The deployment of advanced reactor systems in flexible, remote, and grid-following modes demands higher levels of autonomy in thermal-hydraulic operations. While digital twins (DTs) and machine learning–driven modeling have advanced separately, their integration into deployable autonomous control systems remains limited. This study bridges simulation and practice by developing neural ordinary differential equation (NODE) models trained on the DT-augmented data of a small-scale thermal-hydraulic test bed representing small modular reactor thermal systems. We systematically evaluated the performance of eight configurations, comparing four model architectures across two training setups. The model architecture included gated recurrent unit (GRU) baseline models, pure NODE, multilayer perceptron (MLP)-NODE, and GRU-NODE hybrid architectures. The training setup included experimental data and DT-augmented data. We found that the MLP-NODE architecture trained with DT-augmented data achieved the best performance for autonomous control capability. Critically, DT augmentation proved highly effective; this model achieved a 34% lower error than its counterpart trained on experimental data alone. This configuration yielded the lowest root-mean-square error across the actuator channels. Furthermore, validation on held-out transients, including load-following, rapid shutdown, and multistep power changes, demonstrated robust generalization. The model was also computationally efficient, with real-time inference averaging 70 ms per prediction, enabling real-time model predictive control. This approach, combining a continuous-time formulation with the MLP-NODE, offers a practical pathway using DT-augmentation capability toward real autonomous control in advanced reactors.
Lim et al. (Mon,) studied this question.