Non-destructive combustion analysis of biofuels through FTIR and deep learning for elemental composition and HHV determination

• Novel integration of FTIR spectroscopy with deep learning for non-destructive elemental composition analysis HHV determination • Hybrid three-stage framework combining regression, classification filtering, and weight normalization achieving high accuracy (R² > 0.89 for all elements, and 0.93 for HHV) • Successful validation on real-world liquid fuels with HHV determination within ±8% error • Rapid analysis pipeline completing full compositional and HHV determination in under 3 minutes versus 40 minutes for conventional methods Accurate prediction of elemental composition is essential for assessing the combustion quality and environmental impact of biofuels. Traditional CHON (carbon, hydrogen, oxygen, nitrogen) analysis methods, such as elemental analyzers and mass spectrometry, are accurate but destructive, time-consuming, and unsuitable for high-throughput screening. This study aims to address these limitations by developing a deep learning framework for rapid, non-destructive %CHON prediction and higher heating value (HHV) estimation directly from Fourier-transform infrared (FTIR) spectral data. We present a novel pipeline combining AggMap-based two-dimensional (2D) spectral transformation with a hybrid classification-regression CNN architecture. In contrast to conventional approaches that process raw 1D spectra directly, the AggMap algorithm restructures spectral data into 2D feature maps that spatially cluster chemically correlated wavenumber regions, enabling the CNN to more effectively capture relationships between functional group signatures and exploit localized spectral patterns. The hybrid architecture couples element presence detection with percentage quantification to reduce false-positive predictions, followed by weight normalization to ensure physically plausible mass balance. The final model achieved mean absolute errors (MAEs) of 1.84, 0.78, 1.42, 0.59%, and 1423.75 kJ kg -1 for %C, %H, %O, %N, and HHV, respectively, on the gas-phase fuel dataset. When evaluated on real-world liquid fuel samples, the model obtained MAEs of 5.43, 1.98, 3.51, 0.03%, and 1668.83 kJ kg -1 for %C, %H, %O, %N, and HHV, respectively. This cross-phase validation is exploratory in nature and should not be interpreted as definitive evidence of liquid-phase applicability, given the fundamental spectral differences between gas-phase transmission and liquid-phase ATR-FTIR measurements. This approach offers a rapid, interpretable, and scalable pipeline for gas-phase spectroscopic CHON quantification. While preliminary cross-phase validation on liquid fuel samples demonstrates initial transferability, robust liquid-phase deployment will require domain adaptation on large representative ATR-FTIR datasets.