Nuclear energy offers a promising solution to decarbonize the maritime industry. With the increasing urgency to meet climate goals, innovative nuclear reactor technologies are being investigated to serve as a clean alternative to traditional marine fuels. One of the most promising reactor designs for this purpose are small modular reactors based on the high-temperature gas-cooled reactor concept, due to their compact, passively safe design. However, accurately simulating and optimizing this type of reactor requires computationally intensive models. Therefore, there has been an increasing interest in the development of Reduced-Order Models (ROMs) to speed up reactor simulations. This thesis aims to construct a predictive, non-intrusive ROM that not only simulates the time evolution of the coupled neutronics and thermal-hydraulics dynamics of a high-temperature gas-cooled reactor, but can also predict the system’s behaviour under varying parameter sets (e.g. different initial conditions, material properties, or boundary forcing). The ROMs were developed based on a combination of proper orthogonal decomposition and sparse identification of nonlinear dynamics, and tested on a representative high-fidelity Full-Order Model (FOM). The FOM was developed as part of this thesis and simulates the coupled dynamics of the one-dimensional neutron diffusion equations and the one-dimensional heat equation, capturing the interaction between neutronics and thermal hydraulics. The ROMs were built with transient FOM data to predict the evolution of the neutron flux, the temperature and the precursor concentration in the reactor. The ROMs were evaluated on both accuracy and computational efficiency. The ROM developed to simulate the high-fidelity FOM achieved a reduction in spatial dimensions from 14,318 Degrees of Freedom (DoF) to just 8 DoF, corresponding to a reduction factor of approximately 1,750, while maintaining high predictive accuracy. The ROM was trained and tested on a depressurized loss of forced cooling-like transient event, varying the heat transfer coefficient at the boundaries of the temperature domain as an input parameter. The maximum Relative Root Mean Square Error (RRMSE) the ROMs achieved was 9.9 × 10−3. The maximum RRMSE in the power was found to be 0.05, which corresponds to a power of 4.5 kW, and the maximum RRMSE in the temperature was found at 1.5 × 10−3 , translating to an error of 1.2 K. Additionally, the ROM demonstrated accurate predictive performance when tested on transients with time-dependent heat transfer coefficients, despite being trained only on constant-coefficient transients, highlighting its potential for control-oriented applications. The ROM’s accurate predictive capabilities and significant reduction highlight its potential as a valuable tool for reactor design and safety analysis. Additionally, the way the ROM treats the external forcing parameter shows promising results for control applications.