First principles and neural network-driven biochar spectral database: Raman, XPS, IR, and NMR

Abstract

This study integrates density functional theory and neural network algorithms to predict the spectra of 32 representative molecular structures identified in biochar. This approach addresses computational challenges in biochar molecular modeling and results in the generation of a spectral database covering X-ray photoelectron spectroscopy (XPS) (C 1 s, O 1 s, and N 1 s), Raman and infrared (IR) spectroscopy, ¹H and ¹³C nuclear magnetic resonance (NMR), and 2D NMR. The spectra of independent molecules were then aggregated to describe the spectrum of cellulose char produced at 500 °C. Integrating first-principles predictions with machine-learning techniques establishes a connection between the atomic structures of biochar and their corresponding spectroscopic signatures. This work also expands the reliability of experimental data interpretation, providing a robust framework for atomic-level modeling and characterization. The theoretical spectra strongly aligned with experimental data, achieving >90 % agreement for ¹³C NMR and XPS and > 77 % correspondence for IR and Raman spectroscopy. These results demonstrate the enhanced predictive power of theoretical spectra derived from accurate molecular structures. The spectral database and atomistic structures lay the foundation for future research, providing opportunities to develop machine-learning algorithms that can effectively predict theoretical spectra. Furthermore, this approach facilitates the generation of spectra that would otherwise be costly or difficult to obtain experimental.