Abstract: The reactor operation digital twin provides real-time parameter and state estimation for the reactor during operation, providing input for subsequent safety parameter calculations. As one of the core modules of the reactor operation digital twin, the inverse problem solver is crucial to ensure the real-time and accuracy of parameter and state estimation. In the previous work, the inverse problem is solved based on an initial guess of the input parameter, and the quality is highly depend on the initial guess. In order to improve the accuracy and computational efficiency, a global-local search (GLS) method was proposed in this work. Firstly, the reduced order method and KNN (K-nearest neighbor) method were used to build a reduced forward model, with which one can compute the physical field for a given input parameter online. The inverse problem, i.e., the process of calculating from a set of observations the causal factors that produced them, was then solved using GLS. In the global search stage, the KNN method was used to predict an initial input parameter based on the observed values. In the local search stage, the Latin hypercube sampling (LHS) method was used to discrete the local neighborhood of the initial parameters. The theoretical observation values corresponding to the parameters were calculated based on the reduced forward model, which was also trained using KNN. The optimal input parameter was then determined with which the observations were closest to the actual observation values. Numerical tests were conducted on HPR1000 reactor operation digital twin. To simulate the power distribution of HPR1000, the burnup, the power rate, the control rod inserting steps and the inlet temperature of the coolant were selected as input parameters, and the neutronic code CORCA-3D was used to calculate the power distribution with the four-dimensional input parameters. The digital twin was constructed over a wide range of the four-dimensional input parameters. The synthetic observations, calculated with CORCA-3D, were used to simulate the real observations of the in-core self-powered neutron detectors (SPNDs). The inverse problem solvers with different methods were used to infer the four-dimensional input parameters using the synthetic observations. The accuracy and computational efficiency for parameter inference and the related physical field with and without noisy observations were investigated using GLS, KNN and LHS. Numerical results verify that the proposed GLS outperforms other methods, and provides real-time and accurate parameter and state estimation for reactor operation digital twin, laying the foundation for engineering practice.