Integration of a self-attentive neural network and density functional theory for accelerated screening of graphene-based stabilized binary adsorption systems

Abstract

It is crucial to study adsorbed atomic systems in graphene as exogenous atoms or molecules can precisely modulate the properties of graphene and significantly extend its wide-range applications. However, graphene's exceptional sensitivity to environmental conditions complicates the precise control of the spatial distributions and binding modes of adsorbates at the atomic scale. Moreover, the stability and reversibility of these systems often rely on rigorous preparation processes and characterization methods. Currently, no effective method exists for systematically identifying and screening graphene adsorption systems with optimal performance. To tackle these challenges, herein, we integrate first principle calculations with machine learning not only to accelerate material screening and structural design but also to elucidate the key influencing factors of graphene adsorption systems. Our studies reveal that combining first principles with a self-attentive neural network can successfully predict the stability of adsorption systems of the previously unseen graphene samples. Furthermore, the trained self-attentive neural network model predicts the adsorption energy trends of diatom systems involving transition metals and rare earth metals, while clarifying the effects of electronic structures, d orbital occupancy, and ionic radius of the elements on their adsorption behavior. This research is expected to advance the design and optimization of new generation graphene-based electronic devices, sensors, and catalytic systems, providing more efficient and precise pathways for future material designs.