The structural verification of a nuclear piping system with a complex configuration and a large number of elbows, each subjected to its own combination of pressure, thermal, and dynamic loads, requires a preliminary decision on the elastic or elasto-plastic approach, followed by a systematic procedure for the determination of safety margins according to the RCC-MRx design code. This work develops an end-to-end Python framework that automates the mechanical verification of elbow components and applies it to the Ex-Vessel Cooling System of the ITER Electron Cyclotron Upper Launcher. The framework couples an external Python driver with Ansys Workbench 2025 R2 and operates on a parametric design space spanning six nominal diameters (DN15–DN50), six wall schedules and four bend radii, for 144 unique elbow geometries. Five linear-elastic analyses are systematically executed across the full design space using the temperature-dependent material card of SS316L(N), and safety factors are computed deterministically for the sixteen RCC-MRx load categories governing the structural assessment. The resulting dataset of 87,552 verified entries is used to train two surrogate models. The Physics-Bounded Load-Space Interpolator (PB-LSI) is a deterministic interpolator returning code-ready safety factors with exact-node consistency. The EXplainable, Interpolation-augmented, Physics-Lifted Machine Learning model (XIPL-ML) is a histogram gradient-boosted regressor that predicts dimensionless stress amplification ratios with respect to an FEA-validated anchor, confining the machine learning component to learning the amplification law within the elastic domain. Predictions are explained via SHAP attribution decomposed over seven mechanically meaningful features. Cross-domain validation on a held-out DN32 geometry yields a mean absolute percentage error of 5.37%, and the leave-one-domain-out protocol confirms errors in the 5–10% range across the full design space without systematic bias. Future developments include a systematic elasto-plastic campaign, buckling assessment refinement, and extension to higher temperatures and irradiation damage regimes.
La verifica strutturale di un sistema di tubazioni nucleari con configurazione complessa e un elevato numero di gomiti, ciascuno soggetto a una propria combinazione di carichi, richiede una decisione preliminare sull'approccio elastico o elasto-plastico e una procedura sistematica per la determinazione dei margini di sicurezza secondo il codice RCC-MRx. Il presente lavoro sviluppa un framework end-to-end in Python che automatizza integralmente questa catena di verifica, applicato all'Ex-Vessel Cooling System del launcher Electron Cyclotron Upper di ITER. Il framework integra un Python driver esterno con Ansys Workbench 2025 R2 e opera su uno spazio parametrico definito da sei diametri nominali (DN15–DN50), sei schedule standardizzati e quattro raggi di curvatura, per 144 geometrie uniche. Cinque analisi lineari-elastiche vengono eseguite sistematicamente con le proprietà dipendenti dalla temperatura di SS316L(N), e i coefficienti di sicurezza sono calcolati deterministicamente per le sedici categorie di carico RCC-MRx. Il dataset risultante, composto da 87.552 verifiche, è utilizzato per addestrare due modelli surrogati. Il Physics-Bounded Load-Space Interpolator (PB-LSI) è un interpolatore deterministico che restituisce coefficienti di sicurezza code-compliant con consistenza esatta ai nodi simulati. Il modello XIPL-ML è un regressore gradient-boosted che non predice tensioni assolute, ma rapporti di amplificazione adimensionali rispetto a un'ancora fisica FEA-validata, confinando la componente ML all'apprendimento della legge di amplificazione nel dominio elastico. Le previsioni sono interpretate tramite attribuzione SHAP decomposte su sette feature meccanicamente significative. La validazione cross-domain su DN32 mai osservato in training restituisce un MAPE del 5,37%, e il protocollo leave-one-domain-out conferma errori nel range 5–10% su tutto lo spazio DN15–DN50, senza bias sistematico. I principali sviluppi futuri includono la campagna elasto-plastica sistematica, l'affinamento della verifica di buckling e l'estensione a temperature elevate e regimi di danno da irraggiamento.
From digital twin to physics-guided surrogate model for the structural verification of a nuclear piping system
Galli, Matteo
2025/2026
Abstract
The structural verification of a nuclear piping system with a complex configuration and a large number of elbows, each subjected to its own combination of pressure, thermal, and dynamic loads, requires a preliminary decision on the elastic or elasto-plastic approach, followed by a systematic procedure for the determination of safety margins according to the RCC-MRx design code. This work develops an end-to-end Python framework that automates the mechanical verification of elbow components and applies it to the Ex-Vessel Cooling System of the ITER Electron Cyclotron Upper Launcher. The framework couples an external Python driver with Ansys Workbench 2025 R2 and operates on a parametric design space spanning six nominal diameters (DN15–DN50), six wall schedules and four bend radii, for 144 unique elbow geometries. Five linear-elastic analyses are systematically executed across the full design space using the temperature-dependent material card of SS316L(N), and safety factors are computed deterministically for the sixteen RCC-MRx load categories governing the structural assessment. The resulting dataset of 87,552 verified entries is used to train two surrogate models. The Physics-Bounded Load-Space Interpolator (PB-LSI) is a deterministic interpolator returning code-ready safety factors with exact-node consistency. The EXplainable, Interpolation-augmented, Physics-Lifted Machine Learning model (XIPL-ML) is a histogram gradient-boosted regressor that predicts dimensionless stress amplification ratios with respect to an FEA-validated anchor, confining the machine learning component to learning the amplification law within the elastic domain. Predictions are explained via SHAP attribution decomposed over seven mechanically meaningful features. Cross-domain validation on a held-out DN32 geometry yields a mean absolute percentage error of 5.37%, and the leave-one-domain-out protocol confirms errors in the 5–10% range across the full design space without systematic bias. Future developments include a systematic elasto-plastic campaign, buckling assessment refinement, and extension to higher temperatures and irradiation damage regimes.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_Galli.pdf
solo utenti autorizzati a partire dal 06/07/2027
Descrizione: Testo tesi
Dimensione
29.55 MB
Formato
Adobe PDF
|
29.55 MB | Adobe PDF | Visualizza/Apri |
|
2026_07_Galli_Executive Summary.pdf
solo utenti autorizzati a partire dal 06/07/2027
Descrizione: Executive Summary
Dimensione
3.01 MB
Formato
Adobe PDF
|
3.01 MB | 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/10589/261158