Machine Learning-Enabled ZnS/In2O3 Heterojunction Sensors for ppb-Level Room-Temperature Hydrogen Detection

Real-time, ppm-to-ppb-level detection of hydrogen (H2) remains essential for industrial safety, given its flammability and widespread use in energy applications. Portable H2 sensors, however, continue to suffer from insufficient sensitivity and reliance on elevated operating temperatures. Here, cubic bixbyite In2O3 nanocubes, synthesized by radio frequency sputtering at 600 °C, were functionalized via drop-cast decoration with ZnS nanoparticles to form an n-n heterojunction for H2 sensing at room temperature (∼32 °C). The ZnS/In2O3 architecture leverages a high surface area, oxygen vacancy-mediated adsorption, and catalytic spillover from ZnS to enhance H2 interaction and charge carrier modulation. Pristine In2O3 showed a relative response of 11.23% to 50 ppm H2, which increased to ∼23% upon optimal ZnS loading (48 wt %), with a detection limit of 86.9 ppb and response/recovery times of ∼90 s/∼488 s. To enable quantitative concentration estimation, a two-stage machine learning framework was developed using six transient response features (Smax, Trsp (90%), Trec (90%), ∫Rdt, mean dR/dt, and ∫R/s). Permutation importance identified Trec (90%), mean dR/dt, and ∫R/s as dominant predictors, while principal component analysis captured 91.8% variance with clear concentration-dependent clustering. Among four regressors evaluated via 5-fold cross-validation, Gaussian process regression achieved the best static prediction (root-mean-square error (RMSE) = 4.044 ppm, R2 = 0.945), while a 15 s sliding-window multilayer perceptron model with Bayesian-optimized hyperparameters enabled real-time tracking (RMSE = 3.837 ppm, mean absolute error = 1.263 ppm, and R2 = 0.924) across 0.5-50 ppm, with sub-5 ppm resolution well below the 4 vol % lower flammability limits. These results establish a combined materials and data-driven strategy for robust, ambient-temperature H2 monitoring.