, J. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev. B 1996, 54, 11169–11186.pubmed.ncbi.nlm.nih.gov/9984901
- [4]Jinnouchi, R.; Karsai, F.; Kresse, G. On-the-fly machine learning force field generation: Application to melting points. Phys. Rev. B 2019, 100, 014105.DOI
- [5]Jinnouchi, R.; Lähnsteiner, J.; Karsai, F.; Kresse, G.; Bokdam, M. Phase Transitions of Hybrid Perovskites Simulated by Machine-Learning Force Fields Trained on the Fly with Bayesian Inference. Phys. Rev. Lett. 2019, 122, 225701.DOI
- [6]Parrinello, M.; Rahman, A. Crystal Structure and Pair Potentials: A Molecular-Dynamics Study. Phys. Rev. Lett. 1980, 45, 1196–1199.DOI
- [7]Parrinello, M.; Rahman, A. Polymorphic transitions in single crystals: A new molecular dynamics method. J. Appl. Phys. 1981, 52, 7182–7190.DOI
- [8]Kresse, G.; Joubert, D. From ultrasoft pseudopotentials to the projector augmented-wave method. Phys. Rev. B 1999, 59, 1758–1775.DOI
- [9]Perdew, J. P.; Burke, K.; Ernzerhof, M. Generalized Gradient Approximation Made Simple. Phys. Rev. Lett. 1996, 77, 3865–3868.pubmed.ncbi.nlm.nih.gov/10062328
- [10]Grimme, S.; Ehrlich, S.; Goerigk, L. Effect of the damping function in dispersion corrected density functional theory. J. Comput. Chem. 2011, 32, 1456–1465.DOI
- [11]Monkhorst, H. J.; Pack, J. D. Special points for Brillouin-zone integrations. Phys. Rev. B 1976, 13, 5188–5192.DOI
- [12]Jinnouchi, R.; Karsai, F.; Verdi, C.; Asahi, R.; Kresse, G. Descriptors representing two- and three-body atomic distributions and their effects on the accuracy of machine-learned inter-atomic potentials. J. Chem. Phys. 2020, 152, 234102.DOI
- [13]Tyagi, V.; Pols, M.; Brocks, G.; Tao, S. Tracing Ion Migration in Halide Perovskites with Machine Learned Force Fields. J. Phys. Chem. Lett. 2025, 16, 5153–5159.DOI
- [14]Näsström, H.; Becker, P.; Márquez, J. A.; Shargaieva, O.; Mainz, R.; Unger, E.; Un- old, T. Dependence of phase transitions on halide ratio in inorganic CsPb(BrₓI₁₋ₓ)₃ perovskite thin films obtained from high-throughput experimentation. J. Mater. Chem. A 2020, 8, 22626–22631.DOI
- [15]Michaud-Agrawal, N.; Denning, E. J.; Woolf, T. B.; Beckstein, O. MDAnalysis: A toolkit for the analysis of molecular dynamics simulations. J. Comput. Chem. 2011, 32, 2319–2327.
- [16]Maginn, E. J.; Messerly, R. A.; Carlson, D. J.; Roe, D. R.; Elliot, J. R. Best Practices for Computing Transport Properties 1. Self-Diffusivity and Viscosity from Equilibrium Molecular Dynamics [Article v1.0]. LiveCoMS 2018, 1, 6324.DOI
- [17]Batzner, S.; Musaelian, A.; Sun, L.; Geiger, M.; Mailoa, J. P.; Kornbluth, M.; Molinari, N.; Smidt, T. E.; Kozinsky, B. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nat. Comm. 2022, 13, 2453.DOI
- [18]Musaelian, A.; Batzner, S.; Johansson, A.; Sun, L.; Owen, C. J.; Kornbluth, M.; Kozinsky, B. Learning local equivariant representations for large-scale atomistic dynamics. Nat. Comm. 2023, 14, 579.DOI
- [19]Marronnier, A.; Roma, G.; Boyer-Richard, S.; P edesseeau, L.; Jancu, J.-M.; Bonnassieux, Y.; Katan, C.; Stoumpos, C. C.; Kanatzidis, M. G.; Even, J. Anharmonicity and Disorder in the Black Phases of Cesium Lead Iodide Used for Stable Inorganic Perovskite Solar Cells. ACS Nano 2018, 12, 3477–3486.DOI
- [20]Thompson, A. P.; Aktulga, H. M.; Berger, R.; Bolintineanu, D. S.; Brown, W. M.; Crozier, P. S.; in ’t Veld, P. J.; Kohlmeyer, A.; Moore, S. G.; Nguyen, T. D. et al. LAMMPS - a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Comp. Phys. Comm. 2022, 271, 108171.DOI

















