Uncertainty-aware AI-enabled framework for predicting hydrogen dispersion in confined spaces to enhance operator situational awareness
Authors: Niclas Flehmig, Mary Ann Lundteigen, Shen Yin
Organizations: Department of Mechanical and Industrial Engineering, Norwegian University of Science and Technology, Trondheim, Norway · Department of Cybernetics Engineering, Trondheim, Norway
We present a two-stage AI-enabled framework for enhancing operator situational awareness during compressed hydrogen bunkering in maritime confined spaces. In the event of a leak, the framework first estimates the unknown leakage rate from early sensor readings and then forecasts how the hydrogen cloud will disperse over the next time steps, providing operators with a predicted concentration map together with confidence bounds. The framework is based on two independent AI models building one AI system. The AI system is designed as a decision-support tool. It augments sensor readings, and thus enhances operator situational awareness and supports decision-making. This enables preventive safety barriers to be deployed before hazardous concentrations are reached, while an independent emergency shutdown system remains the ultimate safety layer. The system design incorporates safeguards to ensure the AI system operates within its competence envelope. Trained and tested on an open dataset of confined-space hydrogen releases, the framework achieves a mean absolute error of 0.052~g/s in leakage-rate estimation and ≈0.018 in hydrogen-concentration prediction, approaching the sensor precision bound of ±0.01. The hydrogen-concentration prediction provides well-calibrated uncertainty estimates that allow operators to judge the reliability of each forecast. These results suggest that the framework may provide early information about hazardous concentration build-up, potentially supporting preventive actions by operators.
Figures & tables
Figure 1: Generic system diagram for compressed H2 bunkering system based on [ 37 , 30 , 8 ]
Figure 2: Generic architectural pattern for AI system(s) in safety-related systems based on [ 9 ] . Raw data inputs first pass through an input guard that validates format, range, and integrity. The data then feed the AI system(s), a monitoring component that observes both inputs and internal AI processes, and a supervisory component representing the human operator. The AI outputs are aggregated by a voting component before being relayed to the downstream process. The monitoring component informs the supervisor, who retains authority to override the output of the AI system(s).
Figure 3: The placement of the hydrogen sensors in the experimental setup based on [ 11 ] .
Figure 4: Data processing steps for the hydrogen concentration dataset with their practical safety implications. Top row: (1) Raw data ( n=429514 ) showing extreme right skew. (2) Trimmed data ( n=182181 ) after removing idle periods before and after the exposed leakage interval; transformation step to remove non-leakage related data that would introduce a bias. (3) Log-transformed trimmed data revealing a bimodal structure with distinct not-exposed (near-zero) and exposed sensor populations; transformation reveals high number of not-exposed sensors that should be excluded from the dispersion forecast. Bottom row: (4) Not-exposed vs. exposed split using a threshold of 0.009 H 2 volume fraction, consistent with the sensor accuracy of ± 1%. (5) Distribution of the exposed sensors before transformation; represents relevant sensors for the forecast. (6) Log-transformed exposed subset showing an approximately Gaussian distribution; transformation enables Gaussian likelihood within the GPyTorch framework but comes with information loss.
Feature
Mean
Std
Min
Max
time t (s)
74.28
55.98
0.00
239.95
leakage rate m^ (g/s)
0.35
0.32
0.03
1.27
x (m)
0.46
0.20
0.00
0.90
y (m)
2.17
1.61
0.00
4.68
z (m)
0.49
0.31
0.00
0.80
H 2 conc. c (%)
0.06
0.06
0.00
0.31
Table 1: Descriptive statistics of the trimmed dataset ( n=182181 ).
Figure 5: Control structure and system design of the safety-related system within a bunkering system for compressed H2 . Filled boxes represent the ESD system that is independent and the safety function of the system. Cross-hatched boxes represent component that are introduced to enhance reliability of the AI system based on [ 9 ] . The sensors are part of the ESD system and their readings are replicated for the use of the AI system.
Figure 6: Performance of the Stage 1 MLP ensemble for leakage rate estimation. Left: Predicted versus true mass flow for all test-set samples (MAE = 0.052 g/s); the dashed line marks perfect agreement. Middle: Per-experiment MAE, showing reliable estimation for low-to-moderate leaks but elevated error for a few high-flow scenarios. Right: Distribution of prediction errors (true minus predicted), with mean μ=−0.023 g/s; the negative bias indicates a slight tendency toward under-prediction.
Figure 7: Predictive mean (solid line) and 95% confidence interval (shaded) for the two best SVGP models at three sensor positions, with ground truth (green crosses). (a) Gaussian likelihood: accurate mean but overconfident uncertainty. (b) Beta likelihood: wider, better-calibrated bands that capture the ground truth more reliably, especially at m^≈1.0 g/s where training data are sparse.
Method
Likelihood
MAE
RMSE
NLL
VNNGP ( k=16 )
Gaussian
0.0248
0.0359
17.5418
SVGP ( M=6000 )
Gaussian
0.0207
0.0255
12.6712
SVGP ( M=4000 )
Beta
0.0209
0.0267
−1.2255
SVGP ( M=6000 )
Beta
0.0176
0.0226
−1.5340
Table 2: Test-set performance of the best Stage-2 GP models with light additive kernel.
Figure 8: Calibration of the GP dispersion model (Beta likelihood) on the held-out test experiments (exposed sensor readings, n = 14,694; 1000 predictive Monte Carlo samples per point). (a) Empirical versus nominal coverage of the predictive quantiles. Red markers show PICP for the central 50–99% intervals. The curve above the diagonal indicates conservative prediction intervals. (b) PIT histogram centered at 0.508 but narrower than uniform, confirming mild over-coverage.
Table 4: Sample counts for datasets generated with different exposure thresholds. The exposed share is 64.1 %, 63.8 %, and 63.6 % for thresholds 0.0085, 0.0090, and 0.0095, respectively.
Configuration
MAE
RMSE
NLL
Two-stage (MLP est. m^ )
0.0249
0.0318
−1.421
Oracle (true m )
0.0176
0.0236
−1.534
Table 5: Performance comparison of two-stage approach versus oracle.
Figure 9: Hydrogen concentration measured at six representative sensors (1, 8, 10, 15, 20, 24) over all tests. All sensors exhibit a rapid transient rise during the first approximately 50–100 s, followed by a quasi-steady-state plateau. The post-transient behavior is largely independent of time, indicating that temporal evolution is primarily relevant during the initial dispersion phase. For the majority of the leakage duration, sensor position and mass flow rate dominate the concentration mapping.
Figure 10: Overall distribution of leakage rates in the trimmed dataset. The empirical distribution is heavy-tailed, with dominant peaks near 0.1 g/s and in the 0.4–0.5 g/s band. Data become scarce above 0.8 g/s and at the lowest rates, which limits model accuracy at the extremes.
The integrity of hydrogen infrastructure relies on reliable leak detection, performed almost exclusively via manual tracer gas sniffing in electrolyzer manufacturing. Although mandated by standards, the lack of spatial probe guidance instructions leaves detection reliability entirely to operator execution, further compromised by sensor signal delays. This study quantifies how sniffer trajectory kinematics affect detection reliability at small-scale pipes and fittings, a near-field regime largely neglected by macroscopic dispersion research. Using a robotically guided test bench to eliminate operator variability, static concentration fields and dynamic trajectory passes were acquired across representative geometries under standardized leak rates (5 vol% hydrogen in nitrogen) and varying scanning velocities. Results demonstrate that scanning velocity and spatial probe orientation strongly dictate detectability. Conventional linear trajectories frequently miss leaks under dynamic conditions, causing severe false negatives. Conversely, geometry-specific routing, such as circumferential plunging paths around sealing points, maintains a high safety margin. From these observations, geometry-specific routing rules and a reduction-factor model for dynamic signal loss are derived. The findings reveal that current standard operating procedures pose a tangible safety risk. To operationalize these rules, a proof-of-concept software pipeline is presented, generating validated trajectories directly from 3D models for visualization in assistance systems.
Christian Masuhr, Arne Wendt, Thorsten Schüppstuhl
Institute of Aircraft Production Technology (IFPT), Hamburg University of Technology (TUHH), Denickestraße 17, Hamburg, 21073, Germany
Hydrogen infrastructure in enclosed environments, such as parking facilities for fuel cell vehicles, presents significant safety challenges due to hydrogen's low ignition energy and wide flammability range. Current monitoring systems are largely reactive, detecting leaks only after hazardous concentrations have formed. This study develops a computational framework for proactive sensor placement optimization by integrating computational fluid dynamics (CFD), genetic algorithm (GA) optimization, and a DeepSets neural surrogate. A CFD database of 180 scenarios was generated for a representative 50 m x 30 m x 3 m garage, covering multiple leak positions, rates (1-150 g/s), and ventilation conditions (ACH = 3-10 per hour). Sensor placement was optimized using a multi-objective GA and compared with uniform, random, and surrogate-assisted approaches. The GA achieved a detection rate of 96.1% within 60 s and reduced blind areas to 0.12%, corresponding to an approximately 5% improvement in composite fitness over a uniform baseline. The DeepSets surrogate reproduced near-optimal configurations with a fitness gap below 0.01 while reducing CFD evaluations by 89% and computational time by two orders of magnitude. Detection performance and spatial coverage remained comparable to the GA, demonstrating that surrogate-assisted optimization can retain solution quality while enabling rapid design iteration. Overall, the results show that CFD-informed optimization improves detection effectiveness and reduces sensor requirements compared to conventional layouts. The proposed framework supports scalable deployment and provides a foundation for integrating optimized sensor networks with digital twin systems for real-time monitoring and risk assessment.
Fangnian Wang, Nicholas Tan Jerome, Thomas Jordan +1
Karlsruhe Institute of Technology, Hermann-von-Helmholtz-Platz 1, 76344 Eggenstein-Leopoldshafen, Germany.
Artificial intelligence systems often disclose uncertainty, yet they rarely make clear what response that uncertainty should trigger. Most uncertainty visualizations encode uncertainty in model outputs, leaving users to discern the most appropriate course of action. A second region of the design space--uncertainty in the decision process itself, including how multiple uncertainty conditions compose into an oversight response-- remains comparatively underexplored. We address this gap with two coupled contributions. First, we introduce an uncertainty-to-action binding framework that composes multiple uncertainty conditions into a single oversight response under a precedence policy with a contextual safety modifier. That response concerns whether and how an AI-supported decision may proceed, not the substantive domain decision itself. Second, we present ActionCue, a process-transparency visualization that renders that composition explicit. We demonstrate the approach through a three-way comparison with confidence-only and data-level uncertainty displays, using worked cases from healthcare, credit assessment, and disaster forecasting. Together, the framework specifies how uncertainty conditions are resolved into an oversight response, and the visualization makes that resolution inspectable rather than implicit.
Chisom Anyabolu, Akshat Dubey, Georges Hattab
Robert Koch Institute · Robert Koch Institute Freie Universitat, Berlin