How false data affects machine learning models in electrochemistry?

False data is detrimental to the prediction of machine learning in chemistry. But some models are more tolerant to noise than others. The selection of machine learning models in electrochemistry is based on only the data distribution without concerning the quality of the data. This study aims to provide a discussion of the failure input data in electrochemistry, which demonstrated using heteroatom doped graphene supercapacitor data. The electrochemical data were tested with 12 standalone models including XGB, LGBM, RF, GB, ADA, NN, ELAS, LASS, RIDGE, SVM, KNN, DT, and our "stacking" model. By gradually adding the false data into the pool, the models were then trained on both noisy and ground truth data to obtain various error metrics (MAE, MSE, RSME, MAPE, and R2). The linear regression was then fitted on the errors to obtain the slope and intercept, which refer to noise sensitivity and base accuracy, respectively. Hence, this study utilized contour plots, SHAP, and PDP to explain how the error affects the electrochemical feature including prediction and analysis. It is found that linear models handle the false data well with an average MAE slope of 1.513 F g−1, but it suffers from prediction accuracy (MAE intercept of 60.20 F g−1). This is due to improper model selection for this type of data (average R2 intercept of 0.25). The "Tree-based" models fail in terms of noise handling (average MAE slope is 58.335 F g−1), but it can provide higher prediction accuracy (average MAE intercept of 30.03 F g−1) than that of linear models. Tree-based models also fit well to the data (average R2 intercept of 0.9516). This suggests that the linear based model can be well described the relationship between capacitance and surface area. While the "Tree based" model can be used for handling the other electrochemical features e.g. amount of heteroatom doped, current density, and so on. Miscellaneous models such as SVM, KNN, and NN, are moderately robust to noise (average MAE slope of 25.956 F g−1) and provide moderate accuracy (average MAE intercept of 41.306 F g−1). The models also fit moderately well to the data (average R2 intercept of 0.546). To address the controversy between prediction accuracy and error handling, the "stacking model" was constructed, which not only shows high accuracy (MAE intercept of 24.29 F g−1), but it also exhibits good noise handling (MAE slope of 41.38 F g−1and R2 intercept of 0.86), making stacking models a relatively low risk and viable choice for electrochemist. This study presents that untuned NN is not suitable for electrochemical data, and improper tuning results in a model that is susceptible to noise, which directly affects the misleading in the electrochemical discussion. Thus, "STACK" models should provide better benefits in that even with untuned base models, it can achieve an accurate and noise tolerance. Overall, this work provides insight into machine learning model selection for electrochemical data, which should aid the understanding of data science in chemistry and energy storage context.