Intelligent Prediction of High-Enthalpy Gas Environments for Arc Heaters Based on the Bayesian Optimization Attention-Aware Network and Transfer Learning

Abstract Arc heaters are the core devices for simulating the aerothermal environment of hypersonic vehicles. To meet the demand for rapid and high-accuracy prediction of the high-enthalpy gas environment in segmented arc heaters, this study proposes an intelligent prediction method based on a Bayesian Optimization Attention-Aware Network (B-AAN) integrated with transfer learning. First, a feature-attention module is embedded into a multilayer perceptron to perform sample-dependent reweighting of the input features before they are passed to the shared multi-output prediction network. Subsequently, Bayesian optimization is employed to tune the model’s hyperparameters. Meanwhile, SHapley Additive exPlanations analysis is introduced to reveal the contribution mechanisms of different input features to the prediction results. Finally, a model transfer strategy is further incorporated to address the prediction of high-enthalpy gas environments under new operating conditions. The proposed model was validated using a multi-operating-condition database established from an arc heater experimental platform. The results show that, under conditions of different model initialization random seeds and fixed data partitions, this method achieves high prediction accuracy for arc voltage, arc chamber pressure, and enthalpy, with prediction errors below 3%. The transfer learning results further indicate that the model can still maintain high prediction accuracy and stability after fine-tuning with only a limited number of samples under new operating conditions. Compared with conventional machine learning methods, the method proposed in this paper provides more balanced predictive performance across the three outputs. Overall, the proposed B-AAN method combined with transfer learning provides an effective solution for the intelligent operation and maintenance of arc heaters.