Rapid Identification of Chemical Warfare Agents by Artificial Neural Network Pruning of Temperature-Programmed Microsensor Databases

Rapid identie cation of three chemical warfare (CW) agents and a CW agent simulant has been achieved by analyzing the responses of an array of four microhotplate conductometric sensors with tin oxide and titanium oxide thin sensing e lms. Analyte concentration values in the range of nmol/mol (ppb) to mmol/mol (ppm) were also determined. Calculating the ratios of the response onset and recovery time constants of the different sensor materials at different temperatures, when operated in the e xed temperature sensing mode, clearly identie ed each CW agent. Training artie cial neural network (ANN) models from an 80-component response database (four sensing e lms at 20 operating temperature steps), obtained in the temperature-programmed sensing operating mode, led to successful individual analyte recognition and four separate agent concentration models to provide the concentrations of the target compounds. Recursive elimination of the less relevant inputs and ANN model re-training identie ed the 5 to 12 inputs that are sufe cient to identify and quantify the CW agents. The information obtained through pruning allows one to reduce the microsensor scan time by 40% to 80% and provides insight into the nature of the most critical gas‐ solid interactions for detection.