Machine Learning for Materials-Degradation Prediction
Data-driven models predict how neutron irradiation and thermal cycling degrade fusion materials, guiding material choice and component lifetime estimates.
The degradation problem
Fusion neutrons and heat progressively alter structural and plasma-facing materials: displacement damage, helium and hydrogen production, embrittlement, swelling, and changes in thermal and mechanical properties. Predicting these changes across the many candidate materials and conditions is a large, expensive experimental and modeling task.
Where ML helps
- Predict property changes (hardness, ductility, conductivity) from composition and dose
- Interpolate across irradiation temperature, dose rate, and fluence
- Screen candidate alloys before costly irradiation campaigns
- Combine data from fission, ion-beam, and fusion-relevant sources
Models range from regression on curated property databases to graph and descriptor-based methods that encode composition and microstructure. They accelerate screening by flagging promising compositions and by highlighting where the data are sparse and new experiments are needed.
Data scarcity
Fusion-relevant irradiation data are limited because facilities that produce the right neutron spectrum are rare. Models must therefore lean on physics-based features and transfer from related irradiation environments, while being explicit about the uncertainty this introduces. Overconfident extrapolation to fusion spectra is the main hazard.
Physics grounding
Pure data fits can violate known trends, so models are often constrained or combined with mechanistic damage models. This hybrid approach keeps predictions consistent with radiation-damage physics while using data to calibrate the parts that theory does not pin down.
Lifetime and degradation estimates feed component design and safety cases for concepts like the Hyperion breeder, whose blanket and structure face intense neutron flux. These are computational predictions for materials in an unbuilt machine, validated against dedicated irradiation testing rather than treated as measured service performance.