# A machine learning approach to thermal conductivity modeling: A case study on irradiated uranium-molybdenum nuclear fuels

## TL;DR

This paper introduces a deep neural network model to predict the thermal conductivity of irradiated uranium-molybdenum nuclear fuels, bypassing traditional microstructure-based models and enabling better predictions from limited historic data.

## Contribution

It presents a novel machine learning approach for thermal conductivity modeling that does not require detailed microstructural information, improving predictive accuracy and generalization.

## Key findings

- Achieved approximately 4% mean absolute percent error on validation data.
- Model generalizes well to unseen data, demonstrating robustness.
- Deep learning is viable for material property prediction with limited data.

## Abstract

A deep neural network was developed for the purpose of predicting thermal conductivity with a case study performed on neutron irradiated nuclear fuel. Traditional thermal conductivity modeling approaches rely on existing theoretical frameworks that describe known, relevant phenomena that govern the microstructural evolution processes during neutron irradiation (such as recrystallization, and pore size, distribution and morphology). Current empirical modeling approaches, however, do not represent all irradiation test data well. Here, we develop a machine learning approach to thermal conductivity modeling that does not require a priori knowledge of a specific material microstructure and system of interest. Our approach allows researchers to probe dependency of thermal conductivity on a variety of reactor operating and material conditions. The purpose of building such a model is to allow for improved predictive capabilities linking structure-property-processing-performance relationships in the system of interest (here, irradiated nuclear fuel), which could lead to improved experimental test planning and characterization. The uranium-molybdenum system is the fuel system studied in this work, and historic irradiation test data is leveraged for model development. Our model achieved a mean absolute percent error of approximately 4% for the validation data set (when a leave-one-out cross validation approach was applied). Results indicate our model generalizes well to never before seen data, and thus use of deep learning methods for material property predictions from limited, historic irradiation test data is a viable approach.