Smart Miniature Mass Spectrometer Enabled by Machine Learning

Similar to smartphones, smart or automatic level is also a critical feature for a miniature mass spectrometer. Compared to large-scale instruments, miniature mass spectrometers often have a lower mass resolution and larger mass drift, making it challenging to identify molecules with close mass–charge ratios. In this work, a miniature mass spectrometer (the Brick-V model) was combined with intelligent algorithms to realize rapid and accurate identification. This Brick-V mass spectrometer developed in our lab was equipped with a vacuum ultraviolet photoionization (VUV-PI) source, which ionizes volatile organic compounds (VOCs) with minor fragments. Machine learning would be especially helpful when analyzing samples with multiple characteristic peaks. Four machine learning algorithms were tested and compared in terms of precision, recall, balanced F score (F1 score), and accuracy. After optimization, the multilayer perceptron (MLP) method was selected and first applied for the automatic identification and differentiation of ten different fruits. By recognizing the pattern of multiple VOCs diffused from fruits, an average accuracy of 97% was achieved. This system was further applied to determine the freshness of strawberries, and strawberry picking at different times (especially during the first 24 h at room temperature of winter) could be well discriminated. After building a database of 63 VOCs, a rapid method to identify compounds in the database was established. In this method, molecular ions, fragment ions, and dimer ions in the full mass spectrum were all utilized in the machine learning program. A satisfactory prediction accuracy for the 63 VOCs could be achieved (>99%).