All ML is data-dependent and, therefore, essential for building the right ML model dataset^[13]. In addition to the amount of data, the quality of the data is equally important^[14]. Any situation such as small amount of data, low quality of data, data errors, and data sets that are difficult to replicate will seriously affect the prediction results of ML, resulting in unexplained deviations or even errors in the relevant conclusions^[15]. Therefore, the first step in developing a reliable ML model is to construct a high-quality dataset. At present, the rise of first principles and the establishment of materials databases provide ML with access to high-quality data sets, while computational or experimental methods have accumulated a large number of material property data.However, most of these data are not uniform or incomplete, so before establishing the initial data set, we should first clarify the research problem, preprocess it through data cleaning or artificial screening, and establish a high-quality data set that meets the target problem. Similar to experimental measurements, ML itself should follow good standards and protocols to reduce bias in data processing and prediction.

As the input variables of ML model, descriptors contain most of the physical/chemical properties of materials, which determine the efficiency of ML model and the final results, especially in large-scale experiments. The conversion of descriptors is another important problem in the application of ML in LBs. It is necessary to convert some characteristics and attributes of materials into vectors composed of numbers that can be understood by computers. At present, there are two methods to achieve this: (1) by artificially converting and characterizing some complex material properties (such as atomic coordination number and crystal structure) into digital descriptors through Python library tools (Matminer); (2) using deep learning methods. Tian et al. Proposed a generalized crystal graph convolutional neural network (CGCNN) framework (Fig. 1), which constructs a convolutional neural network directly on the crystal graph generated by the crystal structure by studying the simplest form of crystal representation, that is, the connection of atoms in the crystal.Then after training on the data of using materials project, CGCNN achieves similar accuracy as DFT in DFT calculation, and finally compared with the experimental data of eight different attributes, which shows the generality of the method^[16]. There are many possible descriptors/parameters to be considered in the ML process, and if the training data set is small at this time, it will lead to overfitting. Therefore, future ML research directions in the area of LBs should focus on developing algorithms that specifically target small datasets (e.g., hierarchical ML, reinforcement learning, etc.)^[17].

A good functional form that maps material properties to material descriptors (representing structure-property relationships) helps to find new materials with potential properties. From the experimental point of view, in traditional experimental research, the identification of descriptors mainly depends on physical intuition and chemical theory for artificial screening, while with the improvement of computing power, the optimization of ML algorithm and the promotion of massive material data, a large number of relevant information can be captured, and the features can be extracted as part of the model^[18]. The whole process can be summarized as obtaining the original input, that is, according to the needs of experimental design, obtaining the required data by various means, such as atomic radius, band gap, valence and other related target attributes, and sorting them into tabular form for subsequent processing, and then converting the original input into descriptors or features that can be learned by ML algorithm.With the help of Python Materials Genomics (pymatgen), it can download the required material structure and related properties in batches, and also provide a large number of data processing and calculation after VASP calculation. Finally, the ML model is trained on the data, and the performance of the model is drawn and analyzed. Ideally, descriptors should be irrelevant, as a large number of relevant features can hinder the efficiency and accuracy of the model.

Traditional computational simulation methods, including theoretical calculation and simulation techniques based on DFT, MD, Monte Carlo (MC) and phase-field methods, are used to explore more microscopic phase and spatial components based on the microstructure of materials, and to test whether the conclusions obtained from experiments can be explained by microscopic processes^[19]. It can be said that the traditional computational simulation method is a bridge between material theory and experiment. In contrast, ML can be seen as a powerful complement to traditional computational methods. Its main task is to determine existing physical laws or undiscovered rules based on data and rules, and to build predictive models by evaluating data sets, so as to help researchers understand and analyze the key factors affecting model output.

However, each means of computational simulation is limited. For example, the scale of DFT simulation is small, and the simulation structure error is large in the environment of high temperature, high pressure and strong magnetic field. MD simulation relies on a more accurate potential function, which has the problem of insufficient sampling in the ensemble. Dierk's group proposed a machine learning-based high-entropy alloy design method, starting from experiments and fully combining ML, DFT, thermodynamic calculations and experiments to find two new invar alloys from millions of candidates^[20]. At the same time, this method also proves the feasibility of using sparse experimental data, which greatly reduces the time of traditional material design methods.

Experiment, simulation and data-driven tools have become three indispensable methods in current scientific research, and have shown great potential in the field of battery materials by combining ML methods with traditional computational simulation methods, rationally dividing the work, and giving full play to the ability of ML to analyze and process data.

Supervised learning is the most widely used and effective tool in LBs research. This kind of algorithm (Figure 3) can analyze the linear or nonlinear relationship between input variables and output variables from the data itself, which can be continuous, unfixed and non-unique values^[25]. Using preprocessed and labeled data sets, define certain variables as inputs and expected variables as outputs.

At the same time, according to the type of output, it can be divided into classification problem and regression problem. The purpose of the former is to predict the category of the sample through the input, while regression is mainly used to predict the true value of a variable, and its output is not a specific category, but an actual continuous value^{[26⇓~28][29]}. Three ML algorithms, linear regression, polynomial regression, and kernel methods, are commonly employed to construct numerical models based on the data used in the training process^[30]. While the algorithmic architecture used to develop such models varies with the ML method used, any supervised ML algorithm is a process of relating some output ( \overrightarrow{{\mathrm{y}}_{\mathrm{i}}}) to a numerical model ( {\stackrel{➝}{x}}_i) of some input^[31]. ML is the process of fitting a model with small error through the training and learning of the sample data set, and at the same time making the model have the ability to fit the new sample data set, which is called "generalization ability", and is also a very important property in ML.

Unsupervised learning uses unlabeled data sets without providing a specific output value and wants to capture some hidden structure from the input data. The application of this algorithm in LBs is mainly divided into two aspects: clustering and dimensionality reduction. For example, for dataset classification or for data feature reduction to identify the most important variables. Therefore, the essence of unsupervised learning is clustering, which does not know which categories the data will be divided into at the beginning, but only learns according to the characteristics of the data itself, so as to infer the internal structure of the data.Systematically de-disclosing the dataset and evaluating the quality of the trained model is relatively easy for supervised models (the most commonly used models in the battery domain), but it is very challenging for this algorithm^[32][33]. At present, the data generated in the process of battery research is growing exponentially. These data are preprocessed by clustering or dimensionality reduction to identify the most important variables, and then researchers provide chemical/physical meaning for the classification results of the data.

As shown in (Table 1), in addition to supervised and unsupervised learning algorithms, there are a variety of algorithm comparisons, and the choice of algorithm plays a key role in the results. There are four main tasks of ML algorithms: classification, regression, clustering, and probability estimation. Selecting the appropriate algorithm can not only output the reliability of the results, but also shorten the computing time. Several of the most common algorithms are discussed below.

Both experiments and computational simulations, including ML, are modeled cognitions of the laws of nature, and both have their own solutions to problems. Only by understanding the principles and boundary conditions of both sides, can we take advantage of the advantages of both sides to bring help to the field of battery materials research and development. There is no corresponding relationship between ML model and experiment, and they are independent of each other. The selection of ML model and parameters is subjective, which will lead to the final prediction results not meeting expectations. This discrepancy is normal.Because there is no ML model that can perfectly represent complex experimental situations, the variable analysis process and the final prediction effect of ML only give researchers an overall trend, the ultimate goal of which is to explain the experimental phenomena^[41]. Hao et al. Optimized the synthesis parameters of cathode materials through the guidance of ML model. When constructing ML model, although there is an algorithm with poor prediction performance,However, in the construction process, it is found that four characteristic parameters have a greater impact on the prediction results. Combined with the experimental trend, the effectiveness of the ML method is proved, which supplements the time for the traditional Edison method to find important parameters and accelerates the development of materials^[42].

In general, ML results are inconsistent with experiments, which probably means that researchers ignore key variables in the process of experiments/simulations, while ML can quickly find potential laws, replace or cooperate with traditional experiments and computational simulations, and achieve higher accuracy in predicting the phase or microstructure of materials.

Cathode materials are the key to the development of LBs, and finding cathode materials with ideal energy density and low cost is the premise to meet the demand of advanced LBs^[54]. ML is widely used to predict the performance of cathode materials for rechargeable batteries. For active electrode materials, the main characteristics of concern are discharge capacity, capacity retention, volume change, coulombic efficiency, voltage and so on. ML can be used as a tool to construct a model with good correlation between various factors and battery properties by setting appropriate input variables and output variables, and to identify the best characteristic variables to predict its performance under the condition of ensuring the prediction accuracy^[26][11][55].

High voltage cathode materials are the key components of high energy density LBs, which can improve the voltage plateau of LBs, and some of them can work up to 5. 0 V. However, it takes a lot of time and economic cost for researchers to screen reliable cathode materials by experimental verification. In order to solve this problem, Lin et al. Developed a ML tool based on convolutional neural network (CGCNN) to predict the voltage characteristics of batteries, so as to screen out more stable high-voltage cathode materials^[56]. They screened more than 130,000 inorganic Materials from two representative materials databases, the Materials project and AFLOW, used these data to build a new data set, and used the data set to predict about 80 possible high-voltage candidate materials with low toxicity, high density, high capacity and other characteristics. The results show that the hybrid prediction potential combined with Materials project and ML prediction voltage is in good agreement with the experimental measurement results of 10 known battery cathode systems, while about 70 other unreported candidate Materials are predicted at the same time to be verified by future experiments^[56]. Shang et al. Trained organic molecules with known redox potentials by ML, and successfully synthesized cathode materials for lithium-ion batteries by ML. Monomer BC_z-PH with high redox activity was successfully predicted by molecular access system (MACCS) and SMILES. Compared with Li⁺/Li, the discharge voltage was significantly improved to 4.5 and 4.8 V, with good cycling stability, and high cycling charge-discharge specific capacity was achieved^[57].

Furthermore, in the study of lithium-sulfur batteries, the adsorption capacity of cathode host materials for polysulfide (LiPS) is calculated by DFT. There are many possibilities for the adsorption sites, and different adsorption sites will lead to different calculation results. This process is costly and time-consuming^[58]. To this end, Hai et al. Used transfer learning to successfully predict the adsorption of Li₂S₆ at any position on a substrate material on the basis of a small data set (Fig. 5), and the resulting model had a fairly low mean absolute error (below 0.05 eV)^[59]. The proposed data-driven method, with accuracy comparable to DFT calculations and significantly shorter screening time for AB₂-type sulfur host materials, provides a highly accurate and fast solution for other high-throughput calculations and material screening based on adsorption energy prediction.

To sum up, the ML method shows higher accuracy and efficiency than the traditional material property prediction method, and the combination of DFT calculation and ML advanced technology and algorithm is becoming an important means for automatic screening of compounds.Although the artificial choice of ML algorithm and the construction of the sample data set will greatly affect the final performance model, no single ML algorithm can be suitable for all applications^[60][61]. Therefore, the best solution is always a comparison of multiple ML algorithms, which is a common approach for all ML applications.

Solid-state electrolytes (SSEs) have been widely studied because of their excellent safety characteristics, but it is still challenging to develop SSEs with high ionic conductivity, good interfacial properties, excellent mechanical properties, and electrochemical stability^[62][63]. In general, the DFT method is used to explore the ion transport characteristics of electrolytes, but its feasibility and timeliness are poor when applied to large systems^[64]. Jalem et al. First combined the computational data with the ML algorithm and discovered a new olivine oxide solid electrolyte with low ionic conductivity. The ML method began to be well known. ML provides a more convenient and efficient method for the study of the ionic transport characteristics of solid electrolytes, which will accelerate the research and development of advanced solid electrolytes^[65][66].

Ionic conductivity is one of the most important evaluation indexes of SSE, and SSE with superior ionic conductivity and interface stability is an ideal material for stable all-solid-state lithium metal batteries.Predicting the properties of battery materials is also an important function of ML method. Through the selection of descriptors and the screening of algorithms, the conductivity can be well predicted^[67][68]. Sendek et al. Reported a neural network for predicting the fast conduction ability of lithium ion at room temperature, and constructed a convolutional neural network using five descriptors as input data to train the atomic structure^[69]. Eleven crystalline compounds were identified from 317 candidate materials by ML method, and they found that Li₅B₇S₁₃, Li₃InCl₆, and Li₂B₂S₅ have high ionic conductivity, among which the ionic conductivity of Li₅B₇S₁₃ is expected to be 0.074 S/cm, which greatly reduces the complexity of artificial screening. ML-based simulation results show that the logarithmic average of conductivity is improved by a factor of 44 compared to random guessing and manual calculation. However, the input data of this method contains less attribute information, and the number of screened materials is small, so the accuracy of the trained ML model is low. Similarly, Feng et al. Used a neural network potential to simulate materials composed of Li, Zr/Hf, and Cl (Fig. 6), and used a random surface walk method to identify two potential unique layered halide SSEs.A halide solid electrolyte with high lithium ion conductivity and excellent compatibility with lithium metal anode was obtained by ML prediction, and a stable lithium plating/stripping performance record of 4000 H was set^[70].

At present, all studies focus on ion mobility (migration energy or conductivity) as the target attribute, but the mechanical properties of SSE films also have a great impact on the performance of batteries. In the past, the fabrication of SSE films was mainly based on the evaluation of conductivity.Artificial control can not be considered uniformly, and the ML method trained by experimental data is an effective method to predict the ion transmission performance of SSE. A large number of seemingly useless data generated in the process of repeated experiments.In fact, it provides accurate and reliable data for ML. As shown in Figure 7, Chen et al. Used ML to obtain a data set from the data obtained in the experiment, and then used three ML algorithms, namely principal component analysis (PCA), K-means clustering and SVM, to explore the relationship between performance and experimental variables. Finally, a SSE film with a thickness of 40 μm was produced and successfully cycled for 100 times^[71]. At the same time, the mechanical properties of SSE also have an important impact on the performance of the battery. How to predict the best mechanical properties has always been a difficult problem. Jo et al. Developed a ML-based regression model, obtained 12361 Materials from the Materials project as a training set, and represented them by their respective chemical structure descriptors^[72]. The developed ML model exhibits remarkable accuracy, with mean absolute errors of 11.8 and 15.3 GPa for the shear and bulk moduli, respectively. The model was subsequently applied to predict the mechanical properties of 2432 Na-SSEs and validated by first-principles calculations. Finally, an ideal material screening platform is developed by adding a minimized data set for the optimization process. This method fully demonstrates how the ML method guides the experiment to achieve the desired effect, and provides a new research idea for the future study of electrolytes.

However, considering the complexity of chemical systems, it is still difficult to fully calculate the total conductivity with the existing computing power alone, because most conductors are added with polymers, which makes the calculation more complicated^{[8,26,73⇓⇓~76]}. To solve this complex interactive computation, Yang et al. Used a polymer database to train a gradient boosting model, which showed an accuracy of 90% on training data and 81% on test data^[77]. The experimental data are similar to the predicted ionic conductivity, which also provides a guide for ML to explore complex chemical systems.

For liquid electrolytes, the basic principle is quite different from that of solid electrolytes, and the unique highly disordered characteristics make the study of liquid electrolytes less convenient than that of solid electrolytes. Therefore, there are fewer examples of ML applications in the field of liquid electrolytes than in solid electrolytes, but the important role of ML in the study of liquid electrolytes can not be ignored.

The physicochemical properties of electrolytes are largely influenced by the structure of the solvent. From the simplest experimental point of view, researchers have been working on new solvent, salt and additive formulations to construct stable electrode-electrolyte interfaces.The traditional trial-and-error method not only blurs the concept between structure and performance, but also is time-consuming, while ML can guide or help the study of liquid electrolytes and help researchers get more constructive conclusions^[78]. The ML method was used to match the Fourier transform infrared (FTIR) spectral signature of an unknown electrolyte composition with the same signature of a FTIR spectral database with known composition,This new method can quickly, cheaply and accurately determine the concentration of the main components in the electrolyte of lithium-ion batteries, which provides more solutions for the current simulation and experimental research of liquid electrolytes^[79]. On the other hand, the understanding of the solvation structure of electrolytes is generally observed through experimental characterization, which makes it difficult to understand the liquid transport properties at the atomic level. The rise of first-principles and MD can solve the above problems well, but their high computational cost and limited simulation scale limit the development of current research. Scherer et al. Used the ML method to construct the ML force field for liquid electrolyte simulation, and used the deep neural network method to make up for the lack of accuracy of density functional theory and the efficiency of classical force field^[80]. Similarly, Nakhaei-Kohani et al. Developed a structure-property model for the transport properties of ionic liquids to guide the design of ionic liquids.The data set used collects the conductivity of ionic liquids from 2625 laboratories in a wide temperature range (25 ~ 484.1 K), and the conductivity of ionic liquids is predicted by four ML models, which provides a more accurate and reliable way to predict the properties of ionic liquids^[81].

All in all, the application of ML to electrolytes is broad, and from a computational point of view assists people to think about experimental studies, from experimental parameter prediction to electrolyte multi-scale simulation,Such application examples open up a new way for more effective experimental research of electrolytes, and provide new solutions for simulating complex liquid electrolytes and finding the relationship between the structure and performance of electrolytes.

DFT calculation and MD simulation are widely used in battery research. As ML enters the field of computational materials science, there are more and more successful cases of ML method combining DFT calculation and MD simulation. For example, by developing a new ML method to reduce the energy and force information of the system to assist DFT calculation and MD model, in the high-precision calculation of large systems, the simulation of 100 million atoms has been realized through the deep learning algorithm, and the 1 ns simulation of the system requiring 100 million atoms can be completed in one day, which greatly reduces the time cost of calculation. For another example, based on the calculation data of DFT and MD to predict certain characteristic properties of materials, the result data of MD simulation can be used as the training data set of ML to provide accurate MD simulation prediction for the whole sequence space of electrolytes. ML not only has great advantages in analyzing large data sets and establishing effective causal relationships for the rational design of materials or methods, but also has great potential in promoting the development of molecular simulation modeling^[82].

MD simulation can predict the trajectory of selected atoms moving with time in different systems of lithium metal batteries by controlling the physical force field of interatomic interaction. Through MD simulation, the basic formation of initial SEI or artificial SEI can be known, including the migration of components and Li⁺ in it^[83]. At the same time, MD simulations can be used to explain how current density, temperature affect the electrolyte composition and lithium dendrite formation factors in the microworld^[84]. However, for the material simulation modeling of multi-atomic systems, the accuracy of traditional molecular dynamics depends on the selection of empirical force fields (FFs) and the related physical intuition, and the cost of its high accuracy is a small atomic system and a huge time cost. Although MD simulations driven by FFs have achieved long trajectories of millions of atoms in microseconds, the accuracy of the interatomic potential depends on the selection of parameters, which need to be empirically fitted and then applied to the calculation. The low transfer of the interatomic potential hinders the possibility of large-scale calculation^[85]. In this regard, the application of ML strategy to ab initio molecular dynamics simulation can not only improve the simulation speed by several orders of magnitude, but also greatly expand the scale of scalable systems, while the emergence of ML methods is expected to change the development method of force fields^[86]. Chmiela et al. employed ML methods to construct more flexible molecular force fields, which enable MD simulations of flexible molecules with up to a few tens of atoms with ab initio quantum mechanical accuracy^[87][88]. Similarly, Chan et al. Developed an automated framework that allows users to create their own models using ML algorithms in the form of advanced sampling and fitting force fields, improved the treatment of reactive charge transfer, and cross-validation and iteration of ML-based potential energy surface (PES) models^[89][90]. The prediction and simulation of the properties and functions of molecular systems need to accurately describe the global PES, but the PES usually lacks transferability and can only be described under appropriate conditions, so how to predict the accurate PES with low-cost methods is a challenge. No Noé introduced ML methods that can accurately describe the global PES of elemental materials and small molecules, resulting in a unique solution, the flow of which is shown in Figure 8, enabling simple ML models to reproduce accurate PES for small molecules, significantly speeding up the evaluation of PES^[91]. At the same time, ML can effectively analyze the data results of MD simulation, and unsupervised learning can complete preliminary clustering or dimensionality reduction analysis for these unlabeled chaotic data to facilitate researchers' understanding.

The time scales that classical MD and DFT calculations can access are limited, and ML optimization algorithms and optimization strategies have great advantages over the training of traditional MD interatomic potential functions^[92]. With the continuous development of computers and the continuous optimization of ML algorithms, ML methods can effectively cut into all aspects of the computational simulation process of battery materials, which is expected to bring a deeper understanding of the working mechanism in the electrochemical process, accelerate the high-precision simulation of large-scale systems, and thus significantly accelerate the design process of battery materials^[93].

Although big data reduces the difficulty of obtaining data for researchers, in the field of laboratory-scale material design, the acquisition of research data sets becomes extremely difficult due to the high cost of calculation/experiment. On the other hand, the data obtained by computational simulation lacks important characteristic parameters because it can not fully simulate the experimental conditions, so it takes a long time and is inefficient to collect the required data sets, and the final predicted variable correlation results still need to be verified by a large number of experiments. Therefore, how to carry out high-throughput experiments (HTE) based on ML and accelerate material design through experimental design tools and automated experimental platforms has become a new direction in the field of materials science in the future.

Traditional experimental methods are still largely based on human knowledge and intuition and low-throughput experiments. Effective exploration of unexplored chemical space requires intelligent automation technology, and HTE can carry out a large number of parallel experiments through advanced automation platforms to complete the designated experimental work according to the set rules.Therefore, chemicals and resources can be effectively utilized, and data results can be quickly obtained for chemical reaction characteristics of interest. The advantages of standardization of technical parameters and large volume also generate a large number of accurate and high-quality data for the ML model^[94].

ML is highly similar to HTE in that both extract valuable information through extensive data analysis. HTE includes many aspects, such as representative intelligent design and experiment selection, searching and optimizing in a large number of samples with different parameters and conditions to accelerate the whole search process^[95]. The effective combination of machine learning and high-throughput computing in materials science, including the processing of high-volume information and the automation of experiments in a feasible large-scale repetitive manner, allows experiments to be conducted faster without sacrificing the quality of results, with the help of ML for data collection and processing. For example, for liquid electrolytes, the highly disordered nature of their formulations makes the ML method difficult to apply due to their numerous combinations of solvents and additives. Kafle et al. Rationally designed the most commonly used solvent components in lithium-ion batteries through a high-throughput screening method, and proved that the low-temperature power capability of lithium-ion batteries can be significantly improved^[96]. Whitacre et al. Constructed a system named Otto, with the process shown in Fig. 9, to achieve HT automated formulation and on-line characterization of liquid water battery electrolytes^[97]. Compared with traditional low-throughput experiments, the platform can prepare 140 electrolytes in 40 H. After that, the ML method automatically evaluates the collected data set to realize the inverse material design. The optimal electrolyte was found to be a novel dianionic sodium electrolyte with a wider electrochemical stability window than the baseline sodium electrolyte.

The combination of HTE and ML effectively enables automated experimental design, online characterization, and rapid parallel experimental data analysis, which accelerates the discovery and development of battery materials with high repeatability and efficiency. However, the construction of advanced automation platform requires powerful hardware and software conditions, and these difficulties still need continuous exploration and efforts. Large-scale practical applications in the future will undoubtedly change the mode of chemical research and provide a huge impetus for the development of LBs.