Deep learning for temperature field prediction in a multi-sector swirl-stabilized combustor

Accurate prediction of combustor exit temperature fields is crucial for optimizing temperature uniformity and ensuring the durability and stability of downstream turbine components in aero-engines. This study investigates the potential and limitations of deep learning in modeling temperature fields, especially in the presence of limited experimental data and staging combustion. To this end, we introduce a physics-interpreted deep learning framework for predicting exit temperature fields in a tri-sector centrally-staged swirl combustor. A comprehensive experimental dataset of temperature fields was generated under varying fuel-to-air ratios (FAR), fuel staging ratios (SR), main stage swirl numbers, and single/dual-stage combustion modes. Two deep learning models, a fully connected network (FCN) and a Transformer network (TNN), were combined with the SHAP additive module to predict two-dimensional spatial temperature values based on FAR, SR, swirl number, and spatial coordinates. Results demonstrate that both models can reasonably capture the complex wavy multi-hotspot structure of the temperature field. The FCN model demonstrated superior generalization capability for unseen operating conditions with a mean relative error below 0.025. Furthermore, the FCN model provided more accurate predictions of overall temperature distribution factor (OTDF) compared to conventional semi-empirical models. Shaplye additive analysis demonstrate that swirling intensity dominates temperature pattern. Stronger swirling intensity is associated with negative Shapley values that may reduce local peak temperature values and promote a more uniform temperature field. We note limitations in prediction accuracy for the near-wall regions and dual-stage conditions. Nevertheless, these results demonstrate that this physics-interpreted deep learning framework can characterize temperature pattern in a cost-effective manner, offering a rapid method for optimizing aero-engine combustor performance during the design phase.