An end‐to‐end artificial intelligence platform enables real‐time assessment of superionic conductors

Superionic conductors (SCs) exhibiting low ion migration activation energy (Ea) are critical to the performance of electrochemical energy storage devices such as solid‐state batteries and fuel cells. However, it is challenging to obtain Ea experimentally and theoretically, and the artificial intelligence (AI) method is expected to bring a breakthrough in predicting Ea. Here, we proposed an AI platform (named AI‐IMAE) to predict the Ea of cation and anion conductors, including Li+, Na+, Ag+, Al3+, Mg2+, Zn2+, Cu(2)+, F−, and O2−, which is ~105 times faster than traditional methods. The proposed AI‐IMAE is based on crystal graph neural network models and achieves a holistic average absolute error of 0.19 eV, a median absolute error of 0.09 eV, and a Pearson coefficient of 0.92. Using AI‐IMAE, we rapidly discovered 316 promising SCs as solid‐state electrolytes and 129 SCs as cathode materials from 144,595 inorganic compounds. AI‐IMAE is expected to completely solve the challenge of time‐consuming Ea prediction and blaze a new trail for large‐scale studies of SCs with excellent performance. As more experimental and high‐precision theoretical data become available, AI‐IMAE can train custom models and transfer the existing models to new models through transfer learning to constantly meet more demands.


| INTRODUCTION
As an important technical route of the next generation of power batteries, all solid-state batteries (ASSBs) have promising prospects.The main advantage of ASSBs, which contain only solid-state electrolytes (SSEs), is that they are less prone to thermal runaway and are expected to have higher energy densities. 1,2However, in general, ASSBs are still in the development stage worldwide and have not yet been commercialized on a large scale.4][5] Therefore, the development of SSEs and cathodes will restrict the popularization and application of ASSBs.Ideal SSEs and cathodes have the common characteristic that they must be superionic conductors (SCs) that have high ionic conductivity (I C , >10 −4 S/cm) with ion migration activation energy (E a ) <0.4 eV. 3,6I C determines the internal resistance and power performance of a battery, and SSEs with a high I C can decrease the concentration polarization during charge and discharge so as to improve the power density. 4,7Cathodes with a high I C possess a high current density and provide rapid charge and discharge characteristics with good reversibility for batteries. 8,9Consequently, identifying materials with high I C (or low E a ) is the key approach to the development of ASSBs.That is, determining the I C or E a for a material is required.
Experimentally, temperature-dependent alternating current (AC) impedance spectroscopy, 10 spin-lattice relaxation (SLR) NMR, 11,12 and operando neutron depth profiling (NDP) 13 are mainly used to measure I C .The results are accurate, but their premises are the synthesis and characterization of materials, electrode fabrication, and a series of high-cost, long-cycle complex technologies.In theory, first-principles calculations based on density functional theory (DFT) can reveal microscopic viewpoints on ion migration and atomic shuttles, providing a relatively efficient approach for obtaining I C .Ab initio molecular dynamic (AIMD) simulations are conducted using the potential energy and forces with ab initio methods. 14The wide applicability of AIMD simulations has been successfully demonstrated in studies of ion migration, 15,16 thermal stability, 17,18 and so forth.3][24][25] These calculations can provide a reference before experiments, but they require a variety of computing techniques, which makes the methods not suited for large-scale development, especially when faced with hundreds of thousands of candidates.
7][28] For example, Sargent et al. 29 identified Cu-Al electrocatalysts using DFT calculations and active ML and reduced CO 2 to ethylene with the highest Faradaic efficiency reported thus far, which demonstrates the value of computation and ML in guiding experimental research.Deringer et al. 30 presented an ML model trained on quantum mechanical computations that can help to describe liquid-amorphous and amorphous-amorphous transitions for a system with 100,000 atoms, and the results shed light on the great potential of ML in material modeling.In a recent study, Shi et al. 31 introduced a new framework of ML prediction for E a with hierarchically encoding crystal structure-based descriptors and took cubic Li-argyrodites as a case study.The framework achieved a high coefficient of determination of 0.82, which inspired us to predict E a using ML.However, to date, ML models that can directly predict E a for arbitrary crystal structures are rarely reported so far, which seriously hinders the design and discovery of SSEs and cathode materials and restricts the development of electrochemical devices such as ASSBs.Thus, it is imperative to establish a precise and efficient method for ion migration E a prediction.
Here, we developed an AI-IMAE (artificial intelligence-ion migration activation energy) platform for predicting ion migration E a for all inorganic compounds in the Materials Project (MP) database.As a typical case, we collected and established a database containing a total of 23,368 potential cation and anion conductors, 32 including Li + , Na + , Ag + , Al 3+ , Mg 2+ , Zn 2+ , Cu (2)+ , F − , and O 2− , where the E a were calculated by CAVD and BVSE methods.For each ion, we designed and trained a specific crystal graph neural network (NN) model for E a prediction; and these models achieve a holistic average absolute error (AAE) of 0.19 eV, a median absolute error (MAE) of 0.09 eV, and a Pearson coefficient of 0.92.We further demonstrated the prediction accuracy of the AI-IMAE platform via known experimental structures.From 144,595 inorganic compounds in the MP database, 33 we discovered 2 compounds containing Mg These candidates all exhibit great thermal stability, high I C , suitable electronic conductivity, and good elastic properties.The overall process of this work is presented in Supporting Information: Figure S1.Although the data set we selected was derived from empirical calculations, the proposed work demonstrates the potential of the AI method for rapid E a prediction, which opens a new path for large-scale research on the ion migration of inorganic compounds and will promote a new wave of research on SSEs and cathode materials.In the future, with the emergence of some experimental or high-precision computational E a databases, users can replace the empirical database in this work with new databases, and use AI-IMAE to obtain more accurate models and prediction results.

| Crystal graph-based deep learning model
In this study, the models are designed and trained using a crystal graph convolutional neural network (CGCNN), which is a scalable and efficient tool for predicting material properties.The main advantage of the CGCNN is that it can extract features for different structures with different spatial groups or atomic stacks, and it has been applied in many studies.In the CGCNN, atoms served as nodes, bonds between the neighboring atoms served as edges, and we searched the neighbors within a 12 Å radius that maximally connected 12 nearest neighbors.Each node is represented by the vector v, which is a one-hot encoding based on the atomic properties (i.e., electron affinity, electronegativity, and covalent radius).Each edge is expressed by the vector u(i, j), where each member is the distance between atoms i and j.Then, the feature vector of atom i (v i ) is updated by a convolution function to learn the final feature vector, including its surrounding environment: ( ) After t convolutions, all feature vectors are calculated by a pooling layer to produce the overall feature vector V c for a crystal.Then, fully connected layers are used to capture the V c between the output attribute E a .Overall, this can be turned into an optimization problem: where W is the weights in NN layers; J is the loss function between the true values and prediction values; and f denotes the function between CGCNN outputs and crystal C. In this study, to choose the best model for each ion, we apply a training-validation strategy to optimize the prediction accuracy.During 1000 iterations, each model is trained with 80% of the data and validated with 10% of the data, and the model with the best validation performance is selected and tested with the remaining 10% of the data.The same key hyperparameters optimized for the nine models are the number of hidden atomic features in the convolutional layer (128), the number of hidden features after pooling (64), the number of convolutional layers (2), and the number of hidden layers after pooling (2).Since the scales of the data of the 9 ions vary, the optimized hyperparameters of the learning rate and batch size during the training process are different (see Supporting Information: Table S1).AI-IMAE provides a convenient and user-friendly E a prediction platform that requires no programming knowledge.Users only need to input a crystallographic information file (CIF) and set several calculation conditions to obtain the corresponding E a , and the predictive process takes only ~0.005 s.
The interface of AI-IMAE contains seven panels: Input, Crystal Information, Crystal View, Calculation Conditions, Run AI-IMAE, Model Optimization, and Output.The Input panel presents a file-browsing function where users can upload CIFs.Crystal Information, such as the lattice constants, space groups, volumes, elements, and so forth, can be obtained from CIFs.The Crystal View panel provides a crystal visualization function to help users more intuitively view the atomic arrangement, and the function includes three modes: ball and stick, polyhedral, and space-filling.Then, users need to set the required parameters in the Calculation Conditions panel before running the program.The transport ion should be determined and unique.For example, if we want to predict the E a of Li + in LiFePO 4 , the well-trained model of Li + should be loaded; furthermore, if we want to predict the E a of the O 2− in LiFePO 4 , the well-trained model of O 2− should be loaded.If the uploaded CIF does not contain the given ion, the program will report an error.The temperature, hop distance, lattice vibrational frequency, and Boltzmann constant are the parameters used to calculate the diffusion coefficient (D) and I C (see Supporting Information: Note S1 for the calculation details).The first two parameters are set by the users, and the last two parameters are generally constants.Next, users can click the "Run AI-IMAE" button to run this prediction task, and the results (E a , D, and I C ) will be given in the Output panel.A prompt of "Superionic conductor!" will appear on the right when the predicted E a is less than 0.4 eV.The relationships among the seven panels are clearly shown in Figure 2.
"Run AI-IMAE," which is based on the nine welltrained NN models, is an execution button.Figure 2 shows the workflow of the AI-IMAE software, including the seven panels shown in Figure 1 and the building process of the NN models.To build well-trained NN models, 23,368 crystal structures covering nine ions were collected and constitute the initial database for AI-IMAE.For each structure, a specific crystal graph was formed to represent the atomic arrangement, and a one-hot encoding graph vector was calculated based on the component and structural information (see Section 2).The graph vectors served as inputs for the NN model, and the corresponding E a served as output attributes.The data set was divided into training, validation, and testing sets, where the first two were used in this stage to determine the best model.The mathematical relationship between graph vectors and E a was captured using convolutional network layers, pooling network layers, and fully connected network layers, resulting in welltrained models.
In addition to the function of E a prediction, AI-IMAE has a panel of model optimizations to train custom models (see Figure 2).As the source database grows, users can upload a new source database of target ions from high-precision calculations and experiments and train custom models.In this process, users need to select a specific ion as the research target and upload the corresponding data (CIFs and E a ).In the "Run AI-IMAE" panel, the "optimizer" corresponds to the optimization algorithm in ML (i.e., Adam or SGD), which is used to minimize the loss function.
Interface of the AI-IMAE platform.Seven panels including "Input," "Crystal Information," "Crystal View," "Calculation Conditions," "Run AI-AIME," "Model Optimization," and "Output" are designed.The Software Copyright has been licensed in China for AI-IMAE with No. 2021SR1336972.AI-IMAE, artificial intelligence-ion migration activation energy.than replace).The figures produced from the model (i.e., training error at each iteration and correlation plots) can be exported by clicking "Export Figures," and the log file can be exported by clicking "Export Log."All exported files will be saved in the assigned path in the Output panel.
In addition, users can choose the "Transfer Learning" function to further optimize the well-trained models.Transfer learning (TL) is an ML strategy (pretraining + fine-tuning) that can first build a "source model" on a large data set (sufficient for both inputs and outputs), store its model parameters, and use them as initializations for another "target model" with a small data set (small for both inputs and outputs).Then, by fine-tuning the parameters of the "target model", a model with similar accuracy as the "source model" is potentially obtained.This reduces the cost of data acquisition for the "target model."TL has been used in NNs (CGCNN-TL) for band gap (E g ) prediction from low-level Perdew-Burke-Ernzerhof (PBE) to high-level Heyd-Scuseria-Ernzerhof (HSE06) functionals, 34 and applied in adsorption energy prediction between different adsorbents and heavy metal ions. 35These works inspire us to apply TL to predict E a from calculation to experimental results in the future.This process takes the initial model as the starting point and then fine-tunes it to obtain the final model.Therefore, the learning rate should not be set too large Input and Calculation conditions are required for execution, and the Crystal view and Crystal information will be obtained automatically.Well-trained neural network models are loaded by clicking "Run AI-IMAE."Models can be further optimized as the source database grows.AI-IMAE, artificial intelligence-ion migration activation energy.so as to avoid model oscillation.TL can effectively overcome the problem of data scarcity and has great potential in transferring from theory to experiments. 36

| AI-IMAE: Database, model, and evaluation
AI-IMAE was implemented by a combination of highthroughput calculations and NNs.A material database of 29,268 potential cation and anion conductors, including crystal structures and E a with a range of 0.0-20.0eV, is collected from a high-throughput calculations platform (SPSE). 25,32SPSE provides three main advances: (1) Geometric analysis is combined with the BVSE method to rapidly simulate the path and energy profile; (2) highthroughput hierarchical screening for SCs; and (3) database containing ion migration properties.This database contains crystal structure information, ion migration channel connectivity information, and 3D channel map for ionic conductors.The Crystal Information is described in CIF format, which can be extracted by the Python Materials Genomics (Pymatgen) API 37 and served as the input for our NN models.The corresponding E a used as the output for model training was calculated using the combination of CAVD 38 and BVSE 39 methods.CAVD builds a network of interstices in a fixed frame, which consists of interstices and channel segments.The narrow point in the channel segment represents the bottleneck.The radius of the migrating ions, the gap, and the size of the bottleneck can be used to identify and rank suitable frameworks for ion migration.BVSE is based on the principle of local electroneutrality, according to which the oxidation state of an ion should be close to the bond valence.By taking into account a Morse-type potential term for cation-anion pairs and Coulomb repulsions between ions and ions of the same charge sign, BVSE extends the bond valence sum method.Thus, the approximate minimum energy path for migrating ions can be found.This geometric analysis and bond valence method was based on empirical computations, and the calculation time for a single structure is a few minutes on supercomputers. 40Many previous studies demonstrate that BVSE calculations are in good agreement with the experimental measurement and NEB results.For example, the E a of LiI calculated by BVSE decreases is 0.55 eV, which is consistent with the experimental results of 0.43(4) eV. 41Kim et al. 42 reported the E a of Li 3 YbCl 6 and Li 2.7 Yb 0.7 Zr 0.3 Cl 6 , to be approximately 0.7 eV and 0.6 eV, respectively, which has a good agreement with the experimental results of 0.5 eV and 0.3 eV; they conclude that the BVSE method allows for an approximate assessment of the relative heights of the barriers.The E a of Li 4 PS 4 I from the BVSE method also closely agrees with experimental data from NMR (0.25 eV close to 0.23 eV) and AC impedance spectroscopy (0.39 eV close to 0.40 ± 0.03 eV), and the corresponding conclusions are also reflected in Li 15 (PS 4 ) 4 Cl 3 . 43,44He et al. 45 proposed an informative method to identify interstices and connecting segments constructing an ion transport network.For LiScGeO 4 , the CAVD-BVSE-based E a (0.39 eV) is similar to NEB-based E a (~0.16 eV); and the E a of 0.51 eV and ~0.37 eV are also obtained for LiLaTiO 4 by two methods.][48] In addition to data reliability, the equilibrium of data distribution is also the key to NN models.Structures with E a less than 0.4 eV are promising SCs; thus, structures with E a far greater than 0.   3 shows the histograms of the E a of nine ions with interval steps of 0.2 eV, where the data points are almost continuous without breakpoints and each step line is relatively smooth without a dramatic rise.These good data distributions and suitable E a ranges help us to predict E a more accurately.Box plots of the E a of nine ions are provided in Supporting information: Figure S2.
0][51] The main advantage of the CGCNN is that features of different structures with different space groups or atomic stacks can be extracted and used to predict the material properties of arbitrary crystal structures.In this study, we adopted a training-validation framework to optimize the model and used the testing set to evaluate the generalization ability of the best model.The ratio of the training, validation, and testing sets was 8:1:1 (see Table 1).Figure 4A shows the AAE of the training set (blue curve) and validation set (orange curve) during 1000 iterations for 9 NN models.Figure 4A shows that all curves tend to be horizontal, indicating that the models have reached a stable state.Moreover, the AAE differences between the training and validation sets are small, which makes it unnecessary to worry about the overfitting problem.The model with a minimum validation AAE during 1000 iterations was selected as the best model and further applied to the testing set.In addition to the AAE, we added two other indicators to measure the final accuracy for each model: the MAE and the Pearson correlation coefficient (R p ).The former is a robust statistic that can statistically accommodate outliers in a data set, and the latter R p is used to measure the correlations between true and predicted values.Figure 4B shows the AAEs, MAEs, and R p in the testing set of nine well-trained models.The holistic average AAE (green histogram) of the nine models is 0.19 eV with a maximum of 0.27 eV and a minimum of 0.16 eV, the MAE (purple histogram) is 0.09 eV with a maximum of 0.14 eV and a minimum of 0.06 eV, and the R p (light red histogram) is 0.92 with a maximum of 0.94 and a minimum of 0.89.The AAEs, MAEs, and R p in Figure 4B demonstrate the high accuracy of each model.The AAEs, MAEs, and R p on the training and validation sets are provided in Table 1. Figure 4C plots the correlations of the predicted E a and calculated E a , the data points of the training (blue circles), validation (green triangles), and testing (plum rectangles) sets are evenly distributed on the diagonal, and only a very small fraction of them have allowable differences, indicating that there is no systematic deviation of the models.More specifically, it is noted that for the Zn 2+ and F − models, the differences between the calculated E a and predicted E a are obvious when E a is greater than approximately 3.0 eV, which is attributed to the distribution of their data.However, this has little impact on predicting E a since the desired E a is less than 0.4 eV, and structures with predicted E a much higher than 0.4 eV wound be excluded.Figure 4D shows the box plots in the testing set for the nine models.The above analysis again proves the high accuracy of the models; thus, these models can be used to predict the E a of nine ions for given structures.
As a mainstream research object, many studies have provided experimental or theoretical analyses of Li + SCs.3][54][55][56][57][58][59][60][61][62][63][64] Figure 5A gives the comparison of the experimental results (gray) and AI-IMAE predictions (blue) for 12 Li + structures and shows that the deviation between the experiments and AI-IMAE predictions is slight and allowable.Table 2 shows that there is a minimum deviation of 0.03 eV (Li 3 GeS 4 ) and a maximum deviation of 0.28 eV (LiZr 2 P 3 O 12 ).Similarly, Figure 5B shows the comparison of the calculations (gray) and AI-IMAE predictions (blue) for the other five structures.The minimum deviation was 0.065 eV (Li 2 TiF 6 ), and the maximum deviation was 0.23 eV (LiYF 4 ).The overall average deviation between AI-IMAE and the experimental and computational results is 0.13 eV, which guarantees the accuracy of the discovery and design of SCs using AI-IMAE.Since the structures or E a of other ions have not been extensively studied, and some materials with experimental E a lack accurate crystal structure information due to doping or recombination, these structures cannot be validated and compared with AI-IMAE.According to the model accuracy in Tables 1  and 2, AI-IMAE can also provide a reliable prediction platform for screening the corresponding SSEs or cathode materials.It is worth mentioning that the proposed AI-IMAE software was modeled based on semiempirical data due to the lack of a large number of high-precision E a .But as a typical case, AI-IMAE provides us a way to predict E a of arbitrary crystal structures.With the expansion and improvement of data in the future, such as the emergence of experimental or high-precision theoretical databases, AI-IMAE will be adapted and play to greater advantage.Furthermore, the calculation of an E a will take minutes for a single crystal using the empirical method, whereas AI-IMAE takes only milliseconds, resulting in a speed of ~10 5 times faster.

| AI-IMAE: Applications
Although CAVD and BVSE methods can calculate E a for a single crystal in a matter of minutes, 40 they still inevitably bring high costs and computational cycles for large amounts of data (>10 5 crystals).In this section, we used the well-trained AI-IMAE software to predict the E a of all inorganic compounds (144,595 materials, date: 01.08.2021) in the MP database 33 and discovered the best   (A) (B) SCs containing target ions (Li + , Na + , Ag + , Al 3+ , Mg 2+ , Zn 2+ , Cu (2)+ , F − , and O 2− ) as promising SSEs and cathode materials.Both SSEs and cathode materials require low ion migration E a and good thermal stability.The former is related to the performance of electrochemical equipment 3,4 while the latter ensures that the material is stable and easy to synthesize.The difference is that SSEs require low electronic conductivity (E C < 3.52 × 10 −5 S/cm, or wider E g , >1.0 eV, see Supporting Information: Note S2) to ensure that the electrons across the electrolyte are negligible and avoid the occurrence of a self-discharge process and the formation of lithium dendrites 3,65,66 ; conversely, cathode materials should have a narrower E g to improve the E C and maintain a good charge-discharge performance of the electrode materials. 8,9Moreover, the elastic properties of the SSEs and cathode materials are crucial to the stability of electrochemical devices.The elasticity model proposed by Monroe and Newman suggests that SSEs with a shear modulus (G) larger than ~8.5 GPa can suppress lithium dendrite on anodes, 67,68 which can also be obtained from the MP database according to the material ID.Similarly, shear deformations easily occur in cathode materials, resulting in poor cycling performance. 69,70Therefore, in this study, we considered four property criteria (I C , thermal stability, electronic conductivity, and shear modulus) of materials to screen SSEs and cathode materials, as shown in Figure 6A.
Figure 6A shows the workflow and screening criteria of good SCs for SSEs and cathodes.From the 144,595 inorganic materials in the MP database, 33 we selected 143,611 compounds containing nine target ions as the initial database.Using AI-IMAE, the E a of the 143,611 materials can be predicted quickly and accurately, and compounds with an E a < 0.4 eV are selected.Since the E a calculation is a very time-consuming work, the screening scope can be greatly narrowed in this step (see Supporting Information: Figures S5 and S6 for details).Afterward, the thermal stability was measured by the energy above the hull (E hull ), which is calculated by using the formation energies of all competing phases and is usually used to quantitatively evaluate the thermal and phase stability of materials.Structures with E hull < 0.1 eV/atom are considered structurally stable and are promising for synthesis. 71Then, compounds with an E g > 1.0 eV were further selected as potential SSE candidates, and compounds with an E g < 1.0 eV were possible cathodes.Finally, SSE candidates with a G > 8.5 GPa and an E g > 1.0 eV were screened out to be good SSEs, and cathode candidates with a G > 28.3 GPa (G = 28.3GPa corresponds to a typical α-Li 2 MnP 2 O 7 cathode 70 ) and an E g < 1.0 eV were selected as good cathodes.For all candidates, we excluded the intermetallic compounds (e.g., Li 2 ZnGe, LiLuHg 2 ) that could not be considered electrolyte and electrode materials. 4,72,73Such a rigorous screening process ensures that the screened SSEs and cathodes have good stability, conductivity (ionic and electronic), and mechanical properties.
Figure 6B shows the polar histogram that gives the number of SC candidates used for SSEs.In the figure, 316 SCs  S6).Specifically, we plotted the scatter map for the 40 Li + SSE candidates by E a , E g , and G, as shown in Figure 6D.LiLuO 2 has an E a of 0.229 eV, an E g of 4.79 eV, and a G of 51 GPa, and LiMgH 3 has a lower E a of 0.159 eV, an E g of 4.04 eV, and a G of 32 GPa.The scatter map for 38 Li + cathode candidates by E a , E g , and G was plotted in Figure 6E.LiTiO 2 has an E a of 0.159 eV, an E g of 0.0 eV, and a G of 92 GPa.All information on promising SSE and cathode candidates, including the formula, E a , E g , and G, are listed in Supporting Information: Tables S2-S19, which provides the optimal solid electrolyte and electrode materials for all nine ions (Li + , Na + , Ag + , Al 3+ , Mg 2+ , Zn 2+ , Cu (2)+ , F − , and O 2− ).The scatter maps of the other eight ions as SSEs and cathode materials can be found in Supporting Information: Figures S8-S15.It is noting that Ceder et al. 74 have experimentally proved that MgY 2 Se 4 is a good SC that can potentially integrate with current state-of-the-art Mg cathodes.It demonstrated that the model we trained is helpful in the screening of E a of unknown materials.
Moreover, we made a comparison of commercially available cathode materials with predicted candidates.The cathode material LiCoO 2 is layered rock salt and has an E g of 2.7 eV, spinel LiMnO 2 has an E g of 1.64 eV, and olivine LiFePO 4 has an E g of 3.9 eV. 33,75Currently, LiNi x Mn y Co z O 2 (NMC, x + y + z = 1) is the state-of-theart choice of cathode materials for Li-ion batteries in the electric vehicle application, the E g of which closes as one introduces more Ni (from ~0.5 eV to 0 eV). 76hese indicate that our predicted cathode materials can match or even exceed commercial ones in electronic conductivity.In addition to E g , the minimum value G we determined also has advantages.Previous studies show that LiFePO 4 has a G of approximately 46 GPa, 33,69 and Sun et al. 76  3 GPa) of G we chose is slightly lower than that of current commercial cathode materials, it was to focus more attention on the discovery of SCs.In addition, among the promising cathode materials we found, there are also many materials with G > 46 GPa (see Supporting Information: Tables S2-S19), such as Li 2 NiO 2 (73 GPa), LiTiO 2 (92 GPa), CaAgO 2 (61 GPa), and CeAlO 3 (93 GPa), which are comparable to commercial cathode materials.
In the future, researchers can use these data to directly select the desired materials for further experimental verification and theoretical design.The screening processes of the other eight ionic conductors are provided in Supporting Information: Figures S7-S15, and more details in Supporting Information: Note S3.Due to the framework of crystal graph NN, AI-IMAE is also suitable for doping structure prediction in all crystal files with the standard crystal file format and normal oxidation state theoretically.The greatest advantage of the AI-IMAE software is that it predicts E a with high precision and takes only ~0.005 s on a single CPU (1.39 × 10 −6 h) for a predictive task.Compared with the calculation time of a few minutes (~0.1 h) for the CAVD and BVSE method, 40 AI-IMAE speeds up the calculation by a factor of ~10 5 .Supporting Information: Figures S5 and S6 show the screening process of SSEs and cathodes.The figures show that AI-IMAE predicts the E a of 143,611 compounds (including 20,237 Li + , 8456 Na + , 3846 Ag + , 7063 Al 3+ , 9163 Mg 2+ , 5684 Zn 2+ , 8910 Cu (2)+ , 10,700 F − , and 69,552 O 2− ) in just 0.2 CPU hours while CAVD and BVSE method takes approximately 1.4 × 10 4 h or ~600 days to complete.Therefore, as for the SSEs and cathode materials studied in this work, AI-IMAE has shown great advantages in terms of time and costs, laying a solid foundation for the large-scale design and discovery of SCs.Importantly, AI-IMAE also foreshadows the application of TL for high-precision data expansion, such as AIMD-based or experimental E a , where more screening costs will be saved and higher precision achieved.We hope that the proposed AI-IMAE software can be used in more applications, not only in solid electrolyte and cathode materials in batteries but also in other energy and material fields, in the future to provide quick and accurate guidance for the discovery of good SCs.

| CONCLUSION
Aiming at the problem that ion migration characteristics are difficult to determine experimentally and theoretically, an end-to-end AI platform AI-IMAE was proposed in this study to cut the computing costs of screening SCs to a minimum while reducing the risk of prediction failure.Here, NN models were trained to make accurate predictions of ion migration E a for all inorganic compounds.We collected a high-throughput database that covers Li + , Na + , Ag + , Al 3+ , Mg 2+ , Zn 2+ , Cu Importantly, our study provides a general method for predicting the E a of materials; as increasingly more highprecision training data (e.g., DFT level and experimental level) become available, users can apply AI-IMAE to train custom models and use TL to transfer the existing models to new models so that the AI-IMAE software can constantly meet more demands.We believe that future work could address the experimental accuracy of applications or extend the above efforts to design and discover more SCs with good performance.

Figure 1
Figure1presents the interface of the AI-IMAE software, which was based on nine well-trained NN models.AI-IMAE provides a convenient and user-friendly E a prediction platform that requires no programming knowledge.Users only need to input a crystallographic information file (CIF) and set several calculation conditions to obtain the corresponding E a , and the predictive process takes only ~0.005 s.The interface of AI-IMAE contains seven panels: Input, Crystal Information, Crystal View, Calculation Conditions, Run AI-IMAE, Model Optimization, and Output.The Input panel presents a file-browsing function where users can upload CIFs.Crystal Information, such as the lattice constants, space groups, volumes, elements, and so forth, can be obtained from CIFs.The Crystal View panel provides a crystal visualization function to help users more intuitively view the atomic arrangement, and the function includes three modes: ball and stick, polyhedral, and space-filling.Then, users need to set the required parameters in the Calculation Conditions panel before running the program.The transport ion should be determined and unique.For example, if we want to predict the E a of Li + in LiFePO 4 , the well-trained model of Li + should be loaded; furthermore, if we want to predict the E a of the O 2− in LiFePO 4 , the well-trained model of O 2− should be loaded.If the uploaded CIF does not contain the given ion, the program will report an error.The temperature, hop distance, lattice vibrational frequency, and Boltzmann constant are the parameters used to calculate the diffusion coefficient (D) and I C (see Supporting Information: Note S1 for the calculation details).The first two parameters are set by the users, and the last two parameters are generally constants.Next, users can click the "Run AI-IMAE" button to run this prediction task, and the results (E a , D, and I C ) will be given in the Output panel.A prompt of "Superionic conductor!" will appear on the right when the predicted E a is less than 0.4 eV.The relationships among the seven panels are clearly shown in Figure2."RunAI-IMAE," which is based on the nine welltrained NN models, is an execution button.Figure2shows the workflow of the AI-IMAE software, including the seven panels shown in Figure1and the building process of the NN models.To build well-trained NN models, 23,368 crystal structures covering nine ions were collected and constitute the initial database for AI-IMAE.For each structure, a specific crystal graph was formed to represent the atomic arrangement, and a one-hot "Jobs" denotes the number of training hardware devices (i.e., 24 CPUs or 5 GPUs).Then, the user can split the training data into training, validation, and testing sets at a given ratio (i.e., 0.8:0.1:0.1 or 0.7:0.15:0.15)or give the number of sets (i.e., 500:50:50)."Learning Rate" and "Batch Size" are the hyperparameters of the optimization model in ML that are used to adjust the optimization process.Next, users can set the maximum training time in hours or set the maximum number of iterations to limit the training time and costs."Start Training" refers to the beginning of training the model, and "Stop Training" corresponds to the end of the training process.Users can click "Replace Model" to replace the previous model well-trained on the target ion or "Export Model" to save the model (rather 4 eV (E a >> 0.4 eV) are redundant samples in the NN model.In this study, considering the size and distribution of the data, we selected structures with E a less than 4.0 eV (0.0-4.0 eV) to constitute the training, validation, and testing sets for NN models.Concretely, from the 29,268 potential cation and anion conductors, we collect 23,368 cation and anion conductors with E a between 0.0 and 4.0 eV, including 2123 Li + conductors, 2294 Na + conductors, 904 Ag + conductors, 1342 Al 3+ conductors, 945 Mg 2+ conductors, 1055 Zn 2+ conductors, 2231 Cu (2)+ conductors, 1390 F − conductors, and 11,084 O 2− conductors. Figure Data distribution of 23,368 cation and anion conductors.(A) Histogram of the E a distribution of 2123 Li + conductors.(B) Histogram of the E a distribution of 2294 Na + conductors.(C) Histogram of the E a distribution of 904 Ag + conductors.(D) Histogram of the E a distribution of 1342 Al 3+ conductors.(E) Histogram of the E a distribution of 945 Mg 2+ conductors.(F) Histogram of the E a distribution of 1055 Zn 2+ conductors.(G) Histogram of the E a distribution of 2231 Cu (2)+ conductors.(H) Histogram of the E a distribution of 1390 F − conductors.(I) Histogram of the E a distribution of 11,084 O 2− conductors.The step lines represent the cumulative percentages.E a , activation energy.

T A B L E 1
Date divisions, AAEs (eV), MAEs (eV), and R P of the training, validation, and testing sets for nine well-trained NN models.

F I G U R E 4
Error analysis of nine deep learning models.(A) AAE curves of the nine models on the training (blue) and validation (orange) sets during the iterations.(B) AAEs, MAEs, and R p of the nine well-trained models on the testing sets.(C) Correlation plots of the predicted E a and calculated E a for the nine well-trained models on their training (blue circle), validation (green triangle) and testing (plum rectangles) sets.(D) Box plots of the predicted E a (yellow) and calculated E a (green) for the nine well-trained models on the testing sets.AAE, average absolute error; E a , activation energy; MAE, median absolute error.

6
Screening of SSEs and cathode materials using AI-IMAE.(A) Workflow and screening criteria of good SCs as SSEs and cathodes.(B) Polar histogram, which shows the numbers of screened SC candidates of nine ions for SSEs.(C) Polar histogram, which shows the numbers of screened SC candidates of nine ions for cathodes.(D) Scatter map of 38 Li + SCs as SSE candidates plotted by E a , E g , and G. (E) Scatter map of nine Li + SCs as cathode candidates plotted by E a , E g , and G.The color bars denote the scale of E g .AI-IMAE, artificial intelligence-ion migration activation energy; E a , activation energy; SC, superionic conductor; SSE, solid-state electrolytes.
(2)+ , F − , and O 2− and establish nine NNs for their predicted E a , achieving a holistic AAE of 0.19 eV, an MAE of 0.09 eV, and a Pearson coefficient of 0.92.Based on the nine well-trained NN models, we actualized the AI-IMAE interactive platform for the quick and accurate prediction of E a .From 143,611 inorganic compounds, the proposed AI-IMAE discovers 2 C Mg , 6 C Na , 6 C Al , 8 C Zn , 26 C Ag , 38 C Li , 40 C Cu , 94 C F , and 96 C O as SSE candidates and 1 C Mg , 2 C Na , 2 C Zn , 6 C Al , 9 C Li , 17 C F , 17 C Ag , 24 C Cu ,and 51 C O as cathode material candidates.The screened SSEs and cathode materials have good thermodynamic stability, desired electronic conductivity, I C , and elastic properties.The AI-IMAE platform not only reduces the traditional time for calculating and screening good SSE and cathode materials from the MP database by approximately 600 days but also blazes a new trail for the large-scale study of SCs, which will greatly promote the development of electrochemical energy storage devices.
Comparison of the AI-IMAE predicted, experimental and NEB calculated E a .(A) Experimental results (gray) versus AI-IMAE predictions (blue).(B) Calculation results (dark green) versus AI-IMAE predictions (blue).For E a with no definite value, we take the mean as its value.For example, the E a of Li 10 SnP 2 S 12 is given by (0.27 + 0.6)/2 = 0.435 eV.AI-IMAE, artificial intelligence-ion migration activation energy; E a , activation energy; NEB, nudged elastic band.
T A B L E 2 E a Comparison of AI-IMAE predictions, experiments (Exp, E ), and NEB calculations (Cal, C ) for 17 known structures.Note: The deviations are the absolute differences between the AI-IMAE predictions and experimental or calculated results.Average deviation: 0.13 eV.Abbreviatiions: AI-IMAE, artificial intelligence-ion migration activation energy; NEB, nudged elastic band.C Obtained by NEB calculations.E Obtained by experiments.