Electrode Performance Prediction Of Electrochemical Energy Storage Devices Based On Machine Learning

In recent years,electrochemical energy storage devices have been widely used in many fields such as electronic equipment,new energy vehicles and power systems.It has been a hot research topic in the scientific and industrial communities to continuously improve the performance of electrochemical energy storage devices,such as energy density,power density,safety,rate characteristics and service life,so as to broaden the scope of their applications.As well-know,many macroscopic and microscopic influencing factors affect the performance of electrochemical energy storage devices and they are coupled with each other,which brings difficulties and challenges to the optimal design of electrochemical energy storage devices.The relationship between the structure and performance of electrochemical energy storage devices has been analyzied by using computational simulation methods such as molecular dynamics or density functional theory,which can avoid the mutual interference of various factors in the experiments and achieve performance optimization.However,the current simulation methods mainly focus on the atomic and microscopic scales,and it is difficult to simulate and analyze the macroscopic morphology and performance of electrochemical energy storage devices or materials.So,it is still high interest to develop a new multi-scale（including atom,nano-size,micro-size and macroscopic morphology）computational simulation method to construct the quantitative relationship between structure and performance of electrochemical energy storage devices to provide theoretical guidance and basis for their structure design and performance optimization.Recently,machine learning methods have been used to optimize and design of electrochemical energy storage devices and their electrode materials.However,there are still some problems and challenges as follows.Firstly,the prediction accuracy and efficiency of present machine learning computing models are still poor;Secondly,the transparency and interpretability of existing models are insufficient,and there is a lack of integration with physical and chemical mechanisms,which limits the further acquisition of in-depth insights from predictive models that meet theoretical expectations and practical applications;Thirdly,present machine learning computing models are still unable to analyze non quantitative factors and parameters,such as the morphology and crystal structure of electrode materials,which are important influencing factors on performance of electrochemical energy storage devices.Based on above problems,new machine learning models are constructed.The effect of electrode material morphology,manufacturing parameters,electrochemical cycles,charge/discharging current density on performance（eg.capacity,cycle stability,etc.）of lithium-ion batteries（LIBs）and supercapacitors（SCs）are investigated in detail.The specific research content and innovation are shown in following.（1）Generally,machine learning is a black box model with poor interpretability.Although the existing models can initially establish the quantitative correlation between the preparation process and performance of LIBs electrodes,yet,they are still unable to systematicall analyze the potential correlation and interdependence between them.Here,a new machine learning model is established,which is based on Gaussian progress regression（GPR）algorithm with RQ+SE combination kernel function.It can explore the relationship between active material mass content,solid-to-liquid ratio,viscosity,comma gap and mass load of active materials in electrode.The results show that this model not only has better prediction accuracy than previous models（R²,RMSE,and MAE are 1,0.84322,and 0.6833,respectively）,but also can obtain the optimal slurry formula to optimize manufacturing parameters under different comma gap by combining feature importance,partial dependence plot,shapley values and sensitivity analysis.These provide a theoretical basis and guidance for electrode production and reducing experimental costs.（2）Traditional machine learning methods still have low accuracy in predicting the health of LIBs due to the phenomenon of local capacity regeneration.In addition,it also is difficult to predict the SOH and RUL of LIBs simultaneously by one machine learning model.To overcome above problem,an EMD-PF-GPR fusion model is constructed by combining with data-driven and mechanistic modeling.The results show that the fusion model has strong robustness and generalization ability,and can identify the common capacity regeneration phenomena in actual use process.At a train to test ratio of 50%to50%,the model can synchronously predict the SOH and RUL values of NMC electrodes,with maximum errors not exceeding 1.5%and 1 cycle,respectively.（3）Generally,machine learning method is difficult to analyze the effect of micro-structure,surface functional groups,and testing of carbon materials on performance of carbon based double-layer supercapacitors.To overcome this problem,here,a carbon based double-layer supercapacitor dataset with about 1360 sets of data is firstly established through literature research.Furthermore,a GJO-GPR prediction model is developed for the quantitative relationship between carbon-based electrode structure characteristics and specific capacitance.The results indicate that the model not only has higher prediction accuracy than other models reported in previous literature,it also can obtain some influencing factors and mechanisms that cannot be obtained in experiments.The GJO-GPR model reveal that the specific capacitance value of carbon based electrodes does not always increase with the increase of specific surface area.The electrode rate performance under high current density is related to the relationship between micropores and mesopores.It should be pointed out that electrode performance can be optimized by balancing the surface area of micropores and mesopores of electrode.（4）Previous experimental research has indicated that morphology of active materials is the key role for performance of supercapacitors.However,up to now,machine learning computing models cannot simulate the effect of morphology of active materials on performance of supercapacitors due to its non-numerical features.Here,the Ni Co₂O₄ array with controllable morphology directly grown on the surface of foam nickel is synthesized and lots of data about Ni Co₂O₄ array with controllable morphology are collected.A new calculation model based on GOA-GPR algorithm is developed.The effect of various roles（including morphology）on specific capacitance of supercapacitors are investigated.Furthermore,combining with feature coding technology and SVR algorithm,a quantitative functional relationship between electrode morphology characteristics and capacity retention rate is also investigated under different charging and discharging cycles.The results show that the constructed model can accurately predict the cyclic stability of electrodes with different morphologies reported in present and previous works.