Nanoporous Gold as a VOC Sensor, Based on Nanoscale Electrical Phenomena and Convolutional Neural Networks

Abstract

Volatile organic compounds (VOCs) are prevalent in daily life, from the lab environment to industrial applications, providing tremendous functionality but also posing significant health risk. Moreover, individual VOCs have individual risks associated with them, making classification and sensing of a broad range of VOCs important. This work details the application of electrochemically dealloyed nanoporous gold (NPG) as a VOC sensor through measurements of the complex electrical frequency response of NPG. By leveraging the effects of adsorption and capillary condensation on the electrical properties of NPG itself, classification and regression is possible. Due to the complex nonlinearities, classification and regression are done through the use of a convolutional neural network. This work also establishes key strategies for improving the performance of NPG, both in sensitivity and selectivity. This is achieved by tuning the electrochemical dealloying process through manipulations of the starting alloy and through functionalization with 1-dodecanethiol.

Keywords: convolutional neural network; frequency response; nanomaterials; nanoporous; volatile organic compound.

Figure 1

SEM micrographs of electrochemically dealloyed samples: (a) Ag₇₇Au₂₀ (NPG—nanoporous gold); (b) Ag₇₇Au₂₃Pt₃ (NPG-Pt).

Figure 2

Proposed circuit model for NPG.

Figure 3

Saturated frequency responses of NPG (a) Resistance Frequency Response; (b) Imaginary Frequency Response under exposure to a saturated vapor pressure of acetone, ethanol, hexane, methanol, water, and a control in dry air. Resistance changes varied from 1 × 10⁻⁵ to 2.03 × 10⁻⁵ Ω for the different model compounds. Changes in the capacitance are observed by changes in the saturated reactance responses, via the increasing corner frequency in saturated responses relative to the dry air control. Only frequencies up to 7.5 kHz are plotted for clarity.

Figure 4

CNN Architecture for classification of the five model compounds.

Figure 5

Confusion matrix of test set predictions made by the convolutional neural network (CNN) (a classification network) based on frequency response data on NPG.

Figure 6

CNN architecture for regression of the five model compounds.

Figure 7

Regression model performance for NPG sensing of the five model VOCs studied. The reduced sensing performance towards nonpolar compounds, such as hexane, in comparison to polar compounds, such as water, can be seen by the smaller deviation in the response. Nonlinear bias can be observed by the skew at high concentrations.

Figure 8

Saturated frequency responses (a) Resistance Frequency Response; (b) Imaginary Frequency Response of NPG-Pt under exposure to a saturated vapor pressure of acetone, ethanol, hexane, methanol, water and a control in dry air. An increase in the resistance response can be observed in comparison to NPG, with changes ranging from 1 × 10⁻⁴ to 4.03 × 10⁻⁴ Ω for the different model compounds. Greater shifts in the corner frequency of the reactance response are also observed.

Figure 9

Confusion matrix of test-set predictions made by the CNN classification network based on frequency response data on NPG-Pt.

Figure 10

Regression Model Performance for NPG-Pt sensing of model VOCs. An improved performance compared to NPG can be observed across the model compounds, through smaller variations, due to the higher surface area of transduction.

Figure 11

Saturated frequency responses of tNPG-Pt (a) Resistance Frequency Response; (b) Imaginary Frequency Response under exposure to a saturated vapor pressure of acetone, ethanol, hexane, methanol, water and a control in dry air. In contrast to NPG and NPG-Pt, the saturated response for the all compounds except water demonstrate a decreasing resistance relative to the dry air control, due to the indirect sensing mechanism of the thiol layer. Resistance changes vary from 0.5 × 10⁻⁴ Ω–1.42 Ω. Reactance changes are also observed in tNPG-Pt samples allowing for multivariate sensing.

Figure 12

Confusion matrix of test set predictions made by the CNN classification network based on frequency response data on tNPG-Pt.

Figure 13

Regression Model Performance for tNPG-Pt sensing of model VOC. A significant improvement in the sensing of nonpolar compounds can be observed by the reduced variance in sensing performance, in compounds such as hexane. This performance improvement comes at the trade-off of decreased sensitivity towards polar compounds, such as water.