We have designed three-electrode mixed-potential sensors for real-time detection of NOx and NH3. The sensors were manufactured by ESL ElectroScience on an insulator-yttria-stabilized zirconia (YSZ)-insulator substrate. Dense electrodes of La0.8Sr0.2CrO3 (LSCO), Au0.5Pd0.5 (Au/Pd), and Pt and a porous layer of YSZ electrolyte have been deposited by screen printing. The Au/Pd+Pt pair is sensitive to NH3 under open circuit conditions, while under a negative current bias, the sensitivity of the LSCO+Pt pair to NOx is enhanced and the sensitivity to NH3 is suppressed. The tunable sensitivity allows us to quantify the concentration of NOx and NH3 separately. These mixed potential sensors are also expected to have fast response times, sufficient robustness in the exhaust gas conditions, and are manufactured in processes that are readily scaled up for mass production.
We take advantage of the ability of artificial neural networks (ANNs) to learn continuous functional relationships without human intervention to decode the sensor signals in the presence of cross interference. ANNs have also been successfully used in a broad range of mobile applications which makes them an attractive tool for developing portable analyzers and on-board vehicle diagnostic systems. We have previously demonstrated this technique can quantify mixtures of NOx, C3H8, and CO.3 The ANN model used for NOx/NH3 in Figure 1(a) consists of a three-layer feed-forward structure with 6 input neurons, 16 hidden layer neurons, and 2 output neurons. A peak test error in Figure 1(b) of less than 5% was achieved for decoding the signals from NOx/NH3 mixtures in the 50-200 ppm range. For practical use in automotive systems, stability on the order of 103 hours must be studied. Long term signal drift and changes in impedance response were monitored under repeated exposure to 50-200 ppm of NOx, C3H8, and NH3 in alternating biased and unbiased conditions at 530oC over a period of over 100 days.
 G. Busca, L. Lietti, G. Ramis, F. Berti, Appl. Catal. B Environ., 18, 1–36 (1998).
 M. Koebel, M. Elsener, M. Kleemann, Catal. Today, 59, 335–345 (2000).
 L. Tsui, A.D. Benavidez, P. Palanisamy, L. Evans, F.H. Garzon, ECS Trans., 75, 9–22 (2016).
Figure 1. (a) Artificial neural network model used to decode the signals from a three-electrode sensor operated in open circuit and unbiased mode. (b) Test error for the ANN shows a peak at 2.5% and 90% of the data points are confined to less than 15% error.