Machine learning ushers in a new era for advanced nuclear materials research

Shanghai, August 21, 2025 — Nuclear energy is widely recognized as one of the most promising clean energy sources for the future, but its safe and efficient use depends critically on the development of robust nuclear fuels and structural materials that can endure extreme environments. A newly published review in AI & Materials highlights how machine learning (ML) is transforming this field, enabling scientists to accelerate discoveries, optimize performance, and overcome long-standing challenges in nuclear materials research.

The article, titled “Machine learning in research and development of advanced nuclear materials: a systematic review for continuum-scale modelling,” was authored by Chaoyue Jin and Professor Shurong Ding from the Institute of Mechanics and Computational Engineering at Fudan University. It provides an in-depth systematic review of how ML has been applied to nuclear fuels and structural materials, particularly at the continuum scale.

Why Nuclear Materials Matter

Inside a nuclear reactor, materials are exposed to intense neutron irradiation, high temperatures, mechanical stresses, and corrosive chemical environments. These conditions cause dynamic and coupled thermo-mechanical responses, such as swelling, embrittlement, creep deformation, and loss of thermal conductivity. If not properly understood and controlled, such behaviors can compromise reactor safety.

Traditionally, characterizing these effects has required expensive, time-consuming, and sometimes dangerous irradiation experiments. While theoretical modeling and large-scale numerical simulations have provided valuable insights, the complexity of multi-scale interactions often limits their predictive accuracy.

This is where machine learning comes in.

The Role of Machine Learning

Over the past decade, ML techniques have demonstrated remarkable potential in materials science, from alloy design to catalyst discovery. In the nuclear field, ML is now emerging as a key tool to:

• Analyze complex microstructures: Convolutional neural networks (CNNs) are used to identify grain boundaries, porosity, and irradiation damage in fuels and claddings, extracting patterns that are invisible to the human eye.

• Predict thermal conductivity and mechanical behavior: Deep neural networks (DNNs) and regression models can rapidly estimate properties such as thermal conductivity or yield strength, accelerating the evaluation of material performance.

• Optimize processing and fabrication: ML models link manufacturing parameters with final material microstructures, enabling researchers to fine-tune processes like annealing or rolling to achieve superior performance.

• Integrate with physics-based models: By combining ML with finite element simulations (FEM), researchers are building hybrid frameworks that can generate synthetic datasets, reduce reliance on costly experiments, and ensure that ML predictions remain physically meaningful.

Challenges and Limitations

Despite its promise, the review emphasizes that applying ML in nuclear materials research is not without hurdles. The scarcity of high-quality irradiation datasets remains a major bottleneck, as collecting reliable in-reactor data is both dangerous and expensive. Moreover, selecting the right features that faithfully capture the underlying physics—such as porosity, grain boundary evolution, or irradiation dose—is a non-trivial task.

Another challenge lies in the interpretability of ML models. “Black-box” predictions are insufficient for nuclear applications, where safety and reliability are paramount. The authors argue that future efforts must focus on hybrid ML–physics approaches, embedding physical laws and mechanistic insights directly into ML frameworks.

Looking Ahead

The review outlines several promising directions for future research:

• Hybrid frameworks that integrate ML with physical constraints and governing equations.

• Time-dependent modeling that captures temporal and spatial correlations in materials under irradiation.

• Physics-informed dataset generation using FEM to overcome the scarcity of experimental data.

• Inverse design approaches, where ML can suggest the optimal composition or microstructure to achieve a desired performance.

According to Professor Shurong Ding, “Machine learning is not a replacement for physics-based understanding, but a powerful partner. By combining the strengths of both, we can significantly shorten the development cycle of advanced nuclear materials and enhance reactor safety.”

About the Article

• Title: Machine learning in research and development of advanced nuclear materials: a systematic review for continuum-scale modelling

• Authors: Chaoyue Jin, Shurong Ding*

• Journal: AI & Materials, 2025(2):0012

• DOI: https://doi.org/10.55092/aimat20250012