Machine Learning-Accelerated Discovery of Oxygen-Bridged Layer-Dependent Double-Atom Catalysts on G-C16N5 for Electrochemical Nitrogen Reduction

This is the authors' abstract. We don't add key points for this paper.

Nitrogen reduction reaction (NRR) electrocatalysis suffers from limited active-site tunability and poor activity-selectivity control. By combining density functional theory (DFT) with machine learning (ML), herein, we propose an oxygen-bridged, layer-dependent double-atom catalyst design on g-C 16 N 5 . A bilayer TM 1 -O-TM 2 normal/inverse sandwich motif was constructed using 18 transition metals, generating 324 candidate structures to accelerate catalyst discovery. Among the tested algorithms, Gradient boosting regression delivered the best predictive accuracy and enabled efficient identification of highly active candidates. DFT validation identified Mo–O–W as the optimal bilayer catalyst, achieving a limiting potential of −0.29 V via the distal pathway. Constant-potential calculations reveal pH-/potential-dependent NRR energetics on Mo–O–W, enhancing activity under acidic conditions and demonstrating that electrochemical potential modulates adsorption via frontier-orbital reorganization and changes in d-state occupation. Extending the design to pseudo-bilayer, bilayer, trilayer, and quadlayer architectures yielded a library of 1296 layer-dependent structures. Promising candidates include pseudo-bilayer Re–W (−0.40 V), trilayer Mo–O–W–O–Mo (−0.17 V), and quadlayer Mo–O–Nb–O–Nb–O–Mo (−0.53 V). SHapley Additive exPlanations analysis identifies θ d1 , layer number, M 1 , E i1 , and E A2 as the most influential descriptors governing NRR activity. This work establishes oxygen-bridged layer engineering as an effective route for tuning double-atom catalysts and provides a practical DFT-ML framework for accelerated discovery of high-performance NRR electrocatalysts.