A risk-constrained machine learning method for reproducible and sensitive gas-sensing material screening.

Gas sensors are increasingly needed for rapid detection of diverse air pollutants, yet sensing material development remains labor-intensive due to the wide variety of target gases. Data-driven approaches based on machine learning (ML) offer a promising route to accelerate this process. However, materials with high predicted sensitivity do not necessarily exhibit sufficient device-to-device reproducibility, and current sensitivity-only ML methods may therefore yield candidates with limited practical applicability. Herein, we develop a risk-constrained ML method for reproducibility-aware screening of sensing materials for MEMS gas sensors. Using metal-loaded SnO2 as a prototype, different metal species and loading amounts are integrated into MEMS sensors via a unified fabrication process and tested toward multiple gaseous pollutants, yielding a process-controlled dataset of about 9000 measurements. To account for reproducibility, each material is fabricated as paired MEMS devices, with D1 used for model training and D2 used for validation. A two-stage risk-constrained model first quantifies reproducibility risk and then incorporates it into sensitivity modeling and candidate ranking. Reproducibility Risk was learnable from material and operating-condition descriptors, achieving a leave-condition-out receiver operating characteristic area under the curve (ROC-AUC) of 0.911 for stability discrimination. Incorporating predicted Reproducibility Risk into sensitivity modeling improved D2 prediction performance (R2 = 0.790; low-risk MAE = 0.103) and supported reproducibility-aware candidate ranking. Using this method, optimized sensing materials are identified for nine hazardous gases and integrated into MEMS sensors, with the highest response exceeding 10,000 (toward CH3SH at 1 ppm). Four additional device replicates are fabricated for each gas-material combination to evaluate reproducibility, demonstrating device-to-device variations typically below 20%. This work establishes a practical, reproducibility-aware screening method for M/SnO2-based MEMS gas sensors under a fixed dispensing-based fabrication workflow, and provides a transferable design principle for incorporating reproducibility into data-driven sensing-material discovery.