Multi-objective optimization of supercritical carbon dioxide Brayton cycles using Bayesian physics-informed neural networks: A comprehensive analysis of energy, exergy, economic, and environmental performance

This study presents a physics-informed data-driven framework for the multi-objective optimization of supercritical CO₂ (sCO₂) Brayton cycles, aiming to simultaneously enhance thermodynamic performance, reduce the levelized cost of electricity (LCOE), and minimize carbon emissions. A high-fidelity dataset was generated using the Ebsilon process modeling. This dataset serves as the foundation for training a Bayesian physics-informed neural network (B-PINN), which integrates governing physical laws into its learning process to ensure prediction consistency with first-principles constraints. By combining physics-informed loss functions with uncertainty quantification through Monte Carlo Dropout, the B-PINN framework achieves robust multi-objective performance prediction. To address the inherent trade-offs among competing optimization objectives, based on the NSGA-II, a novel adaptive output-weighting strategy is proposed, which significantly improves regression performance across all output targets. Sobol sensitivity analysis reveals that the turbine inlet temperature plays a dominant role, while strong interactions between the working fluid ratio and compressor inlet temperature indicate complex nonlinear behavior. Compared to the bi-objective optimization, the four-objective strategy yields improvements of 7.55 % and 7.82 % in exergy and thermal efficiency, respectively, reduces the LCOE by 6.5 USD/MWh, and maintains carbon emissions below 95 t. In addition, the Pareto-optimal solutions obtained via NSGA-III are systematically mapped to distinct operational priorities. This classification provides actionable design guidelines tailored to varied industrial contexts, thereby enabling informed decision-making for scalable and sustainable sCO₂ power system implementation.