Multi-target neural network predictions of MXenes as high-capacity energy storage materials in a Rashomon set

Machine learning models are sensitive to the hyperparameters used by the algorithms and, depending on the method and settings, will place different importance on the (structural) features when predicting (property) labels. This is crucial for machine learning in high-performance materials such as MXenes, known for their design flexibility and increasing complexity. To provide unequivocal structure/property relationships for materials science applications, a universal feature importance profile is needed, which can be obtained by using a Rashomon set of high-performing models, even for so-called “black-box” models. Presented here is a method for developing universal feature importance profiles that can be combined with any neural network to provide an interpretation that is insensitive to factors such as the number of layers or neurons. This approach gives more comprehensive insights into the importance of features and allows researchers to control specific property ranges to accommodate their research interests or needs without compromising performance.