A generalized machine learning framework to predict the space-time yield of methanol from thermocatalytic CO2 hydrogenation

Identifying Descriptors for Promoted Rhodium-Based Catalysts for Higher Alcohol Synthesis via Machine Learning

“…Such endeavors, which ensure that the predictions are explainable and in alignment with the existing domain knowledge, will instill a greater sense of trust and acceptability within the community, specifically from an experimentalist point of view. For this purpose, the agnostic ML explainer, SHAP, was coupled with the random forest algorithm to determine the impact of the input features on the prediction of STY HA . , The overall influence of each feature was calculated by normalized SHAP values (Figure ).…”

“…Furthermore, all of the input features used in the study were checked for correlation among themselves. Correlated variables tend to increase the model complexity without adding significance to its prediction capabilities . The correlation among the input features was calculated by Pearson’s correlation coefficient (PCC) given in eq . PCC = ∑ ( x i − x mean ) ( y i − y mean ) ∑ false( x i − x mean false) 2 ( y i − y mean ) 2 where PCC is the Pearson’s correlation coefficient; x i and x mean are the sample and mean values of the x variable, respectively; and y i and y mean are the sample and mean values of the y variable, respectively.…”

“…Causal analytics, commonly referred to as feature importance, is the process of deducing the most important input features that have contributed to the model training and its final prediction . We clarify that the term causal analytics does not establish a mechanistic relationship between the input features and the target variable but rather helps to determine correlations between them . Herein, causal analytics was performed by employing the SHapley Additive exPlanations (SHAP) methodology on the best-performing algorithm, a posteriori feature-importance approach to make ML models interpretable .…”

“…Two ensemble tree-based ML algorithms, namely, random forest and gradient boosting trees, were developed on the training data and comparatively evaluated for their performance to predict the STY HA on the test data. Ensemble tree-based algorithms have recently gained popularity for their applications in heterogeneous catalysis. ,, In contrast to the popular ML algorithms like neural networks and support vector machines, or the inherently explainable algorithms like linear or symbolic regression, they offer greater accuracy when dealing with small datasets (ca. 200–500 datapoints), along with ease of model tuning and faster convergence. , These attributes of the ensemble tree-based algorithms motivated their selection in this study.…”

Data-driven design of electrocatalysts: principle, progress, and perspective

“…179 On the other hand, data mining algorithms can also be used to extract data from existing published papers, and these data come from real materials and real test conditions. 180 Of course, we still need to explore how to integrate "real data" from different data sources, which places high demands on data preprocessing and standardization procedures.…”