Explainable prediction of N-V-related defects in nanodiamond using neural networks and Shapley values

Abstract

Although the negatively charged nitrogen-vacancy (N-V − ) defect in nanodiamonds is desirable for a variety of biomedical applications, a range of other defect complexes involving nitrogen and/or vacancies can also exist, depending on their relative stability. Using machine learning, a re-usable model is developed to predict the likelihood of a particular defect complex being stable at a given depth below reconstructed or hydrogen-passivated surfaces. A neural network is used to generate a system of equations that can be easily implemented in any workflow, and explainable artificial intelligence (XAI) methods are used to provide insights into which structural features and defect configurations are most responsible for the model prediction. It is found that, although the number of nitrogen atoms present in the defect is the most important feature determining the defect likelihood, the most influential data instances are the unlikely defects, providing a type of baseline for comparison. • An accurate neural network predicts the likelihood of N-V defects in nanodiamond • Explainable AI identifies important features and influential defect configurations Although the negatively charged nitrogen-vacancy defect in nanodiamonds is useful, a range of other defect complexes involving nitrogen and/or vacancies can also exist. Here, Barnard uses machine learning to develop a re-usable model to predict the likelihood of a particular defect complex being stable at a given depth below a given surface.