[Background] In nuclear thermal propulsion reactor engineering

the real-time prediction of multi-physics temperature fields under transient operating conditions has long been hindered by the inherent computational inefficiency and high dimensionality of coupled neutronics-thermal-hydraulic simulations. Traditional computational fluid dynamics (CFD) methods

while accurate

struggle to meet real-time decision-making requirements due to their prohibitive computational costs. [Purpose] This study aims to develop a lightweight digital twin framework for achieving real-time

high-fidelity temperature field predictions in nuclear reactors

thereby enabling rapid safety assessment and control strategy optimization. [Methods] By constructing a lightweight architecture of "encoding–operator learning–decoding"

temperature field data of 230

000–1.8 million dimensions are compressed into a 100-dimensional latent space. An improved Deep Operator Network (DeepONet) is used to learn the dynamic evolution within this latent space

and a physics-constrained decoder is employed for high-precision reconstruction. [Results] Validation using a nuclear thermal coupling benchmark dataset generated by the computational fluid dynamics software OpenFOAM shows: in 40-second iterative predictions

the average relative error of the fuel region temperature field is below 1%

and the coolant temperature field error is below 0.5%; in 100-second long-step predictions

the maximum relative error remains under 0.5%. Each iteration of the predictive model takes less than 200 seconds

achieving over 500 times acceleration compared to traditional CFD methods; a single long-step prediction takes 79.23 seconds

which is shorter than the prediction interval

thus meeting real-time prediction requirements. [Conclusions] This study provides a lightweight digital twin solution for transient condition simulation and safety control of nuclear thermal propulsion reactors.