The increasing use of hydrogen in various applications, from vehicles to fuel cells, necessitates highly sensitive and reliable hydrogen sensors, especially in high-humidity environments. Current hydrogen sensors based on resistive, electrochemical, catalytic, and thermal conductivity principles often lack the humidity tolerance required for safe and efficient operation. Optical nanoplasmonic sensors, utilizing localized surface plasmon resonance (LSPR) in metal nanoparticles, offer a promising alternative due to their high sensitivity and selectivity. However, these sensors are also susceptible to deactivation in humid environments. This study addresses this limitation by developing a novel nanoplasmonic hydrogen sensor that combines optimized material selection (Pd70Au30 alloy nanodisks), elevated operating temperature, and advanced data analysis techniques based on deep learning (Deep Dense Neural Networks (DDNN) and Transformers) to achieve superior performance in humid air. The high flammability of hydrogen-air mixtures makes accurate, humidity-tolerant detection critical for safety and process monitoring in fuel cells, electrolyzers, and other hydrogen-related technologies. The paper emphasizes the importance of overcoming the humidity challenge to enable widespread and safe adoption of hydrogen technologies.

Existing hydrogen sensors based on various transduction principles (resistive, electrochemical, catalytic, thermal conductivity) are reviewed. While some achieve high sensitivity and selectivity in dry conditions, their performance often degrades significantly in humid environments. Recent research on optical nanoplasmonic sensors, utilizing LSPR in Pd or Pd-alloy nanoparticles, shows promise but has limitations concerning humidity tolerance. Studies have shown that plasmonic Pd80Co20 sensors lose response magnitude in high humidity conditions. The need for humidity-tolerant sensors compliant with standards like ISO 26142:2010, which specifies requirements for sensor robustness across a range of relative humidity (RH) levels, is highlighted. The lack of plasmonic hydrogen sensor research addressing humidity tolerance above 40% RH is also discussed as a crucial gap in the current literature.

The study uses Pd70Au30 alloy nanodisks fabricated using hole-mask colloidal lithography. The sensor's performance is systematically evaluated under various conditions, including different relative humidity (RH) levels (0-80%), hydrogen concentrations (0.06-1.3%), and operating temperatures (30-130 °C). A detailed protocol based on ISO 26412:2010 is employed for testing sensor response and robustness. The core testing involves multiple hydrogen concentration pulses in synthetic air at various RH levels. The effect of sensor temperature on humidity resistance is investigated. Initially, the sensor response is analyzed using a traditional approach, measuring the peak centroid shift (Δλpeak) of the LSPR peak. Subsequently, machine learning algorithms, specifically Deep Dense Neural Networks (DDNNs) and Transformers, are employed for data analysis. These algorithms process the entire spectral response of the sensor, rather than just a single peak parameter, to improve accuracy and sensitivity. The DDNN is initially used for shorter data sequences, while the Transformer is employed for the long-term stability testing to handle longer time-series data. The limit of detection (LoD) is calculated using both the traditional Δλpeak method and the deep learning-based methods. Sensor robustness is evaluated based on the ISO 26142:2010 standard, which specifies response variation tolerance within the RH range of 20-80%. Long-term stability is assessed via an extended test lasting approximately 142 hours in 80% RH. The study also explores the performance of the deep learning models when operating outside their original training conditions (intermediate RH and lower H2 concentrations), showcasing the importance of model retraining for optimal performance.

The study reveals that humidity significantly deactivates the Pd70Au30 nanoplasmonic hydrogen sensor at ambient temperatures. Operating the sensor at elevated temperatures (80-130 °C) mitigates the negative effects of humidity. The use of deep learning (DDNN and Transformer models) significantly improves the sensor's limit of detection (LoD) in humid air, reaching 100 ppm (0.01%) H2 at 80% RH. This exceeds the US Department of Energy's target of 1000 ppm. The sensor's response is linear with respect to hydrogen concentration across a wide range and operating temperatures above 80°C. The DDNN-based readout improves LoD and enables the sensor to meet the ISO 26142:2010 robustness standard down to 0.06% H2 at 80 °C. Long-term stability testing (142 hours at 80% RH) shows no performance degradation. Retraining the deep learning models with data from intermediate RH conditions and lower hydrogen concentrations further enhances the sensor's performance and expands its operational range. The study also shows that while deep learning models may perform less optimally outside their training conditions, retraining effectively addresses this issue. This allows for adaptable data treatment tailored to different sensing environments without changing the hardware. The sensor shows excellent long-term stability, with no signs of performance loss after 140 hours of continuous operation in 80% RH. The use of a Transformer model, which is well suited to process longer time series data, significantly improves the stability and reproducibility of the long-term test results.

The findings demonstrate the successful combination of materials engineering (Pd70Au30 alloy), operating conditions (elevated temperature), and data analysis (deep learning) to create a highly sensitive and robust nanoplasmonic hydrogen sensor suitable for high-humidity environments. The sensor’s performance significantly exceeds existing technology and meets crucial industry standards. The use of deep learning is crucial for achieving the low limit of detection in the presence of noise and non-linear behavior, highlighting the potential of AI-driven advancements in sensor technology. The ability to retrain the deep learning models for different operating conditions adds to the versatility of the sensor. The study’s results have implications for improving hydrogen safety and monitoring in various industrial and consumer applications.

This study demonstrates a significant advancement in nanoplasmonic hydrogen sensing, achieving a 100 ppm LoD in 80% RH. The combination of elevated temperature operation and deep learning-based data analysis overcomes the limitations of humidity interference, exceeding the US DoE target. The sensor shows robust performance and long-term stability. Future work should focus on optimizing sensor response time and exploring the applicability of this approach to other gas sensing applications and deep learning models.

The study primarily focuses on synthetic air and might not fully capture the complexity of real-world gas mixtures. The long-term stability test, while extensive, could be further extended for even more comprehensive evaluation. The performance of the deep learning models is dependent on the quality and breadth of the training data; retraining is necessary when operating conditions change. While the study demonstrates impressive performance, further validation in real-world scenarios is needed before widespread practical deployment.