Importance of Structural Features and the Influence of Individual Structures of Graphene Oxide Using Shapley Value Analysis

Abstract

The application of machine learning (ML) to materials chemistry can accelerate the design process, and when coupled with a detailed explanation, can guide future research. Shapley value analysis is a complementary approach capable of providing a comprehensive analysis of the underlying reasons behind a structure/property relationship. In this study, we have used data sets of graphene oxide nanomaterials generated using electronic structure simulations to train ML models with outstanding accuracy, generalizability, and stability to predict the formation energy and the Fermi energy and applied Shapley value analysis to understand the results. Feature important profiles that rank the value of structural characteristics to each property confirmed that the underlying structure/property relationships are relatively simple and scientifically intuitive, even though the ML models need complex information to achieve high performance. We have also reported instance influence profiles that rank the value of each individual graphene oxide structure to the training process. Feature/instance interactions are also investigated to explain which structural characteristics make particular structures influential, revealing that the most influential structures typically have very high or very low concentrations of H or O. Since the range of concentrations is typically chosen by researchers based on domain knowledge at the outset, this highlights that extreme care should be taken when gathering training data as these decisions will have a very big impact on the final model once trained. In general, the reproducible workflow demonstrated here can be applied to any similar materials data set to make reliable model-agnostic predictions of how the structural characteristics and individual structures contribute to the prediction of functional properties.

Interpretable Machine Learning Model on Thermal Conductivity Using Publicly Available Datasets and Our Internal Lab Dataset

“…For instance, the spread of the Se data points or variance distribution of its SHAP values is larger for the UWAT_TE data compared to the temperature feature. For the other datasets, the temperature feature is considered the most important one by SHAP due to the wide distribution of its individual data points with SHAP values from the center SHAP value 0.0, as SHAP orders the feature importance based on the higher variance distribution of data points from the center of the SHAP value …”

Machine learning electron-phonon interactions in 2D materials

“…Since ensemble tree models are not intrinsically interpretable, 59 we choose to use SHAP (SHapley Additive exPlanations) values as an extrinsic tool for statistical interpretability 43 . SHAP has been recently utilized to understand structural features that lead to electrocatalysis in N-doped carbon materials 60 and electronic structure of graphene oxide 61 . Briefly, SHAP trains a given model with every permutation of input features and generates predictions on the data.…”

Self-assembly prediction of architecture-controlled bottlebrush copolymers in solution using graph convolutional networks

“…Single chain properties are screened by Shapley additive explanations (SHAP) values to enhance the model performance through optimization of input features. 40,41 The correlation between the single chain properties and self-assembly morphologies is also analyzed using SHAP values. Phase diagrams of bottlebrush copolymers with given side chain sequences are predicted by single chain properties obtained from GCNs.…”