A3M-CIR: An Attention-Aware Adversarial Masking Based Self-Supervised Chemometric Framework for Robust Infrared Spectral Analysis

Abstract Infrared (IR) spectroscopy provides rich molecular fingerprint information and is widely used for qualitative and quantitative chemical analysis across biomedical, environmental, and materials science applications. However, its practical utilization is often constrained by instrument-dependent variability, baseline drift, acquisition noise, and limited labeled data. While deep-learning-based chemometric methods have improved performance over traditional preprocessing-driven pipelines, their reliance on supervised learning limits robustness and transferability across heterogeneous measurement conditions. To address these challenges, we introduce A3M-CIR, an attention-aware adversarial masking-based self-supervised chemometric framework for robust infrared spectral analysis. The framework performs large-scale self-supervised pretraining on diverse unlabeled IR spectra to learn transferable representations without explicit calibration or manual annotation. Its core component, the Attention-Aware Adversarial Feature Masking Block (A3FMB), suppresses diagnostically salient absorption regions to generate informative contrastive views that emulate realistic perturbations, while a lightweight Dual-Mixing Attention (DMA) encoder captures both global spectral context and local continuity of one-dimensional vibrational signals. Extensive evaluations under linear-probing and fine-tuning protocols demonstrate consistent improvements over representative self-supervised baselines, and systematic signal-to-noise ratio analyses further validate robustness and transferability under varying noise conditions. By explicitly addressing spectral variability, measurement noise, and limited supervision as intrinsic challenges of infrared spectroscopy, this work establishes A3M-CIR as a practical and extensible foundation for data-driven infrared spectral analysis.