Temperature-Modulated MEMS Gas Sensor Array Enabled by Physics-Informed Deep Learning for Thermal-Runaway Characteristic Gas Discrimination.

Gas monitoring is a promising method for early warning of thermal runaway, yet existing sensor systems suffer from insufficient miniaturization, inadequate discrimination ability, and severe drift-induced performance degradation. This work presents a 2 × 2 integrated single-chip MEMS gas sensor array with temperature modulation and deep learning for discriminating thermal-runaway characteristic gases. The array comprises four distinct temperature zones achieved by varying the line width of the heating electrodes, with the entire array measuring 1.8 × 1.8 × 0.5 mm3. Furthermore, it is capable of detecting four critical thermal-runaway characteristic gases (H2, CO, CH4, and CO2) and exhibits distinct temperature-modulated response patterns. To achieve reliable multi-gas discrimination, we develop a Hierarchical Kolmogorov-Arnold Mixture-of-Experts (HK-MoE) model. Under single-gas conditions, the model achieved initial F1-scores of 99.53% for target gas presence detection, 99.88% for gas classification, and 98.09% for concentration determination. While performance experienced some degradation following a 30-day continuous operational aging, the model consistently demonstrated satisfactory resilience and significantly outperformed conventional baselines and state-of-the-art Transformer-based time-series models. Further validation using multi-component gas mixtures yielded overall F1-scores above 99.83% for all three tasks, demonstrating reliable mixed-gas discrimination capability. This integrated framework offers a compact solution to drift-induced instability, demonstrating its potential for reliable and early-stage battery safety monitoring.