Sequential construction of stable nitrogen-oxygen compounds using high-throughput quantum mechanical calculations and customized machine learning model

Nitrogen and oxygen are the two most abundant elements in the atmosphere, yet stable compounds composed solely of these elements are relatively scarce. Conceiving novel stable nitrogen-oxygen compounds remains a formidable challenge for current experimental and theoretical research. In this study, we developed a sequential construction strategy to design 168 nitrogen-oxygen compounds with distinct structural innovation, followed by high-throughput quantum mechanical calculations with the highest possible accuracy. From the resulting 7820 structural and property parameters, we created a customized machine learning model that outperforms universal models in accuracy with 13.8% greater robustness across various data splits, achieving stable and high performance on small datasets. Data-driven analysis revealed the energy and electron-related characteristics as key factors in regulating thermodynamic stability, while physics-driven insights uncovered that electron delocalization and hyperstatic constraints fine-tune mechanical firmness. Among the designed nitrogen-oxygen compounds, 106 are expected to be more stable than the known compound N₂O₄, out of which 61 are expected to be even more stable than N₂O₅. Furthermore, their energy densities surpass those of all currently used nitrogen-oxygen oxidizers by 8.3-16.8%, highlighting our newly proposed compounds potential for use in rocket bipropellant systems. Our developed machine learning platform features a user-friendly graphical interface for easy assessment and may be of interest to researchers in other fields, including chemical industry and energy sectors.