DeepMIR: A Hybrid Convolutional Neural Network-Transformer Framework for Accurate Identification of Target Components from Mid-Infrared Spectra of Mixtures

Accurate identification of components from mid-infrared (MIR) spectra of mixtures remains a significant challenge in analytical chemistry due to severe spectral overlap and instrumental variability. To address this, we present DeepMIR, a deep learning framework designed for targeted component identification in complex mixtures using user-defined spectral libraries. DeepMIR uniquely integrates a convolutional neural network for extracting high-fidelity local spectral features and a transformer encoder to capture global, long-range dependencies across the spectrum. This hybrid architecture enables robust identification of target components even when reference and mixture spectra are acquired using different techniques, such as transmission and attenuated total reflectance. The model is trained and validated on a data set of over 67,000 synthetically augmented spectral pairs, achieving an accuracy of 99.5% on the test set, with statistical significance confirmed by t test analysis (p < 0.05). Extensive real-world validation demonstrates high accuracy across diverse scenarios, 94.8% for complex liquid solvents, 93.9% for solid pigment mixtures relevant to cultural heritage, and 99.1% for commercial blended textiles. Dedicated limit-of-detection studies establish reliable detection thresholds of 20% v/v in liquids and 10% w/w in solids. To ensure accessibility, DeepMIR has been deployed as an open-access web server, providing a powerful and practical tool for the scientific community that significantly outperforms traditional library search methods in both accuracy and reliability.