Transformer-based deep learning architecture for multivariable radioactive source term inversion.

Area of Science:

• Nuclear Engineering
• Environmental Science
• Artificial Intelligence

Background:

• Radioactive source term estimation is vital for nuclear emergency response and consequence assessment, especially post-Fukushima.
• Accurate determination of release rate, height, and position is necessary for effective nuclear accident management.

Purpose of the Study:

• To develop and evaluate a transformer-based deep learning architecture for multivariable radioactive source term estimation.
• To assess the model's performance across various scenarios, including time-varying release parameters.

Main Methods:

• Utilized the CALMET-LAPMOD coupling model, validated by the Kincaid tracer experiment, to generate comprehensive datasets.
• Developed a Transformer deep learning model with Bayesian optimization for adaptive hyperparameter tuning.
• Constructed datasets for five scenarios: individual parameters (rate, height, location) and coupled variations.

Main Results:

• Achieved high accuracy in source term inversion, with R² > 0.96 for release rate and height.
• Demonstrated an average location prediction error of 1.19 km at a 95% confidence level.
• In the coupled scenario, R² remained > 0.92 for release rate and location, with R² of 0.72 for height.

Conclusions:

• The transformer-based deep learning model shows excellent performance for radioactive source term estimation.
• Feature ablation analysis highlights the importance of high-concentration monitoring points for optimizing network layout.
• The findings provide valuable insights for improving nuclear emergency preparedness and consequence assessment.