Accelerating Plasmonic Hydrogen Sensors for Inert Gas Environments by Transformer-Based Deep Learning

Abstract

The ability to rapidly detect hydrogen gas upon occurrence of a leak is critical for the safe large-scale implementation of hydrogen (energy) technologies. However, to date, no technically viable sensor solution exists that meets the corresponding response time targets set by stakeholders at technically relevant conditions. Here, we demonstrate how a tailored Long Short-term Transformer Ensemble Model for Accelerated Sensing (LEMAS) accelerates the response of a state-of-the-art optical plasmonic hydrogen sensor by up to a factor of 40 in an oxygen-free inert gas environment, by accurately predicting its response value to a hydrogen concentration change before it is physically reached by the sensor hardware. Furthermore, it eliminates the pressure dependence of the response intrinsic to metal hydride-based sensors, while leveraging their ability to operate in oxygen-starved environments that are proposed to be used for inert gas encapsulation systems of hydrogen installations. Moreover LEMAS provides a measure for the uncertainty of the predictions that is pivotal for safety-critical sensor applications. Our results thus advertise the use of deep learning for the acceleration of sensor response, also beyond the realm of plasmonic hydrogen detection.