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

Area of Science:

• Materials Science
• Sensor Technology
• Machine Learning

Background:

• Developing new gas sensing materials for air pollutant detection is crucial but labor-intensive.
• Current machine learning (ML) methods often overlook device reproducibility, limiting practical applications of predicted materials.

Purpose of the Study:

• To develop a risk-constrained ML method for screening gas sensing materials that accounts for device-to-device reproducibility.
• To accelerate the discovery of reliable sensing materials for Micro-Electro-Mechanical Systems (MEMS) gas sensors.

Main Methods:

• A unified fabrication process created ~9000 measurements from paired MEMS devices using metal-loaded SnO2.
• A two-stage risk-constrained ML model quantified reproducibility risk and integrated it into sensitivity modeling.
• Material and operating-condition descriptors were used to predict reproducibility risk.

Main Results:

• The reproducibility risk model achieved a ROC-AUC of 0.911 for stability discrimination.
• Incorporating reproducibility risk improved prediction performance (R² = 0.790) and enabled reproducibility-aware ranking.
• Optimized materials were identified for nine hazardous gases, with device-to-device variations typically below 20%.

Conclusions:

• The developed method provides a practical, reproducibility-aware screening approach for M/SnO2-based MEMS gas sensors.
• This work offers a transferable design principle for integrating reproducibility into data-driven sensing material discovery.