AI-augmented models for thermofluid behavior prediction in cold-climate-specific molten salt volumetric solar thermal receivers with integrated thermal energy storage

While molten salt volumetric solar thermal receivers with integrated storage are considered techno-economically feasible for industrial heat decarbonization in cold climates, accurately predicting their performance remains challenging, limiting the design and development of commercial-scale units. At high operating temperatures (>400°C), their thermofluid behaviour is governed by complex interactions between radiation-induced natural convection and volumetric heating, making accurate performance prediction difficult. Recent studies have reported mechanistic models to predict the system thermofluid response; however, due to unquantifiable radiation-induced optical property variations, simplifying assumptions, and the lack of accurate Nusselt number correlations, their accuracy was limited at the system boundaries and during long heating times. To overcome these challenges, this study proposes two physics-guided data-driven models. A Physics-Guided Neural Network (PGNN) integrates mechanistic models into its training process to predict spatiotemporal temperature distributions. In parallel, a Physics-Guided Long Short-Term Memory (PG-LSTM) model utilizes past experimental temperature profiles and receiver parameters to forecast vertical temperature distributions in real time. In both modeling approaches, a physics-based loss function derived from the law of conservation of energy is implemented to ensure physically plausible predictions. Validation against experimental data demonstrates strong predictive performance, with the PGNN achieving a Mean Absolute Error (MAE) below 1.0°C while maintaining physical plausibility under boundary sensor drifts expected in high-temperature environments, while the PG-LSTM maintains forecast errors below 2.0°C over 60-300 s. These results highlight the potential of physics-guided modeling approaches as a practical aid for commercial-scale design studies, virtual experimentation, and real-time decision support for next-generation clean heat technologies.