From Empirical Ratio Tuning to Mechanistic Insight: Decoding NiO−ZnO Heterojunction Effects in Gas Sensing via Explainable Machine Learning

P-n heterostructures have been widely recognized as an effective strategy for enhancing the sensing performance of metal oxide semiconductor gas sensors. However, the regulatory mechanism underlying the NiO-ZnO composite ratio and its influence on gas sensing and recognition performance remains poorly understood. In this work, seven NiO-ZnO-based gas sensors with systematically varied molar ratios were designed and fabricated, followed by extensive sensing experiments toward six representative volatile organic compounds. The sensors were comprehensively evaluated in terms of response characteristics, sensitivity, stability, and gas recognition capability within a deep learning-based multitask framework. The results demonstrate a pronounced nonmonotonic dependence of overall sensing performance on the NiO-ZnO composite ratio, indicating that sensor optimization cannot be achieved through simple empirical ratio tuning. Within the constructed deep learning multitask network, the NiO-ZnO sensor with a molar ratio of 0.75 exhibits superior comprehensive performance in both gas species classification and concentration regression tasks. Furthermore, explainable artificial intelligence analysis based on SHapley Additive exPlanations and feature interaction networks reveals that the NiO-ZnO ratio regulates key response-intensity features, particularly the average response and maximum response, thereby reshaping feature-space separability and gas recognition performance. Combined with UMAP visualization, the optimal ratio range (0.6-0.8) exhibits more compact intra-class distributions and clearer inter-class boundaries, corresponding to the best recognition performance. These results establish an intrinsic link between material composition, response feature evolution, and gas recognition capability. These findings elucidate the nonlinear and mechanism-driven role of the NiO-ZnO ratio in modulating gas sensing performance and recognition behavior and provide a new paradigm for the structural optimization and interpretable design of composite gas sensors beyond empirical optimization.