Accurate prediction of nuclide composition evolution during fuel depletion typically requires computationally intensive transport calculations to determine 1-group cross sections for all transmutation reactions in the reaction network. These cross sections are influenced by the neutron energy spectrum, which varies with both depletion and changing state parameters, such as fuel and moderator temperatures. This study investigates the feasibility of a machine learning-based model designed to predict nuclide composition directly from power history and state parameters, bypassing the need to calculate group cross sections and solve the Bateman equation. Model testing on an unseen power and state parameter history showed a mean absolute relative error of 0.471% for 235U. Due to its rapid inference speed, this approach is well suited to applications where traditional transport calculations are prohibitively expensive.

 This thesis investigates the application of advanced machine learning (ML) techniques for reactor physics applications, specifically in the area of lattice and nodal calculations. The research is divided into two main areas. The first focuses on accelerating the generation of few-group cross section (FGXS) data and the second area focuses on the use of ML for accurate predictions of nodal parameters and their uncertainty estimates.   In the first area, ML-based surrogate models were developed to predict multi-group cross section (MGXS) libraries for nuclides commonly found in pressurized water reactors (PWRs), serving as efficient alternatives to conventional tools such as XSPROC. Additionally, a hybrid model incorporating two deep neural networks (DNNs) with a linear blending scheme was proposed to simulate the depletion-driven evolution of nuclide compositions within fuel pellets. To manage the high dimensionality of MGXS data, Principal Component Analysis (PCA) was employed to construct a reduced latent space, which was then mapped using DNNs conditioned on reactor state parameters. Recurrent neural networks (RNNs) were also evaluated for modeling fuel depletion behavior. Two RNN variants—the “Direct NN” and the “Difference NN”—were developed and compared, and a nuclide-specific blending parameter optimized during training was introduced to enhance predictive performance without requiring additional training data.  The second area examines ML-based approaches for nodal data representation. Statistically rigorous model comparisons were performed using the non-parametric Wilcoxon, Nemenyi, McDonald-Thompson (WNMT) test. The study demonstrated that ML models can predict not only mean FGXS parameters but also their associated covariance matrices, capturing aleatoric uncertainty and enabling their integration into full-core uncertainty quantification frameworks. Both polynomial regression and DNN-based models were assessed, alongside the effects of descriptive statistical preprocessing. Results showed a strong dependence of model accuracy on training dataset size, with polynomial regression combined with descriptive statistics yielding the most accurate predictions on the largest dataset.

To compute few-group nodal cross sections, lattice codes must first generate multi-group cross sections using continuous energy cross-section libraries for each material in each fuel cell. Since the processing cost is significant, we propose representing the multi-group cross sections during lattice calculations using a pre-trained deep-learning-based model. The model utilizes a combination of Principal Component Analysis (PCA) and fully connected Neural Networks (NN). The model is specifically designed to manage extensive multi-group cross-section data sets, which contain data for several dozen nuclides and encompass more than 50 energy groups. Our testing of the trained model on a PWR assembly with a realistic boron letdown curve revealed an average relative error of around 0.1% for both fission and total macroscopic cross sections. The average time required for the model to generate the cross sections was approximately 0.01% of the time needed to execute the cross-section processing module.

A deep neural network model is proposed to predict assembly node cross section libraries based on fuel pellet nuclide concentrations and instantaneous state parameters.We tested the impact of two different factors on model p erformance.The first factor is the impact of averaging fuel pellet nuclide compositions and the second factor is the use of a single large DNN model versus the use of multiple DNNs for different cross section p arameters.The combination of the two factors result in four variations of the model architecture.All four architectures are rigourously tested and statistically compared using the Nemenyi posthoc test.The pair-wise Nemenyi results indicates that the averaging of fuel pellet nuclide concentrations had no impact on model performance.However, the results show a clear improvement of the representational model when a single large DNN model is used compared with multiple smaller DNN models.The selected best model with averaged fuel pellet compositions and a single large DNN model was tested on previously unseen data.The model was shown to have high accuracy with less than 1% mean relative error for the fission and total cross section values.

Measurement of the velocity field in thermal-hydraulic experiments is of great importance for phenomena interpretation and code validation. Direct measurement by means of Particle Image Velocimetry (PIV) is challenging in some multiphase's tests where the measurement system would be strongly affected by the phase interaction. A typical example can refer to the test with steam injection into a water pool where the rapid collapse of bubbles and significant temperature gradient makes it impossible to obtain main flow information in a relatively large steam flux. The goal of this work is to investigate the capability of the use of machine learning for the flow reconstruction of the jet induced by steam condensation from sparse temperature measurement with ThermoCouples (TCs). Two frameworks of (i) 'FDD' using pure data-driven modeling and (ii) 'FPINN' combining data-driven and Physics-Informed Neural Networks (PINN) are proposed and investigated. The frameworks are applied to a single-phase turbulent planar jet with data generated by CFD simulations.

To compute few-group nodal data, lattice codes first need to generate multi-group cross-sections for each constituent material within the lattice model. This generation process utilizes continuous-energy cross-section libraries, which is expensive in terms of the computing cost. Moreover, any alteration in the nuclide compositions or other state parameters necessitates the repetition of this process. To reduce the computational demands, we propose the application of a pre-trained representational model. This model, which integrates Deep Neural Networks (DNN) and Principal Component Analysis (PCA) modules, is particularly beneficial in scenarios that require repeated multi-group data processing by the lattice code. In our previous research, we established that such a model could accurately generate multi-group data for fuel pellet materials. In the present study, we have broadened the scope of the model to encompass a more extensive range of materials typically found in pressurized water reactors, including zirc-alloy cladding and borated water moderators. We also show that the model can be trained on a wide spectrum of fuel enrichments. When integrated into lattice calculations, the errors introduced by the deep-learning-based representational model result in less than 1% deviation in the k eff and pin-power distribution. We have further refined the model to assess also the neutron fluxes in the fuel pellet and borated water. This refined model was subsequently employed to perform a flux-weighted collapse and generate few-group cross-section libraries for lattice calculation. The few-group libraries generated in this manner exhibited high accuracy and gave a low average k eff error and minimal errors in pin power distribution.

This study presents a comparative analysis of regression techniques for modeling few-group cross-section libraries as a function of state parameters. A key advancement is the prediction of both point estimates and covariance matrices, enabling the capture of aleatoric uncertainty in nodal data. We evaluated two regression approaches: polynomial-based linear regression (PR) and deep neural networks (DNNs). Model parameters were optimized using negative log-likelihood minimization and a descriptive statistics transformation was explored to improve accuracy.

Results indicate that model performance strongly depends on the number of training samples ($N_{\text{branch}}$). At $N_{\text{branch}} < 512$ PR and DNN models performed best, while for $N_{\text{branch}} > 512$ the transformed polynomial regression (tp-PR) model outperformed others. The tp-PR model accurately predicted $k_{\text{inf}}$, with mean error as low as 0.015\%. These findings showcase the feasibility of using linear regression and DNN models to predict mean and covariance estimates of few-group cross-section parameters for nodal codes.