Answers are grounded in this paper and cite their sources.
• Includes a 20,000-case spent-fuel decay heat and pool thermal response benchmark. • Physics-informed networks penalize thermal and decay-heat inconsistencies. • Joint in-distribution normalized error is 52.2 % below the standard neural ensemble. • The selected model reaches 9.25 % mean absolute percentage error on calorimetry. Rapid spent fuel pool (SFP) assessment requires fast, reliable estimates of decay heat and thermal response across heterogeneous fuel inventories and degraded cooling states. This study establishes a new, reproducible benchmark by sampling three-cohort pressurized water reactor spent-fuel inventories, deriving decay heat from an open Serpent2 depletion library, and propagating the resulting heat load through a reduced-order subcooled pool energy balance. The proposed physics-informed neural network (PINN) embeds heat-balance consistency, physically correct transient direction, and decay-heat monotonicity into supervised learning. These constraints directly couple the heat-load and temperature targets and discourage physically inconsistent predictions under domain shift. To the authors’ knowledge, this is the first openly reproducible study to evaluate a problem-specific PINN for joint SFP decay-heat and bulk-temperature prediction using independent calorimetry and explicit domain-shift tests. The latter is compared on independent and identically distributed (IID) and domain-shift conditions with engineering and statistical baselines, multilayer perceptron (MLP) ensembles, physics-feature MLPs and decay-heat residual hybrids. Architectures and hyperparameters are selected using validation scores; performance is assessed through repeated holdouts and external CLAB calorimetry. Direct Serpent interpolation achieves 5.65 % MAPE on 16 measured 17 × 17 CLAB assemblies, reduced to 1.73 % by the reported 27.1 W bias correction. Across five random seeds, the selected PINN MLP attains an IID mean normalized RMSE of 0.031, reducing IID mean normalized RMSE by 52.2 % relative to the standard MLP ensemble, while achieving 9.25 % external decay-heat MAPE. The results demonstrate that problem-specific physics constraints materially improve rapid, simultaneous heat-load and thermal-response prediction.
No takes yet. Share an insight, caveat, or question.
Pinheiro et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: