CFD-informed deep learning for gas concentration field reconstruction and hazard mapping in longwall goaf

Accurate and rapid prediction of gas concentration fields and hazard-zone evolution in longwall goaf environments is critical not only for preventing methane-related explosions and oxidation-driven spontaneous combustion, but also for efficient drainage management and methane abatement. However, conventional tools such as computational fluid dynamics (CFD), while capable of delivering accurate spatial distributions of goaf gas concentration, require substantial computation time, specialised commercial software, and expert knowledge for model setup and interpretation. These constraints limit their practicality for routine hazard assessment or responsive operational planning. To address this gap, this study presents an integrated framework in the domain of longwall goaf hazard assessment to couple high-fidelity CFD simulations with a multi-output deep neural network for near-instantaneous, CFD-equivalent predictions. A scalar-to-spatial multilayer perceptron (MLP) encoder transforms a single gas emission rate into a latent spatial representation, which is decoded by five task-specific U-Nets to produce oxygen and methane concentrations, oxidation and explosive zones, and synergistic hazard maps. Trained on 2000 CFD-generated scenarios covering 10–20,000 l/s emission rates, the model achieves PSNR > 40 dB, SSIM > 0.99, and R² > 0.99, while reducing computation time from ∼6 h to < 1 s. Sensitivity analysis shows that beyond an emission sampling interval of 20, SSIM drops sharply at interval 40, with CH₄ and synergistic hazard predictions more sensitive to sparsity than O₂ and oxidation outputs. Compared with a continuous conditional generative adversarial network (CcGAN), the hybrid model demonstrates superior structural coherence, stable and consistent responses to changes in emission rates, and strict boundary adherence, whereas CcGAN exhibits boundary distortions, excessive smoothing, and irregular local reversals. These advantages, together with its computational efficiency, position the hybrid model as a practical tool for generating high-resolution gas concentration and hazard maps on demand, enabling timely risk evaluation and proactive mitigation in underground longwall mining. It could also support more energy-efficient methane drainage and utilisation by guiding the targeted localisation of favourable drawdown zones and informing data-driven methane management strategies.