Numerous previous studies have investigated how different synthesis parameters affect the chemical properties of catalysts and their performances. However, traditional trial and error optimization in comprehensive multiparameter spaces that is driven by chemical intuition may cause influencing factors to be artificially ignored. Hence, we introduce machine learning to provide insights by feature ranking based on data sets. Taking zeolite imidazole framework-derived oxygen reduction catalysts as an example, computing results reveal that pyridinic nitrogen species are strongly related to catalytic performance. Besides pyrolysis temperature, pyrolysis time, which has not been set as variable by the vast majority of studies, is discovered to be decisive at the synthesis level. Guided by these predictions, the insights of the algorithm are verified by control experiments. The characterization results and interpretable model reveal an ignored mechanism. Continuous processes that successively affect pyridinic species, including the loss of Zn–N species, formation of Fe–N species, and conversion into graphitic N species, resulted in a volcano-like relationship between the half-wave potential and the pyrolysis time. This work not only provides insights into catalyst design but also proves that machine learning has the ability to mine key factors and mechanisms concealed in complex experimental data to boost the optimization of energy materials.

Recent researches have proven that incorporation of machine learning could significantly shorten development cycle of energy materials. However, this rising multidisciplinary field still needs a standard research paradigm instructing how...

Traditionally,alarger number of experiments are needed to optimize the performance of the membrane electrode assembly (MEA) in proton-exchange membrane fuel cells (PEMFCs) since it involves complex electrochemical, thermodynamic,a nd hydrodynamic processes.H erein, we introduce artificial intelligence (AI)-aided models for the first time to determine key parameters for nonprecious metal electrocatalyst-based PEMFCs,t hus avoiding unnecessary experiments during MEA development. Among 16 competing algorithms widely applied in the AI field, decision tree and XGBoost showed good accuracy (86.7 %and 91.4 %) in determining key factors for high-performance MEA. Artificial neural network (ANN) shows the best accuracy (R2 = 0.9621) in terms of predictions of the maximum power density and ad ecent reproducibility (R2 > 0.99) on uncharted I-V polarization curves with 26 input features.H ence,m achine learning is shown to be an excellent method for improving the efficiency of MEA design and experiments.