Advancement of solid state electrolytes (SSEs) for all solid state batteries typically focuses on modification of a parent structural framework for improved conductivity, e.g. cation substitution for an immobile ion or varying the concentration of the mobile ion. Therefore, novel frameworks can be disruptive by enabling fast ion conduction aided by different structure and diffusion mechanisms, and unlocking optimal conductors with different properties (e.g. mechanical properties, sintering needs, electrochemical stability) than previously published. Herein, we perform a high throughput survey of an understudied structural framework for sodium ion conduction, Na8−xAxP2O9 (NAP), to understand the family's thermodynamic stability, synthesizability, and ionic conduction. We first show that the parent phase Na4TiP2O9 (NTP) undergoes a structural distortion (with accompanying conductivity transition) due to unstable phonons from a pseduo-Jahn Teller mode in the 1D titanium chains. Then, screening of cation-substituted structural candidates with ab initio and machine-learned potential calculations reveal a number of candidates that are thermodynamically stable, likely synthesizable, and have high predicted ionic conductivities. High throughput experimental trials and subsequent methodology optimization of one Na4SnP2O9 (NSP) highlight collective challenges to the synthesis pathways for sodium phosphate materials via solid state synthesis. Our results demonstrate that NAP is a highly tunable conduction framework whose high temperature conductivity transition has heretofore eliminated it from significant research interest. By expanding the structural toolkit for SSE design, we increase the number of useful sodium ion electrolytes for integration into safe and accessible solid state batteries.
Introduction
First principles design of new solid state electrolytes (SSEs) often incorporates structural framework that are engineered to allow for fast ion conductivity. When new frameworks are reported, they are dissected, categorized, and integrated into high-throughput studies to understand the conductivity pathway and diffusion mechanisms, and how these might be optimized. For lithium SSEs, the significant number of frameworks have allowed studies to distill key features and design principles utilizing high-throughput materials screening and machine learning-based approaches [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]. However, rational development of sodium and other non-lithium ionic conductors is impeded by the relatively smaller pool of reported examples to draw from, and the only handful of studies that have attempted to featurize the frameworks [11] [12]. Therefore, uncovering a single new framework provides a key additional data point for the computational modeling and machine learning studies necessary for data driven sodium SSE development.
The NTP superionic conductor class was first introduced by T. Takahashi, et al. at the International Conference on Solid State Electrolytes in 1981 [13] (visualized in Figure 1). Subsequent studies were released in the 1980s and 1990s to clarify the structure of the low and high-temperature polymorphs, and verify a one-order-of-magnitude change in conductivity upon phase transition near 250 °C [14–19]. The framework presented in NTP falls into a general formula Na8−xAxP2O9 (NAP) and is composed of large, highly-coordinated sodium cages pinned by 1D cation chains. Despite the initial success and ease of synthesis of the original NTP structure, there are only handful of other reported cation substituted NAP phases [16, 20, 21]. However, no studies known to the authors describe detailed the mechanism behind the structural or conductivity transitions, underscoring the limited knowledge and research activity on NTP, despite its high ionic mobility. Moreover, within the original Google set, only a handful of materials were found that had low activation energies for sodium ion diffusion, and NTP had one of the lowest. The lack of reported structural and conductivity investigations, particularly with modern first principles calculations, in conjunction a narrow understanding of the chemical flexibility of the structure, demands an immediate and broad survey of the NAP framework.
Figure 1. Structural visualization of Na4TiP2O9, showing a) multiple formula units with patterned i and ii layers in (001) b) components of the structure, including atoms and their respective Wyckoff sites, titanium octahedra chains, and the two types of sodium polyhedra, 5-n square pyramidal and 8-cn cubic polyhedra.
Historically, investigation of new materials represents a bottleneck towards accelerating new phase discovery; trial-and-error-based experimental synthesis development is limited by time, resources, and expertise, with no guarantee of success. Meanwhile, as new materials are predicted, critical a priori questions for emerging materials families include synthesizability and properties of prospective phases. However, output from first principles calculations and machine learning models can fail to reflect the structures and properties of real-world samples. This gap in knowledge between computationally predicted and experimentally realized results limits the scope that ambitious computational research can reasonably tackle. To address this challenge, the integrated use of computational and experimental tools such as open-access material databases, machine-learned potentials, synthesis science models, and self-driving labs is essential. Feedback between these different strategies fosters more efficient sampling of materials spaces. Moreover, the culmination of positive and negative results can paint a more complete picture of the materials discovery process, introducing new insights towards the synthesis development and phase characterization processes.
We here conduct a survey of the NAP framework to report key components necessary for development, including thermodynamic stability, synthesizability, conductivity predictions and measurements, and the sodium diffusion pathway. Experimental synthesis attempts of 3+, 4+, and 5+ cation substitutions yields successful synthesis of one new material, Na4SnP2O9 (NSP). A closer look into the physics behind previously reported NTP gives crucial insight into why NAPs have a sudden drop in ionic conductivity upon distortion to lower temperature phase, and what modifications might be necessary to suppress this transition.
Methods
Experimental Methods
Solid state synthesis
Reactions were completed with traditional, powdered solid state (ceramic) synthesis. Reagents were purchased from Sigma-Aldrich (Millipore Sigma) and BeanTown Chemical at generally 98% or greater purity. Due to the significant number of reagents used, the full list of precursors used, CAS numbers, manufacturer, and purity are included in the Supplemental Materials. As sodium pyrophosphate is purchased as hydrated Na4P2O7⋅10H2O, it is heated for at least 24 hours at 70 °C under vacuum to remove crystal water before use.
For all syntheses, stoichiometric amounts of reagent are weighed within a 0.1% threshold and then mixed in air. Mixing is done via either wet (ball) milling, pestle and mortar grinding, or dry ball milling. Wet (ball) milling is carried out as described by Z. Lun, et al. [22] First, a slurry is made by spin mixing ethanol and precursors together in plastic vial with 10 zirconia balls for 10 minutes within a Hauschild Speedmixer machine (DAC 150.1 FVZ-K) at 2,000 rpm. The slurry is then immediately pipetted into a crucible. The crucible is heated at 80 °C in air to fully remove the ethanol. Mortar and pestle mixing involves placing powdered precursor in a pestle and mortar, and then grinding them by hand for at least 10 minutes until fully ground. For dry ball-milling, the powder is first carefully ground using a pestle and mortar for 15 minutes in the air. This mixture is then placed in a 50 mL zirconia jar. with 5mm -zirconia balls and milled (SPEX 8000M Mixer/Mill high energy ball mill) for 50 min. The ball milling followed by hand grinding with mortar and pestle are repeated twice, with an additional final hand grinding for 15 minutes. The pestle and mortar steps are found necessary to ensure adequate mixing due to the stickiness of many phosphate precursors. After ball-milling, samples are pelletized into 13 mm pellets with a Carver Manual Bench Top Pellet Press (model 3851-0). Pelletization is completed by applying 4 tons of pressure for 5 minutes. Samples are then heated at a default of 2 °C/min to 150 °C, held for 12 hours (calcination), and then ramped at 5 °C/min to the dwell temperature for 12 hours. Other heating profiles are used where indicated. Natural (uncontrolled) cooling is completed in the furnace without opening the door. After cooling, samples are scraped from crucibles and collected in vials.
A-Lab Synthesis
Exploratory solid state synthesis is performed using the A-Lab, an autonomous and automated solid state synthesis laboratory where full details of the infrastructure are described by N. Szymanski, et al. [23]. Briefly, the A-Lab performs solid state synthesis in three stages reflective of human (manual) synthesis: (i) a precursor preparation stage, (ii) a heating stage, and (iii) a sample recovery and characterization stage. At each stage, there is a 6-axis robot arm that is programmed with high-precision waypoints to move samples through a series of tasks carried out by devices as implemented within AlabOS by Y. Fei, et al. [24] To prepare samples, powders are dispensed and weighed simultaneously by using an automatic dispenser balance (Quantos, Mettler Toledo) into a plastic vial containing 10 zirconia balls. Ethanol is added and the sample is mixed at 2000 rpm for 10 minutes to create a slurry. The slurry is pipetted from the vial and dispensed into a crucible. After drying at 80°C, the sample is left mixed and densified at the bottom of a crucible. The samples are then moved to furnaces and heated according to the programmed profile. The default heating profile is 2 °C/min to 300 °C, followed by 5 °C/min up to the maximum dwell temperature, followed by a 12-hour hold, and then uncontrolled cooling in the furnace to room temperature. Finally, the samples are recovered autonomously by vertically shaking a crucible covered with a plastic cap with a 10-mm alumina ball inside to dislodge (mill) the sample from the crucible wall and generate powder. Finally, the sample is shaken through a steel mesh and pressed onto a disc for automated powder XRD characterization, and the remaining samples are stored inside a uniquely-labeled plastic vial for further characterization.
Characterization
Sample characterization is completed primarily with powdered x-ray diffraction (XRD) on an Aeris Minerals diffractometer (Malvern Panalytical) utilizing Cu Kα radiation. Scans are performed between 10° and 100° 2θ range for 8 minutes, with a step size of ≈0.01∘ and an active length of 5.542° unless otherwise noted. In situ x-ray diffraction targeting Na4SnP2O9 samples were carried out on a Bruker in situ XRD (model: D8 ADVANCE) with Anton Paar HTK 1200N non-ambient heating chamber. Stoichiometric powders are prepared and loaded onto a 0.8 mm depth alumina sample holder (Anton Paar) and inserted into the flow cell. Scans are performed every 10°C during heating within the range of 10° and 60° 2θ, and at a step length of 0.02° with 0.04 s per scan step for 2 minutes (low fidelity scan). Samples are held for 4 hours at a 900°C dwell. Scans are taken every 10 minutes at a step size of 0.02° with 0.2 s per scan step during a total time of nearly 10 minutes (high fidelity scan). During uncontrolled cooling, temperatures are held every 100°C for a low fidelity scan.
Analysis of XRD patterns is completed through Rietveld refinement delivered by Dara and Profex, utilizing the refinement kernel BGMN [25, 26]. Dara is an in-house automated phase identification and Rietveld refinement algorithm for powder XRD pattern analysis which ranks possible phase sets based on a quality of fit figure of merit. Initial structure searches for phase identification are performed by Dara on known crystal structures from the Inorganic Crystal Structure Database (ICSD) [27]. Ultimate phase identification is performed by human analysis with the aid of best case matches from Dara. Final refinements are completed within Profex. The background is fit by Lagrangian polynomials. Crystalline phase site occupancies are assumed to be nominally stoichiometric. Lattice parameters are permitted to vary by 1% when permitted by symmetry. A preferred orientation model with fourth-order spherical harmonics (gewicht=SPHAR4) is used during refinement. The crystalline size parameters k1 and B1 are constrained isotropically in the range 0.0 to 0.1 with initial value of 0. The microstrain effect k2 is constrained between 0 and 0.1. Isotropic thermal displacement parameters are set to vary by up to 0.01 Å2 for oxygen sites.
Electrochemical impedance spectroscopy (EIS) was completed at low and elevated temperatures to assess sodium conductivity. The annealed sample is first ball milled using a high-energy ball mill (SPEX 8000·M Mixer/Mill) with zirconia balls and jar to reduce the particle size. Then the sample is pressed into a 6 mm pellet with 2 tons of weight for 3 minutes and sintered at 900 °C for 16 hours before cooling naturally in the furnace. For room temperature measurements, pellets are sanded finely and attached to two indium foils as ion-blocking electrodes. EIS measurements were performed using an EC-Lab Electrochemistry SP300 system. The measurements were conducted at the initial open-circuit voltage in the frequency range of 7 MHz–10 mHz with the application of a 10-mV signal amplitude. For high temperature measurements, the pellets were Au-coated by 2 minutes sputtering with a gold source. Silver paste (AG-I) purchased from fuelcellmaterials is then applied and platinum electrodes are attached with a platinum mesh during a 15 minute heat at 900 °C. High temperature EIS is performed with a BioLogic potentiostat (VSP-300) within a Barnstead International tube furnace (21100). Measurements are taken every 50 °C up to 800 °C, and every 50 °C during controlled cooling.
Scanning electron microscopy (SEM) energy-dispersive X-ray spectroscopy (EDS) measurements are taken using a Thermofisher Phenom XL G2 system. SEM-EDS data are collected and quantified in automated fashion with Auto-SEMEDS, a software developed in-house. Details on the employed method are reported elsewhere [28]. For each analyzed sample, 200 point spectra are randomly collected from tens of particles. The electron beam energy is fixed at 15 kV and a total of 50,000 counts per spectrum is collected. The spectra are then quantified with the peak-to-background method [29, 30], using the pure precursors as standards. Only compositions with analytical errors less than 10% are considered for this study. From each measured composition – i.e. vector of elemental atomic fractions – precursor molar fractions are obtained via a non-negative least squares optimization algorithm, employing the precursors elemental atomic fractions as fixed vector bases. The obtained coefficients are normalized to a total of 1.
Transmission electron microscopy (TEM) EDS measurements were taken using the ThemIS microscope using a superXG2 detector at 300 kV, housed at the National Center for Electron Microscopy. The beam current is recorded to be ≈0.5 nA using a spot size 3. During EDS collection, the electron counts varied between 2,000 cps to 4300 cps. The dwell time was fixed at 25 μs. Elemental quantification was done using the Thermofisher Velox software (version: 3.7.0.872). Na, Sn, O, and P compositional quantification used K, L, K, and K edges of the elements, respectively. We note that the samples were first imaged at room temperature; however, noticeable beam damage was observed in the HAADF-STEM micrograph of the particle before and after EDS mapping (shown in Supplemental). To prevent beam damage the EDS mapping was obtained under cryo-conditions using a Gatan 915 holder. As EDS is measured by suspending the particles on a foamvar stabilized lacey carbon support Cu grid, both C and Cu EDS peaks were used for deconvolution. The total summed up EDS spectrum is shown in the Supplemental.
Computational Methods
Density Functional Theory Calculations
Density functional theory calculation were carried out with Vienna Ab initio Simulation Package (VASP) leveraging the generalized gradient approximation functional Perdew-Burke-Ernzerhof [31–37]. DFT settings, including K-point density, plane wave cutoff, pseudopotential, magnetic moments, and smearing were those recommended by pymatgen’s MPRelaxSet [38, 39]. However force convergence during relaxation was tightened to 10−2 eV/Å. Energetic corrections to static calculations on optimized structures were completed with the MaterialsProject2020Compatibility class [40, 41]. Phonon calculations at 0 K utilizing the quasiharmonic approximation were executed with the Phonopy package on a 2×2×2 supercell with a minimum lattice parameter length of 10 Å [42, 43]. Phonon band structures sample the Brillion zone with reciprocal space paths recommended by SeeK-path [44, 45]. Sodium-vacancy ordering distributions are found with structures enumerated by pymatgen’s OrderDisorderedStructureTransformation class, wherein symmetrically equivalent structures are removed.
Diffusivity Calculations
Diffusivity calculations for all relevant materials were completed with molecular dynamics calculations with machine-learned potentials created with fine-tuned CHGNet [46]. CHGNet was fine-tuned separately for each structure. Each dataset for fine tuning was generated by collecting DFT structures from on-the-fly MD with FLARE [47] at 5 applied strains (-1, 0, 1, 2, 3%) and 4 temperatures (400, 600, 800, 1000 K) initiated from the DFT relaxed structure. Mean absolute errors for the validation energy, force, and stress are provided in the Supplemental. Molecular dynamics (MD) simulations were run for 50 ps in N-P-T, followed by a minimum of 2 ns under N-V-T conditions. Simulations were performed with the Berendsen thermostat [48].
For each phase, MD simulations were performed at 8 temperatures: 300, 400, 500, 600, 700, 800, 900 and 1,000 K. Na-ion self-diffusion coefficients were computed from the mean squared displacements (MSD) of the MD simulation trajectories, where the squared displacements were averaged over time intervals of duration Δt. Only the portion of the MSD curve with Δt>1 ps and Δt<50% of the total simulation time was considered in the analysis. Activation energies and room temperature diffusivities were extrapolated by fitting the diffusivities to an Arrhenius relation. Ionic conductivities were calculated from Na-ion self diffusivities assuming the Nernst-Einstein relation. Error bars for the diffusivities were computed using the empirical error estimation scheme proposed by X. He et al. [49].
Precursor Prediction Optimization
Precursor prediction models were utilized to suggest and build a list of precursors for NSP. Precursor sets were filtered safety, accessibility, product separability (if a solid byproduct is made), and the model-dependent confidence ranking. Reaction network [50, 51] was performed with the NSP phase set as the target, and with a hyperdimensional chemical system of “B-Br-C-Cl-F-H-K-Li-N-Na-O-P-S-Sn”, and a reaction temperature of 600 °C. 34,880 pairwise reactions were found to have negative or close to negative reaction energy to form the target, and were ranked by additional phase competition metrics. s4 [52] was run with the target composition as input. 10 precursor sets were output, which are modifications to previously reported syntheses for chemically similar targets. The synthesis planning algorithm [53] was executed with the respective target composition as input. 3 precursor sets were found which had negative reaction energies and fulfilled other thermodynamic metrics of synthesizability. A full list of precursor sets chosen, their reaction energy from Materials Project, total DFT energies, and their relative model outputs is shown in the Supplemental.
Results
NAP Structural Model
The generalized Na8−xAxP2O9 structural framework is derived from the high temperature, orthorhombic Bmem (67) Na4TiP2O9 parent phase (HT-NTP, ICSD 39702), visualized in Figure 1. HT-NTP is a layered structure composed of five 1D chains of polyhedra stacked into two distinct layers in the a-c plane. Na+ ions are primarily located in large 8 coordination number (cn) distorted cubic polyhedra cages (Figure 1b). One-dimensional, corner-connected TiO6 octahedra chains (Figure 1b), surrounded by chains of phosphate tetrahedra that alternate with sodium square pyramidal 5-cn polyhedra, scaffold the structure. Layer i holds the titanium octahedra and the Na24b sodium site, while the thicker layer ii contains the 8-cn cubic Na14e and Na44a cages framed by the Na38m-PO₄ columns. The low temperature, monoclinic P2/c (13) polymorph (LT-NTP, ICSD 39901) is topologically similar to HT-NTP, but displays compression and elongation of the TiO6 octahedra, and a significant distortion of sodium polyhedra resulting in changes to Na site occupations [14]. For HT-NTP, high sodium mobility (2.6×10−2Ω−1cm−1 at 300∘C [14]) was previously reported to be promoted by easy transport between and through the large 2D sodium channels in the a-c plane, and is likely supported by partial sodium disorder on the Na2 4b, Na4 4a, and Na3 8m sites [21].
We note that coupled temperature-dependent conductivity-structure transitions define ionic conductivity for all known NAP phases. The conductivity of NTP decreases upon cooling when the orthorhombic phase transforms to the monoclinic phase at approximately 300∘C [14]. Similar conductivity transitions are measured near 200-300∘C for the other two reported NAP materials Na4.6FeP2O8.6F0.4 and Na(Al,Cr)P2O9−xF2x [20].
To better understand the conductivity transition in NTP, we use ab initio calculations to explore specific structural features in NTP that might impact conductivity: the stability and sodium ordering energies of the high T symmetric structure. Figure 2 (left) shows the phonon band structure and (right) variation of the energy with sodium occupational disorder in HT-NTP. We note that the high temperature phase is dynamically unstable due to a number of negative phonon frequencies at the Γ, R, A, and Z points (depiction of the full Brillouin zone path is shown in the Supplemental.Visual inspection of the negative modes at q-points in the Brillion zone elucidates key instabilities: (Γ: 0, 0, 0) and (Z: 0, 0, 0.5) pseudo-Jahn Teller like TiO6 elongation/compression, (A: 0, 0.5, 0.5), Ti-octahedral rotation with dimerization of Na-Na in layer ii, and (R: 0.3006, 0.3006, 0.5) collective wave-like wobbling of the oxygen atoms. Modulation of individual unstable phonon mode followed by DFT structural relaxation is not able to recover the monoclinic ground state (see Supplemental). Therefore, the structural transition to the low temperature monoclinic phase is likely driven by the combination and interaction of multiple unstable phonon modes.
Figure 2. Figure showing (left) HT-NTP lattice dynamics calculations show negative modes at R, A, and Z visualized by the green phonon band (right) energetic variation with sodium occupational disorder in HT-NTP.
We also explore the energetics of sodium ordering. The right panel of Figure 2 depicts the energy per sodium for 300 symmetrically inequivalent sodium occupations in HT-NTP. The lowest energy configuration, found to be a fully vacant Na4 site configuration, is set to zero. Higher energy configurations range from 74.4 meV/Na up to 277.6 meV/Na. In addition, we perform DFT calculations on the LT-NTP phase, where we explore the filling of inequivalent Na sites. In LT-NTP, the Na4 site is distorted such that there are now four inequivalent cation centers in the same 1D chain along the c-axis. Only one is reported to be filled [14]. Cell volume and shape are fixed to the experimental values, while atomic position are allowed to move during relaxation. We find that there is an increase ranging from 447 meV/Na and 969 meV/Na as sodium is moved to other sites previously reported to be vacant see Supplemental. This large energy change with site occupation variation represents a possible barrier for Na mobility in the LT-NTP. Together, we find that the sodium site energy landscape is not flat, and structural distortions to sodium polyhedra impact this landscape substantially.
We verify the synthesis, characterization, and impedance of NTP. NTP samples are made by wet ball milling Na2CO3, NH4H2PO4, and TiO2 precursors, followed by heating at 950°C. PXRD confirms the presence of monoclinic Na4TiP2O9 with some Na4P2O7 (1.05w%) impurity found during refinement (see Supplemental).Temperature-dependent EIS is then performed on a conventionally-sintered pellet of ≈99% phase pure NTP. For comparison to experimental conductivities, machine-learning molecular dynamics (MLMD) diffusivity calculations were completed on the experimentally reported monoclinic and orthorhombic phases. Figure 3 plots the experimentally measured and MLMD calculated conductivities as log(σT) versus 1000/T for the temperature range of 150 to 850°C. EIS measures a conductivity transition near 200-300°C, in agreement with the literature [14]. The low and high temperature conductivity regimes exhibit activation energies of 0.836 eV and 0.449 eV, respectively. MLMD diffusivity calculations predict much lower activation energies of 0.294 eV and 0.280 eV for LT-NTP and HT-NTP, respectively. Furthermore, no conductivity transition is identified the MLMD calculations. Interestingly, MLMD diffusivity calculations demonstrates much higher conductivity and much lower diffusion barriers for both phases than that of the experimental NTP sample Figure 3. This supports the hypothesis that the experimentally observed conductivity transition might have structural origins that occur from the temperature-dependent opening of a percolation pathway.
Figure 3. Computationally predicted and experimentally measured (golden line) temperature-dependent conductivity of NTP. Pink triangles give the computed conductivity in the DFT-optimized single crystal XRD structures for the orthorhombic high temperature phase, and blue squares give computed conductivity in the low temperature monoclinic phase. The dashed gray line indicates the measured conductivity transition, occurring at approximately 555 K.
Our work with NTP shows us that the high temperature NTP structure becomes unstable upon cooling, leading to distortions in Na polyhedra and cation octahedra [11] [12]. This structural instability increases the activation energy for Na motion, limiting the room temperature Na-ion conductivity. For NAP to be a useful solid state electrolyte, it is critical to understand if cation substitution might be able to lower activation energies, increase sodium conductivity, and either stabilize the orthorhombic phase or significantly lower the phase transition temperature. Therefore, we complete a high-throughput exploration of possible A site substitutions within Na8−xAxP2O9.
High Throughput Computation Investigation of Cation Substitution
We evaluate the possibility of cation substitution on the A (Ti) site into orthorhombic NAP through high-throughput density functional theory computations. A total of 24 cation candidates with oxidation states ranging from 3+ to 5+ are chosen, with sodium content adjusted to balance charge: Al, Sc, Ti, V, Cr, Mn, Fe, Co, Ga, Ge, Y, Zr, Nb, Mo, Ru, In, Sn, Sb, La, Hf, Ce, Ta, W, and Bi. When the A cation has a 3+ charge, all Na sites are filled. For a 4+ cation, we arrange the Na ions according to the ground state in NTP (vacant Na4 Wyckoff site). Similarly, sodium occupation for A 5+ was set by keeping Na4 and half of Na3 sites vacant, in line with common occupancies in NTP [17, 18]. To our knowledge, all substituted Na8−xAxP2O9 are unreported in literature. The substituted NAP candidates are structurally optimized with DFT and manually inspected after relaxation to ensure that symmetry, structure topology, and sodium site occupancies does not change.
To evaluate thermodynamic stability, the composition-energy convex hull is generated with pymatgen’s Phase Diagram class [38, 54] utilizing DFT total energies as approximations for 0 K enthalpies. A phase diagram for each substituted candidate is created by querying all entries within the corresponding chemical space from the Materials Project [55]. In the right graph of Figure 4a, we show the energy above the convex hull (Ehull) for each candidate [54, 55]. Fifteen phases shown in blue (inclusive of NTP), are within 30 meV of the convex-vesm hull. While there is no precise energy criterion to separate synthesizable and nonsynthesizable phases, previous work has indicated that many structures with Ehull<30 can be synthesized [56, 57]. One composition, Na4HfP2O9, is on the hull. The remaining nine out of twenty-five phases with Ehull>30 meV/atom (shown in red) are excluded from future screening as it is less likely they are synthesizable.
Figure 4. Results from high throughput computational screening for candidate structures created through substitution of the A site cation. a) left panel: the reaction energy to create the NAP compound from the most stable binary A oxide, sodium carbonate, and ammonium phosphate for (meta)stable phases; right panel: energy above the convex hull. b) Room temperature extrapolated conductivities calculated with MLMD plotted versus ionic radii for (left) 4+ and (right) 5+ cations substituted into the NAP structure. c) Activation energies calculated with MLMD plotted versus ionic radii for 4+ cations.
Na-ion conductivity for each phase is calculated from molecular dynamics simulations with machine-learned potentials created with fine-tuned CHGNet [46] between T=300 K to 1,000 K. Room temperature extrapolated (RT) conductivities for (meta)stable phases with 4+ and 5+ cation substitutions are shown in the left and right panel, respectively, of Figure 4b. Calculations associated with 3+ cations, where all sodium Wyckoff sites are filled, exhibit no Na-ion hopping during the time of the simulation. For example, the probability density and number of hops for Na5AlP2O9 is shown in the Supplemental at 1000 K. No hoping is observed for up to 2 ns, indicating that it is likely an ionic insulator. We find that 5+ cations have high predicted room temperature conductivities: 11.1 mS/cm (Na3TaP2O9) and 12.4 mS/cm (Na3VP2O9). The RT conductivites are slightly lower for the 4+ cations, between 0.01 and 6.5 mS/cm. We plot the activation energy at 600 K versus ionic radius for 4+ cations in Figure 4c. The increase of activation energy with ionic radius is supportive of a possible pillaring effect, wherein structural barriers to conduction are forced open with bigger cations. Smaller ions (Fe, Ga, Mn) are calculated to have activation barriers approximately 0.1-0.2 eV greater than the larger ions.
References
[1]Huo, H.; Bartel, C. J.; He, T.; Trewartha, A.; Dunn, A.; Ouyang, B.; Jain, A.; Ceder, G. Machine-Learning Rationalization and Prediction of Solid-State Synthesis Conditions. Chem. Mater. 2022, 34, 7323——7336, DOI: 10.1021/acs.chemmater.2c01293.DOI
[2]McDermott, M. J.; Dwaraknath, S. S.; Persson, K. A. A graph-based network for predicting chemical reaction pathways in solid-state materials synthesis. Nat. Commun. 2021, 12, DOI: 10.1038/s41467-021-23339-x.DOI
[3]McDermott, M. J.; McBride, B. C.; Regier, C. E.; Tran, G. T.; Chen, Y.; Corrao, A. A.; Gallant, M. C.; Kamm, G. E.; Bartel, C. J.; Chapman, K. W.; Khalifah, P. G.; Ceder, G.; Neilson, J. R.; Persson, K. A. Assessing Thermodynamic Selectivity of Solid-State Reactions for the Predictive Synthesis of Inorganic Materials. 2023.DOI
[4]Bartel, C. J.; Millican, S. L.; Deml, A. M.; Rumptz, J. R.; Tumas, W.; Weimer, A. W.; Lany, S.; Stevanović, V.; Musgrave, C. B.; Holder, A. M. Physical descriptor for the Gibbs energy of inorganic crystalline solids and temperature-dependent materials chemistry. Nat. Commun. 2018, 9, DOI: 10.1038/s41467-018-06682-4.DOI
[5]Malcolm, W.; Chase, J. NIST-JANAF Thermochemical Tables, 4th Edition; American Institute of Physics, -1, 1998.
[6]Andrea Guinto, G. C. Accurate EDS Fitting. tbd tbd, tbd.
[7]Donohue, P. C. Synthesis, structure, and superconducting properties of new high-pressure forms of tin phosphide. Inorg. Chem. 1970, 9, 335–337, DOI: 10.1021/ic50084a032.DOI
[8]Lv, Z.; Li, R. Synthesis, Crystal Structure, and Characterizations of Two Tantalum Phosphates A3TaP2O9 (A=K, Na). Inorg. Chem. 2022, 61, 13554–13560, DOI: 10.1021/acs.inorgchem.2c02186.DOI
[9]Maximov, B.; Bolotina, N.; Simonov, V.; Petříček, V.; Schulz, H. Structure of a modulated monoclinic phase of Na4TiP2O9. Acta Crystallogr. B Struct. Sci. 1994, 50, 261–268, DOI: 10.1107/s0108768193009917.DOI
[10]T. Takahashi, M. S., K. Kuwabara 1981.
[11]Deng, B.; Zhong, P.; Jun, K.; Riebesell, J.; Han, K.; Bartel, C. J.; Ceder, G. CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nat. Mach. Intell. 2023, 5, 1031——1041, DOI: 10.1038/s42256-023-00716-3.DOI