Introduction One challenge for chemical sensing is to discriminate target analytes with unknown concentrations since many chemical sensors respond to more than one chemical analyte at various concentrations [1]. As machine learning advances occur, researchers have seen the benefits of applying the methods for sensor training and pattern recognition [2]. Machine learning can help analyze patterns employing previously measured datasets for better and faster analyte discrimination and can dynamically update and optimize the discrimination algorithm for different sensing backgrounds. However, the approach requires a large amount of data to obtain an accurate and stable system. In this work, we report a camera-based measurement that can monitor multiple plasmonic sensors at the same time. This methodology allows one to monitor and record large amounts of sensor data in a short time. Five different analytes, i.e. acetone, ethanol, methanol, toluene and dry air, have been used in testing plasmonic sensors at varied concentrations from 200 nmol/mol to 10 μmol/mol. Machine learning was applied on the measured dataset to successfully classify and quantify the analytes. Method This work used Cu-BTC metal organic framework (MOF) coated nanohole array (NHA) plasmonic sensors described in detail in reference [3]. Figure 1(a) illustrates a SEM image of part of the Cu-BTC MOF coated NHA. The developed sensing platform works as a plasmonic sensor that shows an optical resonance peak in the visible wavelength range (between 400 nm to 700 nm). When an analyte is adsorbed by the coated MOF, the optical spectrum of the sensor shifts, leading to a color variation due to local refractive index changes [3]. The color changes can be captured by a CMOS camera. The plasmonic samples used in this work each have four sensing areas with the same parameters (size, period, coating material, etc.). By monitoring the four sensors in parallel using a CMOS camera, one can monitor the sensing platforms at the same time and thus can record more sensing responses in a shorter time.Figure 1(b) shows the experimental procedure of this work. First, a sample chip (with four Cu-BTC MOF coated sensing areas) was placed in a gas chamber and then it was exposed sequentially to dry air (reference and carrier gas in this work), acetone, ethanol, methanol and toluene at various concentrations, from 200 nmol/mol to 10 μmol/mol. The sensor responses to each of the analytes were recorded using a CMOS camera and stored as images. The pixel information of the sensing area was separated into two parts. The first part was tagged with analyte ID and the corresponding testing concentration and was sent to an artificial neural network (ANN) for sensor training. The second part was used in testing. The trained ANN predicted the analyte ID and concentration of the untagged input data, and then the predicted result was compared with the real data to verify the accuracy of the system. Results and Conclusions Figure 1(c) reports a classification result for the ANN-based training system using a relatively small proof-of-concept training/testing dataset. A circle means a hit while a cross means a miss. The figure shows that, even for testing with unknown concentrations, the system was able to classify the sensing analyte. The overall accuracy in this testing was 97%. Specifically, the accuracy for each analyte, i.e. air, acetone, ethanol, methanol and toluene were 95 %, 100 %, 97 %, 100 % and 100 %, respectively. Figure 1(d) shows an example of the regression result for acetone vapors. The blue dots show the training data with concentrations at (0, 0.2, 0.5, 1, 2, 5 and 10) μmol/mol. The red dots show the regression results with acetone vapor mixed with dry air at random concentrations (0 to 12) μmol/mol. The system successfully predicted the concentration of the testing data with an accuracy of 95 %. Note: in this experiment, an accurate quantification is defined as a prediction that has less than a 5 % difference from the actual value. The testing accuracy may vary due to different testing datasets and hyperparameters for the ANN, but it is estimated that larger datasets could lead to an even more stable and accurate system. The sensing platform and data analyzing method demonstrated in this work expanded the selectivity beyond that of a single sensor. It enables a novel and efficient chemical sensing method for next-generation smart sensing systems.
Read full abstract
Jun 26, 2020
Mengya Yang + 9
Open Access