Machine learning-based polarization-insensitive biosensor for plant growth regulator detection

Plant growth regulators (PGRs), an important class of phytohormones, have attracted increasing demand for accurate detection, making terahertz (THz) metasurface biosensors a promising sensing platform due to their high sensitivity and label-free capability. To further improve detection accuracy, resonance-absorption matching is commonly employed to enhance near-field effects. However, achieving simultaneous matching between multiple resonances and the absorption peaks of multiple analytes remains challenging, as optimizing one resonance may affect the others, thereby reducing spectral distinguishability and limiting subsequent qualitative and quantitative analysis. In addition, for detection in complex environments, it is necessary to employ polarization-insensitive structures. Therefore, this work proposes a machine-learning-assisted polarization-insensitive metamaterial (PIM) biosensor for the qualitative and quantitative detection of three PGRs: thidiazuron (TDZ), forchlorfenuron (CPPU), and 6-benzylaminopurine (6-BA). First, a multi-objective optimization strategy based on the constrained two-archive evolutionary algorithm was employed to optimize a predefined PIM, and the resonance valleys of the optimized structure were aligned with the absorption peaks of the target analytes at 1.31, 1.60, and 1.80 THz. Subsequently, a two-stage machine learning framework was developed based on the experimental data for analyte identification and quantitative analysis. The proposed framework achieved 100% classification accuracy for the three PGRs on the test set, with an average root mean square error (RMSE) of 3.61 μg and an average coefficient of determination (R 2 ) of 0.9917 for quantitative detection, demonstrating excellent sensing performance. This study provides a promising strategy for the simultaneous detection of multiple target analytes in complex environments.