利用機器學習輔助之Sb摻雜SnO₂感測陣列於室溫下實現選擇性氣體偵測

Abstract

金屬氧化物半導體因其優異的氣體感測特性，在氣體感測領域中具發展潛力。然而，其氣體選擇性較差，難以區分不同的目標氣體。本研究使用自架設的常壓化學氣相沉積(APCVD)系統，製備濃度範圍為0 ~ 4 wt%的Sb摻雜SnO₂薄膜，形成同一基礎材料的氣體感測器陣列，並在室溫下照射紫外光以增強感測能力。經由單一響應特徵與機器學習演算法，能有效辨識 CO、NH₃、H₂和 NO₂這四種工業氣體。 不同的Sb摻雜濃度主要影響材料的電子濃度、氧空缺與薄膜的表面形貌，進而形成不同的感測器「指紋」，改變對不同氣體的響應。本研究採用主成分分析(PCA)提取數據的主要特徵，並利用隨機森林(RF)、貝氏分類(NB)、支援向量機(SVM)與 K近鄰演算法(KNN)這四種演算法進行氣體分類。模型訓練過程中，透過10折交叉驗證（10-Fold CV）來優化各演算法的超參數設定。結果顯示，僅使用三種Sb摻雜濃度的感測器，即可達到100%的分類準確度。顯示結合單一響應特徵與機器學習的Sb 摻雜SnO₂氣體感測器陣列，具備高效氣體識別的潛力，為未來智慧氣體感測技術提供可行方案。

Metal oxide semiconductors have shown great potential for development in the field of gas sensing due to their excellent gas sensing properties. However, their poor gas selectivity makes it difficult to distinguish between different target gases. In this study, a self-designed and built atmospheric pressure chemical vapor deposition (APCVD) system was employed to fabricate Sb-doped SnO₂ thin films with doping concentrations ranging from 0 to 4 wt%. A gas sensor array was consequently built with the same base material, and ultraviolet (UV) light irradiation at room temperature was applied to enhance the sensing performance. The combination of single-response features and machine learning algorithms enabled the effective identification of four industrial gases, namely CO, NH₃, H₂, and NO₂. Different Sb doping concentrations mainly affect the material's electron concentration, oxygen vacancies, and surface morphology of the thin films, thereby creating distinct sensor "fingerprints" and altering their responses to various gases. Principal Component Analysis (PCA) was used to extract the key features from the sensing data. In addition, four supervised machine learning algorithms, namely Random Forest (RF), Naïve Bayes classifier (NB), Support Vector Machine (SVM), and K-Nearest Neighbors algorithm (KNN), were employed for gas classification. During model training, 10-fold cross-validation (10-Fold CV) was adopted to optimize the hyperparameters of each algorithm. The results demonstrate that using only three sensors with different Sb doping concentrations can achieve a classification accuracy of 100%. Furthermore, integrating the Sb-doped SnO₂ gas sensor array with single-response features and machine learning enables highly efficient gas recognition. Such integration represents a promising direction for the advancement of intelligent gas sensing systems.