Critical Heat Flux (CHF) relates to the transition from nucleate boiling to film boiling. This transition leads to a significant decrease in heat transfer efficiency. Consequently, accurate CHF prediction is essential for the safety and performance of water-cooled nuclear reactors where thermohydraulic margins are critical. Over the years, numerous CHF prediction models have been developed, including mechanistic, empirical, artificial intelligence (AI) and machine learning (ML). This work proposes (1) a novel ensemble of neural networks (NNs) model aimed at enhancing accuracy by combining individually optimized NNs with varied architectures and hyperparameters and (2) an interpretable physics-informed neural network (PINN) that integrates governing physical laws into the learning process to improve prediction accuracy. For the ensemble model, systematic procedures are presented to identify and optimize the best NNs and to aggregate them into the optimal ensemble. As for the PINN, the Westinghouse (W-3) correlation, an empirical CHF correlation for water-cooled reactors, is integrated as a physical model to drive the learning process. Such correlation is simple to implement and effective for both subcooled and saturated boiling conditions. To interpret the PINN results, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) are used, providing insights into the influence of input variables compared to standard NN models. The two prediction models developed are validated using experimental CHF data made available by the Working Party on Scientific Issues and Uncertainty Analysis of Reactor Systems (WPRS) Expert Group on Reactor Systems Multi-Physics (EGMUP) task force on AI and ML for Scientific Computing in Nuclear Engineering projects, promoted by the OECD/NEA. The obtained results demonstrate that the proposed models outperform state-of-the-art models, with sensitivity analysis further confirming robustness across various input parameters.
Il Critical Heat Flux (CHF) si riferisce alla transizione tra nucleate boiling e film boiling. Questa transizione porta a una significativa riduzione dell'efficienza di trasferimento del calore. Di conseguenza, una previsione accurata del CHF è essenziale per la sicurezza e le prestazioni di reattori nucleari raffreddati ad acqua, dove i limiti termo-idraulici sono fondamentali. Nel corso degli anni sono stati sviluppati numerosi metodi per predire il CHF, tra cui modelli meccanicistici, empirici, di intelligenza artificiale (IA) e di machine learning (ML). Questo lavoro propone (1) un nuovo modello d’insieme di reti neurali (NNs) progettato per migliorare le prestazioni combinando modelli individualmente ottimizzati con architetture e iperparametri diversi, e (2) un Physics-Informed Neural Network (PINN), che integra la fisica del fenomeno nel processo di apprendimento della rete per migliorarne l'accuratezza. Per il modello d’insieme, vengono presentate procedure sistematiche per identificare i migliori modelli di NN e per aggregarli una volta ottimizzati in un ensemble ottimale. Nell'approccio PINN, la correlazione Westinghouse (W-3), una correlazione empirica del CHF per reattori raffreddati ad acqua, viene integrata come modello fisico nel processo di apprendimento. Questa correlazione è semplice da implementare ed è efficace sia per condizioni di ebollizione sottoraffreddata che satura. Per interpretare i risultati del PINN, vengono utilizzati i metodi SHAP (SHapley Additive exPlanations) e LIME (Local Interpretable Model-Agnostic Explanations), che forniscono informazioni sull'influenza delle variabili di input rispetto ai modelli standard di NN. I due modelli di previsione sono stati validati utilizzando dati sperimentali sul CHF messi a disposizione dal Working Party on Scientific Issues and Uncertainty Analysis of Reactor Systems (WPRS) Expert Group on Reactor Systems Multi-Physics (EGMUP), che ha promosso progetti di IA e ML per il calcolo scientifico in progetti di ingegneria nucleare, promosso dall'OECD/NEA. I risultati ottenuti dimostrano che i modelli proposti superano gli attuali metodi all’avanguardia, e un'analisi di sensitività conferma ulteriormente la robustezza di questi modelli rispetto a vari parametri input.
Critical heat flux prediction by ensemble and physics-informed neural networks
Gatti, Irene
2023/2024
Abstract
Critical Heat Flux (CHF) relates to the transition from nucleate boiling to film boiling. This transition leads to a significant decrease in heat transfer efficiency. Consequently, accurate CHF prediction is essential for the safety and performance of water-cooled nuclear reactors where thermohydraulic margins are critical. Over the years, numerous CHF prediction models have been developed, including mechanistic, empirical, artificial intelligence (AI) and machine learning (ML). This work proposes (1) a novel ensemble of neural networks (NNs) model aimed at enhancing accuracy by combining individually optimized NNs with varied architectures and hyperparameters and (2) an interpretable physics-informed neural network (PINN) that integrates governing physical laws into the learning process to improve prediction accuracy. For the ensemble model, systematic procedures are presented to identify and optimize the best NNs and to aggregate them into the optimal ensemble. As for the PINN, the Westinghouse (W-3) correlation, an empirical CHF correlation for water-cooled reactors, is integrated as a physical model to drive the learning process. Such correlation is simple to implement and effective for both subcooled and saturated boiling conditions. To interpret the PINN results, SHapley Additive exPlanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) are used, providing insights into the influence of input variables compared to standard NN models. The two prediction models developed are validated using experimental CHF data made available by the Working Party on Scientific Issues and Uncertainty Analysis of Reactor Systems (WPRS) Expert Group on Reactor Systems Multi-Physics (EGMUP) task force on AI and ML for Scientific Computing in Nuclear Engineering projects, promoted by the OECD/NEA. The obtained results demonstrate that the proposed models outperform state-of-the-art models, with sensitivity analysis further confirming robustness across various input parameters.File | Dimensione | Formato | |
---|---|---|---|
Thesis_Gatti.pdf
accessibile in internet per tutti
Descrizione: Thesis PDF
Dimensione
4.31 MB
Formato
Adobe PDF
|
4.31 MB | Adobe PDF | Visualizza/Apri |
Executive_Summary_Gatti.pdf
accessibile in internet per tutti
Descrizione: Executive Summary PDF
Dimensione
849 kB
Formato
Adobe PDF
|
849 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/230864