Arxiv:2105.00414V2 [Cond-Mat.Mtrl-Sci] 13 Jun 2021

Predicting highly correlated hydride-ion diﬀusion in SrTiO3 crystals based on the fragment kinetic Monte Carlo method with machine-learning potential

Hiroya Nakata1, a) Kyocera Corporation, Research Institute for Advanced Materials and Devices, 3-5-3 Hikaridai Seika-cho Soraku-gun Kyoto 619-0237, Japan.

Oxyhydrides have drawn attention because of their fast ion conductivity and strong

reducing properties. Recently, hydride ion migration in SrTiO3−xHx oxyhydride crystals has been investigated, showing that hydride ion migration is blocked by slow oxygen diffusion. In this study, we investigate the hydride-ion migration mechanism using a kinetic Monte Carlo approach to understanding the relationship between the hydride and oxygen ions. The difficulties in applying the method to hydride and oxygen ion migration involve complex changes in the ionic migration barrier, which shifts dynamically depending on the characteristics of the surrounding hydride and oxygen ions. We can predict these complex changes using a machine-learning neural network model. The simulation can then be performed using this model to predict the temperature-dependent ionic-migration behavior. We found that our simulation results with respect to the activation barrier for hydride ion diffusion accorded well with those obtained by experiment. We also found that hydride ion migration is affected by slow oxygen diffusion and that oxygen diffusion is accelerated by changes in the ionic migration barriers. The parallel-processing efficiency of our proposed method was 84.92 % for our 1,000-CPU implementation, suggesting that the approach should be widely applicable to simulations of ionic migration in crystals at a reasonable computational cost.

I. INTRODUCTION of studies have investigated replacing a sig-

arXiv:2105.00414v2 [cond-mat.mtrl-sci] 13 Jun 2021 nificant proportion of the hydride ions by ox- Ionic conductors are attracting great in- ide ions to form oxyhydrides8–15. Because of terest because of their potential use in their fast ion conductivity and strong reduc- many electronic devices such as batteries1–4, ing properties, oxyhydrides have potential as superconductors5, photocatalytic devices6, effective charge carriers in traditional electro- and solid fuel cells4,7. Recently, a number chemical applications16–19, and as catalysts a)Electronic mail: [email protected] for ammonia synthesis20. 1 Significant efforts have been made to- been investigated. ward understanding the experimental re- Theoretical approaches also have powerful sults for oxyhydride materials. The mo- tools for understanding hydride ion diffusion. bility of hydride ions has been measured Iwazaki et al.27,28 investigated the complex- for LaSrCoO3H0.70 via quasi-elastic neutron defect electronic structure of H in BaTiO3−δ 21 scattering and for La2xySrx+yLiH1x+yO3y and SrTiO3−δ, and with the simulation re- via impedance measurements.18 The kinetics sults suggesting that the hydride ions were 29,30 of hydride ions in ABO3-type perovskite ma- trapped by VO. Liu et al. investigated the terials are particularly challenging because formation and migration energy of H in the the materials have high conductivity, making BaTiO3−δ and K2NiF4 types of oxyhydrides, impedance measurements difficult to apply. which revealed the electronic configuration of

Hydride ion mobility in SrTiFeO3 has the hydride ions. This first-principles study been measured22 but the experiments could suggested that the activation barrier for hy- 26 not separate fully the hydride ion diffusion dride ion diffusion is 0.17 eV for SrTiO3−δ 30 from the oxide anion diffusion, and the kinet- and 0.28 eV for BaTiO3−δ . Because the ics of the hydride ions were unclear. Hydride simulated activation barrier for hydride ions ion diffusion in ATiO3−xHx could be observed (0.17 eV) was less than that obtained by ex- via the exchange reactions with D13,N23,24, periment (0.30 eV), it is suggested that the and F24. The kinetics of H/D exchange fast hydride ion migration is retarded by slow have also been analyzed by quadrupole mass oxide anion diffusion. Although theoretical spectrometry25. first-principles studies suggest that slow ox- Recently, Liu et al.26 measured the self- ide anion diffusion can retard hydride ion mi- diffusion coefficient for hydride ions by com- gration, how the presence of oxide anions af- bining secondary ion mass spectrometry and fects the kinetics of hydride ion diffusion is a first-principles study, suggesting an acti- still unclear. Therefore, kinetic simulation of vation barrier of around 0.30 eV. Although oxyhydride diffusion remains an interesting this experimental study partially explained subject of investigation. hydride ion diffusion, both experimental and The kinetic Monte Carlo (kMC) theoretical studies on this topic have been method31–33 can be used to predict long- limited in comparison to those dealing with time-scale ionic diffusion in crystals, where the diffusion of oxide anions. Oxygen mi- the ions jump into the nearest site according gration in oxyhydride materials has also not to experimentally obtained or simulated ac-

2 tivation barriers. The kMC method has been and oxide ions in SrTiO3−δ oxyhydrides. To used to predict the kinetics of many types achieve this, the activation barrier for each of ionic migration and chemical reactions hydride and oxide ion is refined based on such as yttria-stabilized zirconia34,35, oxygen changes in potential energy. diffusion in doped ceria36,37, various chem- Such a refinement of the activation barrier ical reactions38–44, material diffusion45–49, using potential energy differences has been electrochemical impedance50, chemical proposed by Koettgen et al.37, and the ap- catalysis51–57, and crystal growth58,59. proach convincingly explains the ionic con- Recently, we proposed an efficient parallel- ductivity of doped ceria obtained by exper- processing approach for predicting ionic dif- iment. Our previous research60 also follows fusion with fragment-based kMC (FkMC)60. the approach of Koettgen et al.37, with the

The method has been applied to SrTiO3−δ method being used to investigate the effect of 61 systems using a simple Manning model as Fe dopant on oxide ion diffusion in SrTiO3−δ. a pilot test. The activation barrier of hydride Here, we extend the FkMC approach to one ions found60 did not accord with experimen- based on machine learning (ML), i.e., FkMC- tally derived activation barriers26, suggesting ML, in which the complex potential-energy that the simple Manning model is inadequate landscape is modeled in terms of ML-derived for predicting ionic diffusion behavior in oxy- potentials, and the kMC simulation is per- hydrides. A major reason for the discrep- formed using these predicted changes in po- ancy may be attributed to inhomogeneous tential energy. potentials, i.e., the previous simulation model First, the original FkMC is reviewed did not consider the complex potential en- briefly and its extension to ML-based poten- ergy surface (PES) composed of both hydride tial correction is described in detail. Second, and oxide ions. Inclusion of such a complex the kinetics of hydride ion migration are eval- oxyhydride PES into the kMC method is not uated using the new FkMC-ML model and straightforward, making the kinetic analysis the differences between the original and the of oxyhydrides a challenging issue in terms of ML-corrected potentials are discussed. We computational science. also compare the evaluated activation bar- The aim of this study is to develop a kMC rier of hydride ion with those obtained by model that can simulate complex ionic mi- experiment26 to demonstrate the validity of gration in crystals, with the method being the proposed approach. Third, the diffu- applied to predicting the kinetics of hydride sion coefficient for the oxide ions is evalu-

3 site ated and the effect of H concentration on ion in perovskite crystal, ni = 8 for all oxy- both hydride and oxide ions is discussed, with gen sites. The transition probability from i j the aim of understanding the mechanism of to j (pi ) is given by ionic migration in oxyhydrides. Finally, the parallel-processing efficiency of FkMC-ML is j j tot pi =ki /r . (3) evaluated to show the effectiveness of the approach. By iteratively selecting the transition from i j to j based on the probability pi , the position II. THEORY AND METHOD of target site i is updated. Likewise, the next

A. Summary of the FkMC method n + 1th step simulation time tn+1 can be ob-

tained from current time tn and the total rate The kMC method is described in detail constant in Eq. 2: elsewhere31–33. Here, we briefly describe the kMC approach to simulating ionic diffusion. ln(r0) t = t − , (4) n+1 n rtot In the jump-diffusion kMC approach, the j transition rate (ki ) from atom i to atom j where r0 is uniform random number from zero can be estimated as " † # to one. E kj = Aexp − ij , (1) i RT Because the total number of events is atom site tot † n × n , the total rate constant r where Eij, R, T , and A are the activa- increases cubically with the size of lattice. tion energy, universal gas constant, tempera- Therefore, the simulation time step tn+1 − ture, and pre-exponential factor, respectively. tn cubically decreases, and the kMC ap- Summing the respective transition rates gives proach will be limited to relatively small an estimate of the total rate constant rtot, ex- systems. To reduce the computational cost pressed as of kMC, several block separation schemes natom nsite i 62,63 tot X X j have been developed , aiming to achieve r = ki , (2) i j good parallel-processing efficiency. We have atom site where n and ni are the total number recently proposed an alternative type of of target atoms in the system and number parallelizable scheme involving atom-based of nearest neighbor sites for atom i, respec- fragmentation60, where the transition rate is tively. For example, in the case of an oxide estimated for each atom i and the maximum

4 rate constant for an atom is deﬁned as

X j ri = ki , (5) site j∈ni max r =max (ri) . (6)

If we adopt dynamic renormalization64, then j the transition probability pi in Eq. 3 can be reformulated as

rmax kj pj = i . (7) i rtot rmax

The transition event selection can then be partitioned into the selection of an atom with FIG. 1. A: schematic illustration of the ef- rmax/rtot and the selection of an event for fect of PES correction using the ML model. B: j max schematic illustration integrating the ML model the independent atom ki /r . With this renormalization, several ions can be updated into kMC simulation. simultaneously , and the parallelization of the presence of other types of ions. The main kMC becomes easier60. In this study, we ex- idea in this study is to consider correcting tend this approach by adopting ML potential PES changes by using an ML model (the red correction. j line in Fig. 1(a)). The transition rate ki can then be reformulated as

B. Modiﬁcation of transition rate " † # E + ∆EML/2 kj,0 = Aexp − ij ij , (8) with ML potential correction i RT

ML The key parameter for kMC simulation is where ∆Eij is the potential energy differ- † the activation barrier Eij from i to j defined ence between the final and the initial vacancy in Eq. 1. In the case of oxide ion migration sites, which can be estimated via the ML in a single SrTiO3 crystal, the PES is almost model. A similar reformulation of the tran- flat, as shown by black line in Fig. 1(a), and sition rate can be found in the recent review † 37 no correction of the activation barrier of Eij of Koettgen et al. . is required. As an example, for the case of a vacancy

However, the ﬂat PES case is quite rare, transition in an SrTiO3 perovskite-type crys- with the actual PES often being aﬀected by tal, the vacancy can jump to any of the

5 at the right edge of the periodic boundary condition (PBC) (the yellow line in Fig. 2). We apply the minimum-image convention for

VO and the PBC lattice is shifted to the blue

line to locate the VO at the center of the PBC (see Fig. 2). The atoms in the oxygen site are then reordered as the first nearest neighbor (red), the second nearest neighbor FIG. 2. Schematic illustration of constructing (green), and the third nearest neighbor (or- the feature vector for oxygen vacancy VO, with ange). They are labeled down to up and left colored labels 1, 2, 3, ··· indicating the elements to right (1, 2, 3, ··· 10 in Fig. 2). Using of the feature vector. this labeling by the minimum image convention, the feature vector X can be filled for eight nearest-neighbor sites (see Fig. 1(b),and any atomic species. In the case of SrTiO Fig. 2). First, the transition rates from i to 3 perovskite-type crystal, all the oxygen sites the other sites (a, b, ··· c) are set from a pre- are symmetrical, making it possible to use a defined simulation parameter. The transition ML a single ML model f to predict the potential rate ki is then refined using the ML model ML energy difference ∆Eij . By inserting Eq. 9 ML ML j,0 ∆Eij = f (X), (9) max into Eq. 8, ki , ri, and r can be updated j where f ML is the ML model and X is the fea- and the transition probability pi can be up- ML dated using Eq. 7. ture vector used to predict ∆Eij . In this study, the training of the ML model is aimed at developing the structure–energy relation- III. COMPUTATIONAL DETAILS ship, with the energy being estimated using first-principles simulation. As noted in the Introduction, we inves- The feature vector is a one-dimension ar- tigated the kinetics of hydride and oxygen ray that represents the atomic configuration ions with the aim of understanding how in the crystal. To show how the vector X the hydride and oxygen ions interact with is prepared, an example is shown in Fig. 2). each other. To achieve this, the FkMC- In this example, we are trying to construct ML method was evaluated for four dif- the feature vector for oxygen vacancy VO, de- ferent levels of hydride ion concentration, picted in blue. The original position of VO is namely for SrTiO2.75H0.25, SrTiO2.65H0.35,

6 and SrTiO2.55H0.45, and SrTiO2.55H0.60. The construct the neural network model, where same concentrations of hydride ions (from the neural network contained four hidden lay- 0.25 to 0.45) were used in the experimental ers of 36 neurons each. To prevent overﬁt- study of Liu et al26, with the simulated acti- ting, Ridge regression (L2 normalization) was vation barriers being compared with the ex- adopted, with a coeﬃcient of 0.001.

perimental results. Although SrTiO2.55H0.60 In the second step, a kMC-ML simula- was not used in those experiments, we in- tion was performed to evaluate the diﬀusion cluded it to evaluate fully how the simulation coeﬃcients of the hydride ion and oxygen results change with an increase in hydride ion for each of the four hydride ion con-

ions. The FkMC-ML simulation involves two centrations (SrTiO2.75H0.25, SrTiO2.65H0.35,

main steps. The first step is to construct the and SrTiO2.55H0.45, and SrTiO2.55H0.60). In ML model using the structure–energy rela- this case, the simulation system size was tion and the second step is performing the 90×90×96 (the size of simulation system is FkMC-ML simulation to predict diffusion co- 100 nm) and 288,000,000 simulation steps efficients. were performed. To evaluate the apparent To construct the ML model, a first- activation energies for the various hydride ion principles simulation (DFT) was performed concentrations, the temperature range was using Quantum Espresso software65,66 with a set to 550 K, 600 K, 650 K, and 700 K, for Perdew-Burke-Ernzerhof functional67,68. We which the diffusion coefficient was experimen- used ultrasoft pseudopotentials69 and the tally measured26. The vacancy concentration cutoff energy for the plane-wave basis set was was set to 0.1 %. The simulation was par- taken to be 300 Ry. The default convergence allelized using 144 central processing units criterion of 1.0D-4 a.u was used for the geom- (CPUs). etry optimizations, and the default conver- The FkMC-ML approach was im- gence criteria 1.0D-6 a.u was used for the self- plemented within the kMC program consistent field calculations of the electronic (written in C++) and the program was

states. The system size of SrTiO3 in the first- parallelized by using a message-passing principles simulation was 3×3×3 and the to- interface. (The FkMC-ML program is tal number of atoms was 135. (More detailed available free of charge from GitHub information about constructing an ML model (https://github.com/hiroyanakata/kMC.v02)). is given in the Results and Discussion sec- The parallel-processing efficiency of tion.) In this study, we used tensorflow70 to the kMC-ML was evaluated for the

7 SrTiO2.75H0.25 case, where the system between vacancies. The size of the dataset for size was 900×900×900 (at a resolution of evaluating the structure–energy relation was about 350 nm), and the kMC-ML simulation 8,000 and, using this dataset, we constructed was performed using 1,000,000 steps. The the ML model used to predict the PES in computational costs were evaluated for the oxyhydrides. cases of 108, 216, 432, 864, and 1,000 CPUs To construct the ML model, the feature and the parallel-processing efficiency was vector X in Eq. 9 is prepared using three evaluated. steps. First, from the trajectory obtained by kMC, the selected transition event (V → O, V → H) from site i to site j is determined IV. RESULTS AND DISCUSSION and the atomic labels are assigned using the minimum image convention with respect to A. Generating the ML model from the center of the V site ( site i) (see the the DFT dataset O Theory and Method section for details). Sec- The reference datasets for the ML model ond, the labels of the atomic species (O, V, were prepared using a standard kMC sim- H) are replaced by the integer array (0, 1, 2) ulation (i.e., without the ML potential cor- to obtain the sequential integer vector X = rection) and 1,000 structures were randomly [1, 0, 0, 2, ··· , 2]. To distinguish the transi- generated in the single kMC simulation run. tion event from the i to j site, we use the neg- The system size was 3×3×3 unit cells, as ative integer labels −1 and −2 for the initial noted in the Computational Details section, i site and final j site, respectively. Finally, and independent kMC simulations were per- we introduce another integer labeling (using formed for four SrTiO3−xHx cases, where x 1 and 2) to distinguish the hydride ion and was 0.25, 0.35, 0.45, or 0.60. For each sim- oxygen ion transitions, with the transition la- ulation, one or two oxygen atoms were re- bel “1” or “2” being put at the top of the placed with vacancies, giving eight sets of feature vector. The completed feature vec- kMC simulations. Note that the insertion tor then becomes X = [1, −1, 0, −2, 2, ··· , 2], of two oxygen vacancies is at a higher con- which contains all the necessary information centration than was used in the actual kMC (reaction type, transition site positions, and simulations, but we used both one-vacancy surrounding atomic species) to determine the and two-vacancy models for the ML dataset change in PES. to include the effect of the repulsion potential The comparison between the ML energy

8 ror expected for experimentally determined activation barriers26. From Fig. 3, we note that the most of the DFT energies are lo- cated in a vicinity of zero, indicating that the PES is nearly flat. Furthermore, the comparison between ML and DFT energy shows a similar trend toward positive and negative changes in PES. These results suggest that the ML model developed is adequate for the kinetic simulation of oxyhydrides. We therefore adopt this ML model for predicting the FIG. 3. Comparison between DFT diffusion coefficients of hydride and oxide ions energy (horizontal axis) and ML prediction in SrTiO3−δ. energy (vertical axis). The colors show the numbers of data for each DFT en- B. Hydride ion diffusion coefficient ergy: (a) the results for SrTiO H , (b) 2.75 0.25 evaluation using FkMC-ML the results for SrTiO2.65H0.35, (c) the results for SrTiO2.55H0.45, and (d) the results for The hydride ion diffusion coefficients were

SrTiO2.40H0.60. evaluated for four oxyhydrides (SrTiO3−xHx, where x= 0.25, 0.35, 0.45, and 0.60), as noted (energy predicted by the ML model) and the in the Computational Details section. To density functional theory (DFT) energy (en- evaluate the effect of ML correction on the ergy estimated by the first-principles simu- diffusion coefficients, results for the standard lation) is depicted in Fig. 3. In Fig. 3, the kMC (i.e., without ML) were obtained and horizontal axis represents the DFT energy these results were compared with those ob- and the vertical axis represents the predicted tained with the help of ML potentials. The ML energy. The DFT energy denotes the en- results for the diffusion coefficients are shown ergy difference in the ionic transition from in Fig. 4, where there is a significant dif- the i site to the j site. As shown in Fig. 3, ference between the diffusion coefficients ob- the ML model reproduces the DFT energy tained with and without the ML correction. quite well and the root mean square error in For hydride ion concentrations of x = 0.25 Eq. 8 is 0.04 eV, which is less than the er- or 0.35, the PES accelerates the diffusion of

9 creased with an increase in the concentration of hydride ions. The largest activation barrier was 0.53 eV for x=0.25, and the smallest was 0.27 eV. The changes of activation barriers can be understood in terms of slow oxygen diffusion. When the hydride ion concentration is small, the diffusion of hydride ions is limited by the surrounding oxygen ions. Therefore, the diffusion of hydride ions is affected by the oxygen ion concentration, with FIG. 4. The diffusion coefficient for hy- the activation barrier increasing with oxygen dride ions, with the red closed squares de- ion concentration. However, without PES noting the diffusion coefficients using kMC- correction, the estimated activation barriers ML, and the blue closed squares denoting are slightly larger than those obtained exper- those using kMC. The horizontal axis rep- imentally (See Table I). Inclusion of ML cor- resents the inverse of temperature and the rection in Eq. 8 changes the activation barrier vertical axis represents the diffusion coeffi- from 0.42 eV to 0.28 eV for x = 0.35 and this

cients: (a) the results for SrTiO2.75H0.25, (b) simulation result with ML correction agrees the results for SrTiO2.65H0.35, (c) the results well with the experimentally derived activa-

for SrTiO2.55H0.45, and (d) the results for tion barrier (0.28 eV). Likewise, for x = 0.45,

SrTiO2.40H0.60. the simulated activation barrier with ML is 0.32 eV, agreeing well with experiment (0.30 hydride ions significantly, whereas the diffu- eV), which suggests that the simulation re- sion coefficient using kMC-ML remains the sults are reasonable.

same for x = 0.45 and is slightly reduced for Note that the activation barriers esti- x = 0.60. mated via FkMC-ML do not change very We can also note the difference in acti- much with the variation of hydride ion con- vation energy with and without PES correc- centration. However, without ML correction, tion. A summary of the activation barriers the activation barrier is significantly changed, with and without ML correction is shown in as noted above, indicating the importance Table I. Without the effect of PES correc- of the ML potential contribution. For the tion, the activation barrier monotonically de- case of x = 0.25, the activation barrier of

10 TABLE I. Apparent activation barrier for hy-

dride ion diﬀusion in SrTiO3−xHx, for concentrations x= 0.25, 0.35, 0.45, and 0.60. ML and non-ML denote the results obtained by kMC-ML and standard kMC, respectively. N/A indicates that results are not available. x non-ML ML Liu et al.26

0.25 0.54 0.27 N/A 0.35 0.42 0.28 0.28

0.45 0.32 0.32 0.30 FIG. 5. The diffusion coefficient for oxygen 0.60 0.22 0.27 N/A ions, with the red closed square denoting the diffusion coefficients using kMC-ML, and the 0.54 eV is close to the oxygen migration bar- blue closed square denoting those using kMC. rier in SrTiO3 (0.6 eV), which suggests that The horizontal axis is inverse of temperature, the rate-determining step for hydride ions is and the vertical axis represents the diffusion oxygen migration. Using ML potentials de- coefficients: (a) the results for SrTiO2.75H0.25, creases the barrier to 0.27 eV, which suggest (b) the results for SrTiO H , (c) the re- that the inclusion of the PES may affect not 2.65 0.35 sults for SrTiO H , and (d) the results for only hydride ion migration but also oxygen 2.55 0.45 SrTiO H . ion migration, with both hydride and oxy- 2.40 0.60 gen ion diffusion being accelerated. Because trations. The results for the diffusion coef- the hydride and oxygen ions interact signifi- ficients are shown in Fig. 5 and the activa- cantly, the diffusion of oxygen should be an- tion energies are listed in Table II. Without alyzed to understand fully how hydride ion PES correction, the activation barriers for diffuse in SrTiO3 crystals. We therefore also oxygen ions do not depend on the concentra- analyzed the kinetics of oxygen diffusion. tion of hydride ions. The activation barriers are around 0.6 eV, which is the same as for C. Oxygen ion diffusion coefficients oxygen migration in a single SrTiO3 crystal.

To understand fully the kinetics of SrTiO3 Inclusion of the ML potential contribu- oxyhydride, oxygen ion diffusion was also tion changes significantly the kinetics of oxy- evaluated for the four hydride ion concen- gen diffusion, particularly for small hydride

11 ion concentrations. When the hydride ion D. Hydride and oxygen ion diffusion concentration is 0.25, the diffusion coeffi- mechanisms cient for oxygen using kMC-ML is ten times larger than that obtained without using ML To help understand the diffusion of hy- potentials. The contribution of ML poten- dride and oxygen ions, the changes in the ac- tials to the diffusion coefficient is less no- tivation energy for changing concentrations ticeable when the simulation temperature is of hydride ions are shown in Fig. 6. In this high, which is reasonable because the small figure, the difference between the blue and changes in potential can be neglected given red lines shows the impact of the PES correc- the increases from thermal fluctuation. Be- tion, summarizing the impact of interactions cause of the temperature dependence, the ap- between hydride and oxygen ions. parent activation energy of oxygen ion diffu- The activation barrier for oxygen ions sion decreases from 0.54 eV to 0.35 eV for is uniformly reduced by considering PES x = 0.25. Similar trends can be observed in changes, with the activation energy being the simulation results for the other hydride around 0.4 eV (see Fig. 6(b)). By contrast, ion concentrations, with the apparent activa- with hydride ion diffusion, the effect of PES tion energies being decreased by around 0.15 correction does depend on the concentration eV using the ML potential contributions for of hydride ions. There is a reduction in ac- each hydride ion concentration. tivation energy for small hydride ion concentrations (x = 0.25 or 0.35), whereas the activation energy of the hydride ions remains TABLE II. Apparent activation barrier for oxy- similar for a higher hydride ion concentration gen ion diffusion in SrTiO3−xHx, for concentra- (x = 0.45). tions x= 0.25, 0.35, 0.45, and 0.60. ML and non- ML denote the results obtained by kMC-ML and At a low hydride ion concentration (x = 0.25), hydride ions cannot diffuse without standard kMC, respectively. involving the diffusion of oxygen ions. Be- x non-ML ML cause the estimated oxygen diffusion is accel- 0.25 0.54 0.35 erated when considering ML potentials, the 0.35 0.58 0.45 estimated diffusion of hydride ions will also 0.45 0.59 0.47 be accelerated significantly. For a moderately 0.60 0.63 0.43 high hydride ion concentration (x = 0.35 or 0.45), hydride ions can diffuse while interact-

12 diffusion is not the primary rate-determining factor in such cases. Therefore, considering the decreased oxygen activation barrier by using ML potentials does not significantly affect the activation barriers for hydride ions. Likewise, at a high hydride ion concentration (x = 0.60), a decreased oxygen activation barrier would not accelerate hydride ion diffusion. In contrast to the other cases, using the ML potentials suppresses the diffusion of hydride ions, and the activation barrier is observed to increase from 0.22 eV to 0.27 eV. In summary, the activation barrier for hydride ions is around 0.3 eV, independent of the hydride ion concentration.

In this section, we have investigated the hydride and oxygen ion diffusion mechanisms by analyzing the activation barriers’ depen- FIG. 6. Activation barriers for hydride ions in dence on the hydride ion concentration. Con- SrTiO3−xHx, for concentrations x = 0.25, 0.35, sidering the inhomogeneous PES created via 045, and 0.6. The red closed squares denote with ML model, the simulation results con- the diffusion coefficients using kMC-ML and the cur with the experimental activation barri- blue closed squares denote those using kMC: (a) ers obtained by Liu et al.26 (see Table I). the activation barrier for hydride ion diffusion This offers an insight into why the experi- and (b) the activation barrier for oxygen ion dif- mentally determined activation barrier does fusion. not depend on the hydride ion concentration. The simulation results indicate that there are ing mainly with other hydride ions. The acti- two factors determining hydride ion diffusion, vation barrier without considering ML poten- namely that increasing the concentration of tials will then also become less dependent on hydride ions accelerates hydride ion diffusion the activation barriers for oxygen ions (0.42 and that using an inhomogeneous PES accel- eV or 0.32 eV), confirming that slow oxygen erates the diffusion of oxygen. These factors

13 aﬀect the rate-determining step, particularly at low hydride ion concentrations.

E. Parallel-processing eﬃciency

Finally, the parallel-processing eﬃciency of FkMC-ML was evaluated for various numbers of CPUs in the range of 64 to 1,000, as noted in the Computational Details section.

The results are shown in Fig. 7. In the fig- FIG. 7. Computational timing and its parallel- ure, the red line is the actual computation processing efficiency for FkMC-ML. The black time, and the black line is the ideally par- dashed line is the ideal computational time, es- allelized performance. Ideally, the computa- timated by using the computational time with tional time should decrease as the number of 64 cores. CPUs increases. We obtained a parallel efficiency of 84.92 % using 1,000 CPUs in a barriers for hydride ion diffusion. comparison with the 108-CPU case. The ef- The FkMC simulation indicates that the fective parallel-processing efficiency (the ra- activation barrier for hydride ions is affected tio by which the program can be parallelized) by slow oxygen diffusion and that the esti- was 99.979 %, suggesting that most aspects mates for oxygen diffusion rates are increased of the simulation had been parallelized. when considering an inhomogeneous PES. Therefore, the inclusion of such PES changes should be an important aspect of meaningful V. CONCLUSIONS FkMC models. In this study, we have investigated hydride In this study, we have developed an and oxygen ion diffusion in SrTiO3 oxyhy- FkMC-based approach that includes neural- dride crystals using a FkMC-based simula- network capabilities, which enables the com- tion. Experimentally obtained observations plex PES landscape to be included eas- about the interaction between oxygen ion and ily in the simulation model. The parallel- hydride ion diffusion were also observed in processing efficiency of the proposed ap- our simulations, with our results concurring proach is promising, suggesting that the ap- with the experimentally obtained activation proach can be widely used for simulating

14 ionic diﬀusion in crystals. We hope that 5Y. Kamihara, T. Watanabe, M. Hirano, our FkMC-ML simulation will aid the under- and H. Hosono, Journal of the American standing of ion diﬀusion mechanisms in crys- Chemical Society 130, 3296 (2008). tals. 6M. I. Litter, Applied Catalysis B: Environ- mental 23, 89 (1999). 7T. Ishihara, Perovskite oxide for solid ox- ACKNOWLEDGMENTS ide fuel cells, Springer Science & Business Media, 2009. We thank the Research Institute for 8A. Janotti and C. G. Van de Walle, Nature Information Technology at Kyushu Univer- materials 6, 44 (2007). sity for providing computational resources. 9S. Koch, E. Lavrov, and J. Weber, Physical This research also used the computa- review letters 108, 165501 (2012). tional resources of the Fujitsu PRIMERGY 10K. Hayashi, S. Matsuishi, T. Kamiya, CX400M1/CX2550M5(Oakbridge-CX) at M. Hirano, and H. Hosono, Nature 419, the Information Technology Center of the 462 (2002). University of Tokyo through the HPCI Sys- 11G. Bouilly et al., Chemistry of Materials tem Research project (Project ID:hp200015). 27, 6354 (2015). 12T. Yajima, A. Kitada, Y. Kobayashi, T. Sakaguchi, G. Bouilly, S. Kasa- REFERENCES hara, T. Terashima, M. Takano, and

1Y. Zhao, L. Xu, L. Mai, C. Han, Q. An, H. Kageyama, Journal of the American X. Xu, X. Liu, and Q. Zhang, Proceedings Chemical Society 134, 8782 (2012). 13 of the National Academy of Sciences 109, Y. Kobayashi et al., Nature materials 11, 19569 (2012). 507 (2012). 14 2J. Suntivich, H. A. Gasteiger, N. Yabu- C. Tassel, Y. Goto, Y. Kuno, J. Hes- uchi, H. Nakanishi, J. B. Goodenough, and ter, M. Green, Y. Kobayashi, and Y. Shao-Horn, Nature chemistry 3, 546 H. Kageyama, Angewandte Chemie 126, (2011). 10545 (2014). 15 3N. Kamaya et al., Nature materials 10, 682 C. Tassel et al., Angewandte Chemie Inter- (2011). national Edition 55, 9667 (2016). 16 4L. Malavasi, C. A. Fisher, and M. S. Islam, Z. Zhang, Y. Zhu, Y. Zhong, W. Zhou, and Chemical Society Reviews 39, 4370 (2010). Z. Shao, Advanced Energy Materials 7,

15 1700242 (2017). 27Y. Iwazaki, T. Suzuki, and S. Tsuneyuki, 17S. Yamaguchi, Science 351, 1262 (2016). Journal of Applied Physics 108, 083705 18G. Kobayashi, Y. Hinuma, S. Matsuoka, (2010). A. Watanabe, M. Iqbal, M. Hirayama, 28Y. Iwazaki, Y. Gohda, and S. Tsuneyuki, M. Yonemura, T. Kamiyama, I. Tanaka, APL Materials 2, 012103 (2014). and R. Kanno, Science 351, 1314 (2016). 29X. Liu, T. S. Bjørheim, and R. Haugsrud, 19Y. Yamazaki, F. Blanc, Y. Okuyama, Journal of Materials Chemistry A 5, 1050 L. Buannic, J. C. Lucio-Vega, C. P. Grey, (2017). and S. M. Haile, Nature materials 12, 647 30X. Liu, T. S. Bjørheim, and R. Haugsrud, (2013). Journal of Materials Chemistry A 6, 1454 20Y. Kobayashi, Y. Tang, T. Kageyama, (2018). H. Yamashita, N. Masuda, S. Hosokawa, 31C. Caron, editor, An Introduction to Ki- and H. Kageyama, Journal of the Amer- netic Monte Carlo Simulations of Surface ican Chemical Society 139, 18240 (2017). Reactions, Springer, Heidelerg in Germany, 21C. A. Bridges, F. Fernandez-Alonso, J. P. 2012. Goﬀ, and M. J. Rosseinsky, Advanced ma- 32G. Murch, American Journal of Physics 47, terials 18, 3304 (2006). 78 (1979). 22S. Steinsvik, Y. Larring, and T. Norby, 33K. E. Sickafus, E. A. Kotomin, and B. P. Solid State Ionics 143, 103 (2001). Uberuaga, Radiation eﬀects in solids, vol- 23T. Yajima et al., Nature Chemistry 7, 1017 ume 235, Springer Science & Business Me- (2015). dia, 2007. 24N. Masuda et al., Journal of the American 34K. C. Lau, C. H. Turner, and B. I. Dunlap, Chemical Society 137, 15315 (2015). Chemical Physics Letters 471, 326 (2009). 25Y. Tang, Y. Kobayashi, K. Shitara, A. Kon- 35K. C. Lau, C. H. Turner, and B. I. Dunlap, ishi, A. Kuwabara, T. Nakashima, C. Tas- Solid State Ionics 179, 1912 (2008). sel, T. Yamamoto, and H. Kageyama, 36S. Grieshammer, S. Eisele, and J. Koettgen, Chemistry of Materials 29, 8187 (2017). The Journal of Physical Chemistry C 122, 26X. Liu, T. S. Bjørheim, L. Vines, Ø. S. 18809 (2018). Fjellv˚ag,C. Granerød, Ø. Prytz, T. Ya- 37J. Koettgen, S. Grieshammer, P. Hein, mamoto, H. Kageyama, T. Norby, and B. O. Grope, M. Nakayama, and M. Mar- R. Haugsrud, Journal of the American tin, Physical Chemistry Chemical Physics Chemical Society 141, 4653 (2019). 20, 14291 (2018).

16 38S. Piccinin and M. Stamatakis, ACS Catal- 50R. Pornprasertsuk, J. Cheng, H. Huang, ysis 4, 2143 (2014). and F. B. Prinz, Solid State Ionics 178, 39C. Wu, D. Schmidt, C. Wolverton, and 195 (2007). W. Schneider, Journal of Catalysis 286, 51E. W. Hansen and M. Neurock, Chemical 88 (2012). engineering science 54, 3411 (1999). 40M. Stamatakis, Y. Chen, and D. G. Vla- 52E. W. Hansen and M. Neurock, Surface chos, The Journal of Physical Chemistry C science 464, 91 (2000). 115, 24750 (2011). 53E. W. Hansen and M. Neurock, Journal of 41L. Yang, A. Karim, and J. T. Muckerman, Catalysis 196, 241 (2000). The Journal of Physical Chemistry C 117, 54K. Reuter and M. Scheﬄer, Physical Re- 3414 (2013). view B 73, 045433 (2006). 42W. Guo, M. Stamatakis, and D. G. Vla- 55M. Stamatakis, Y. Chen, and D. G. Vla- chos, ACS Catalysis 3, 2248 (2013). chos, The Journal of Physical Chemistry C 43W. Guo and D. G. Vlachos, Nature com- 115, 24750 (2011). munications 6, 8619 (2015). 56J. A. Boscoboinik, C. Plaisance, M. Neu- 44S. Lin, J. Ma, L. Zhou, C. Huang, D. Xie, rock, and W. T. Tysoe, Physical Review B and H. Guo, The Journal of Physical Chem- 77, 045422 (2008). istry C 117, 451 (2013). 57L. Kunz, F. M. Kuhn, and 45S. M. Auerbach, International reviews in O. Deutschmann, The Journal of chemical physical chemistry 19, 155 (2000). physics 143, 044108 (2015). 46S. Matera, H. Meskine, and K. Reuter, The 58G. Gilmer, Science 208, 355 (1980). Journal of chemical physics 134, 064713 59T. P. Schulze, Journal of crystal growth (2011). 263, 605 (2004). 47B. Temel, H. Meskine, K. Reuter, M. Schef- 60H. Nakata, Computational Materials Sci- ﬂer, and H. Metiu, The Journal of chemical ence 184, 109844 (2020). physics 126, 204711 (2007). 61J. R. Manning and L. Bruner, AmJPh 36, 48M. Rieger, J. Rogal, and K. Reuter, Phys- 922 (1968). ical review letters 100, 016105 (2008). 62K. Li, H. Shang, Y. Zhang, S. Li, B. Wu, 49R. Pornprasertsuk, T. Holme, and F. B. D. Wang, L. Zhang, F. Li, D. Chen, and Prinz, Journal of The Electrochemical So- Z. Wei, Openkmc: a kmc design for ciety 156, B1406 (2009). hundred-billion-atom simulation using mil- lions of cores on sunway taihulight, in Pro-

17 ceedings of the International Conference for and C. Fiolhais, Physical review B 46, 6671 High Performance Computing, Networking, (1992). Storage and Analysis, pages 1–16, 2019. 68J. P. Perdew, K. Burke, and M. Ernzerhof, 63T. Tada and N. Watanabe, ECS Transac- Physical review letters 77, 3865 (1996). tions 57, 2437 (2013). 69D. Vanderbilt, Phys. Rev. B 41, 7892 64S. Grieshammer, B. O. Grope, J. Koettgen, (1990). and M. Martin, Physical Chemistry Chem- 70M. Abadi et al., Tensorﬂow: A system ical Physics 16, 9974 (2014). for large-scale machine learning, in 12th 65P. Giannozzi et al., Journal of physics: {USENIX} symposium on operating sys- Condensed matter 21, 395502 (2009). tems design and implementation ({OSDI} 66P. Giannozzi et al., Journal of Physics: 16), pages 265–283, 2016. Condensed Matter 29, 465901 (2017). 67J. P. Perdew, J. A. Chevary, S. H. Vosko, K. A. Jackson, M. R. Pederson, D. J. Singh,