Development and application of an artificial neural network for quick estimation of debris bed quenching

Lim, Sokh Kong (author), Uršič, Mitja (mentor), Suikkanen, Heikki (comentor)

In the event of a severe nuclear accident, loss of core cooling can lead to core degradation and melting, which could finally result in the formation of a porous debris bed outside the reactor pressure vessel. Ensuring debris bed coolability through top-flooding is important for preventing the progression of molten core-concrete interaction and the associated generation of non-condensable gases. Although deterministic thermal-hydraulic codes can simulate debris bed quenching, their high computational time limits their application to rapid severe accident assessments. This thesis therefore investigates an Artificial Neural Network as a computationally efficient approach for estimating quenching time. Top-flooding experiments from the FLOAT facility (IKE, Germany), tests F.32 and F.33, were evaluated to investigate the effect of bottom air injection. Experiments were simulated using the MC3D code (ASNR, France). A two-dimensional axisymmetric computational model was developed, and a coarse mesh configuration with a uniform initial debris temperature was selected based on sensitivity analysis. Using Latin Hypercube Sampling, 38000 MC3D simulations were generated by varying the initial debris bed temperature, water volumetric flow rate, and air mass flow rate. The Artificial Neural Network was implemented using PyTorch, with its hyperparameters optimized using Optuna. An ensemble of five independently trained Artificial Neural Network models was used to improve prediction robustness. The results confirmed that air flow increases the total quenching time. Correlation analysis showed that the initial debris bed temperature had the strongest positive correlation with total quenching time. The Artificial Neural Network ensemble achieved a coefficient of determination of 0.903 on the unseen test dataset. While an individual MC3D simulation required several hours of computational time, the trained Artificial Neural Networks predict quenching time in less than 1 second.

• Language: English
• Keywords: nuclear safety, severe accident, debris bed coolability, quenching modeling, artificial neural network
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FMF - Faculty of Mathematics and Physics
• Year: 2026
• PID: 20.500.12556/RUL-189418
• COBISS.SI-ID: 294090755
• Publication date in RUL: 07.10.2026

Secondary language

• Language: Slovenian
• Title: Razvoj in uporaba umetne nevronske mreže za hitro oceno hlajenja plasti razbitkov

Abstract

V primeru težke jedrske nesreče lahko izguba hlajenja sredice vodi v njeno degradacijo in talitev, kar lahko v končni fazi vodi tudi nastanek porozne plasti razbitkov zunaj reaktorske tlačne posode. Zagotavljanje njenega učinkovitega hlajenja s pomočjo zgornjega poplavljanja je pomembno za preprečevanje interakcije staljenega korija z betonom in s tem povezanega nastajanja nekondenzirajočih se plinov. Čeprav lahko z determinističnimi termo-hidravličnimi računalniškimi programi simuliramo hlajenje plasti razbitkov, njihove visoke računske zahteve omejujejo njihovo uporabo pri hitrih analizah težkih nesreč. V magistrski nalogi smo zato raziskali možnost uporabe umetne nevronske mreže kot učinkovitega računskega pristopa za hitro ocenjevanje skupnega časa hlajenja. Vpliv vbrizgavanja zraka z dna smo proučili na podlagi eksperimentov s poplavljanjem plasti iz zgornje strani F.32 in F.33, izvedenih na napravi FLOAT (IKE, Nemčija). Eksperimenta smo simulirali z računalniškim programom MC3D (ASNR, Francija). Razvili smo dvodimenzionalen osno simetrični računski model ter na podlagi analize občutljivosti izbrali konfiguracijo z grobo računsko mrežo in enakomerno začetno temperaturo razbitkov. Z uporabo metode vzorčenja z latinskim hiperkubom smo generirali 38.000 simulacij MC3D, pri čemer smo spreminjali začetno temperaturo plasti razbitkov, prostorninski pretok vode in masni pretok zraka. Umetno nevronsko mrežo smo implementirali z uporabo programa PyTorch, njene hiperparametre pa smo optimizirali z uporabo orodja Optuna. Za izboljšanje zanesljivosti napovedi smo uporabili skupino petih neodvisno naučenih modelov umetne nevronske mreže. Rezultati so potrdili, da tok zraka ovira povečuje skupni čas hlajenja. S korelacijsko analizo smo ugotovili, da ima začetna temperatura plasti razbitin najmočnejšo pozitivno korelacijo s skupnim časom hlajenja. Skupina umetnih nevronskih mrež je na nevidenem testnem naboru podatkov dosegla koeficient determinacije 0,903. Medtem ko je posamezna simulacija MC3D zahtevala več ur računskega časa, usposobljene umetne nevronske mreže napovejo čas hlajenja v manj kot eni sekundi.

• Keywords: jedrska varnost, težka nesreča, učinkovitost hlajenja plasti razbitkov, modeliranje hlajenja, umetna nevronska mreža