CIRSNet: A Lightweight Chemometric Infrared Spectral Network for Spectral Analysis

Fourier Transform Infrared (FTIR) spectroscopy is a powerful, non-destructive analytical technique widely used in pharmaceuticals, chemicals, food safety, environmental monitoring, forensics, materials science, and biotechnology. However, data modeling remains a challenge beyond spectral acquisition, especially for real-time, edge-based applications. Traditional chemometric methods require extensive preprocessing and computationally intensive offline analysis, making real-time spectral interpretation difficult. We introduce Chemometric Infrared Spectra Network (CIRSNet), a lightweight deep learning model optimized for molecular characterization, enabling onboard processing directly on handheld spectrometers. Unlike conventional multivariate statistical methods and ML-based approaches, CIRSNet eliminates the need for external computation while maintaining high predictive accuracy. A major challenge in spectral analysis is its multilabel nature, where one spectrum may correspond to multiple properties, compounded by the highly imbalanced distribution of spectral data. To address this, CIRSNet is designed for imbalanced multilabel classification, ensuring robust performance even for underrepresented classes. Additionally, we integrate a Spectral-Spatial Attention (SSA) module, which enhances spectral representation by capturing both local and global patterns while preserving dominant peak positions and intensities. We evaluate CIRSNet on a publicly available, highly imbalanced spectroscopy dataset, achieving state-of-the-art performance in functional group detection, and validate its generalizability on an odorant property prediction dataset, where it outperforms existing methods. By enabling real-time, edge-compatible FTIR spectroscopy, CIRSNet advances molecular diagnostics, environmental monitoring, and pharmaceutical quality control, making automated spectral analysis more accessible and efficient. The source codes and trained models are available at: https://github.com/durgesgameta/CIRSNet/.