Simultaneous Prediction of Causes and Consequences in Hydrogen-Related Accidents Using Transformer-Based Multi-Task Learning

Hydrogen is emerging as a critical alternative in the global transition to sustainable energy. However, its characteristics - such as high flammability, rapid diffusion, and low ignition energy - pose significant safety challenges in the operation of hydrogen systems. To prevent future accidents, a thorough understanding of hydrogen-related incidents is crucial. This work proposes a novel approach utilizing transformer-based multi-task learning to simultaneously predict accident causes and consequences from unstructured accident narratives within the Hydrogen Incident and Accident Database (HIAD 2.1). By fine-tuning a pre-trained BERT model, the accident narratives are processed to identify root causes while estimating the potential consequences. The proposed multi-task model uses shared representations, enabling efficient learning of both causal and consequence patterns from historical hydrogen accidents. Our results show that multi-task learning enhances the model's ability to generalize across multiple prediction tasks, outperforming single-task models. By suggesting likely causes and consequences, this methodology supports risk identification and assessment, contributing to the safer adoption of hydrogen technologies.

Keywords: Multi-task learning, Hydrogen safety, BERT-based models, Accident analysis.