Learning nonlinear operators in latent spaces for real-time prediction of coolant temperature in small modular high-temperature gas-cooled reactors

In the field of nuclear thermal propulsion reactor engineering, real-time prediction of complex multi-physics temperature fields remains a critical challenge. To address this issue, this study proposes a latent space neural operator (L-DeepONet)-based approach for real-time prediction of temperature fields in nuclear thermal propulsion reactors. A lightweight "encoder-operator learning-decoder" framework is constructed by coupling an autoencoder (AE) with a deep operator network (DeepONet). First, high-dimensional temperature fields are compressed into a 100-dimensional latent space using AE. Subsequently, DeepONet is trained in the low-dimensional space to learn coolant dynamic evolution patterns. Finally, high-fidelity reconstruction of predictions is achieved through the decoder. Validation using OpenFOAM-generated coolant temperature field datasets demonstrates that the method achieves average relative errors below 1% for fuel temperature fields in both 40-second iterative predictions and 100-second long-term predictions, with errors for coolant and cladding below 0.5%. The training time of 79.23-192.83 seconds represents a two-order-of-magnitude acceleration compared to traditional CFD simulations, enabling real-time single-step long-term prediction with error distributions concentrated in gradient-sensitive regions. This work innovatively introduces latent space operator learning into multi-physics modeling of nuclear thermal propulsion reactors and achieves real-time prediction. The framework provides insights for real-time simulation and decision-making under extreme operating conditions, and can be extended to neutronics-thermomechanical coupling scenarios, offering new pathways for digital twin applications in advanced nuclear systems.