Multi-task learning for terahertz metamaterial molecular fingerprint sensing of α-amino acids

• The first multi-task learning (MTL) framework (MetaFP) is proposed for terahertz (THz) metamaterial sensing. • Simultaneous high-accuracy classification and precise quantification of structurally similar α-amino acids. • Employs supervised contrastive learning for cross-angle fingerprint consistency. • Achieves a 42.7% MAE reduction and and 93% classification error reduction over baselines. The terahertz (THz) molecular fingerprints of structurally similar α-amino acids exhibit substantial spectral overlap, which makes them difficult to distinguish using conventional metamaterial sensing. Existing deep learning models further compound this problem by remaining restricted to single-task learning and thus cannot jointly perform accurate identification and quantification. To address these limitations, this work proposes an integrated multi-task learning framework specifically tailored for THz metamaterial sensing. An all-dielectric grating-waveguide sensor based on cyclic olefin polymer (COP) is designed, leveraging the coupling between guided-mode resonance (GMR) and quasi-bound states in the continuum (quasi-BIC) to form an angle-frequency resonance envelope. Built upon this platform, the proposed framework constructs a multi-modal input by fusing physical priors including incident angle, resonance peak information. Multi-angle spectra are treated as complementary views of the same molecular fingerprint, while supervised contrastive learning is employed to capture cross-sample global correlations and enhance fingerprint reconstruction fidelity. A learnable adaptive multi-task loss is introduced to dynamically balance classification and regression during training. Numerical simulation results show a 42.7% decrease in concentration prediction error (MAE) and a 93% reduction in discrimination error versus state-of-the-art baselines. High-accuracy sensing via a single-angle spectrum reduces physical scanning time by > 10×.