Calibration-free and drift-robust AlGaN/GaN HEMT sensor arrays for intelligent pH detection

AlGaN/GaN high-electron-mobility transistors (HEMTs) combine surface charge coupling with excellent chemical stability, making them highly attractive for liquid-phase biosensing. However, when extended to array-level implementations, sensor-to-sensor variation, chip-to-chip inconsistency, and response variation across repeated measurements can collectively undermine reproducibility. In this work, we report a hardware-algorithm co-optimized sensing platform that integrates a monolithically embedded on-chip reference electrode and multilayer metallization to ensure electrical isolation and matrix-addressed signal acquisition. To address array-level variability, a data-driven artificial neural network (ANN) was employed, improving robustness across sensors, chips, and repeated measurements without explicit recalibration. The model showed robust recognition performance across different chips and measurement conditions, with stable within-chip recognition and effective cross-chip transfer, achieving a classification accuracy of approximately 95% when trained on the combined datasets. X-ray photoelectron spectroscopy and Kelvin probe measurements reveal gradual surface-state evolution at the GaN cap layer during the 30-day storage period. By incorporating an attention-enhanced ANN, the degradation in classification accuracy observed on 30-day datasets is alleviated, yielding improved prediction accuracy without recalibration.