Open Access

Gas distribution mapping (GDM) is essential for industrial safety and environmental monitoring, as it enables real-time hazard detection and air quality assessment. Traditional GDM methods, such as kernel-based techniques, struggle to reconstruct complex gas plume dynamics accurately. While deep learning has shown promise for GDM, two critical challenges hinder its practical use: the scarcity of available training data and the incompatibility of conventional architectures with irregular sensor layouts. To address these limitations, we propose GNN-DM, a graph neural network-based model for GDM that incorporates the relational structure of sensor networks to infer high-resolution maps from minimal, irregular inputs. The model is pretrained on synthetic gas dispersion data generated from measured wind data and fine-tuned on two industrial datasets collected on a ferry car deck and in a hot rolling mill. Compared with established GDM techniques, GNN-DM achieves higher accuracy on synthetic and real-world data, highlighting the potential of graph-based learning for practical gas mapping applications.