Deep learning for gas sensing using MOFs coated weakly-coupled microbeams

Gas sensors have been increasingly employed in a wide range of applications such as air quality monitoring, disease diagnosis, potential leakage of exhaust gases in industrial facilities, and food quality control. We propose a novel MEMS gas sensor made of two mechanically-coupled microbeams coated with metal organic frameworks and subject to electric actuation. The objective is to exploit the dynamic features of the microstructure to simultaneously detect the presence of two gases, namely carbon dioxide (CO₂) and methane (CH₄), and estimate their concentrations. A nonlinear mathematical model of the gas sensor is developed and verified in the nonlinear operating regime against experiments reported in the literature. Deep learning methods are integrated with the sensor's model to predict the gases’ characteristics from the dynamic features of the coupled microbeams. To do so, we generate a large set of dynamic responses of the sensor for varying operating conditions and use it to train deep neural networks to capture the nonlinear dependencies in the data. The results show evidence of high prediction accuracy in terms of gas type and concentration estimation. The optimal accuracy is achieved when all dynamic features of the two coupled beams are used as inputs. These include the first six natural frequencies, the RMS, minimum, and maximum values of the dynamic response of the coupled microbeams when actuated by a combination of DC and AC voltages near the veering point. Yet, the use of the dynamic features of one of the two microbeams is found to still predict the concentrations of both gases with a good accuracy level. As such, one can deploy one beam for actuation and the other beam for sensing both gases thanks to their mechanical coupling. We also compare our results to those obtained from classical statistical approaches, namely linear regression and support vector regression. The neural network based approach outperforms these two classical methods, especially when a limited number of data features is used as input.