Interpretability guided transfer learning approaches for tritium pedestal predictions

Artikel i vetenskaplig tidskrift, 2026

We explore transfer learning approaches to extend data-driven pedestal models trained on deuterium (D) plasmas to tritium (T) and DT mixtures. Specifically, we use models pre-trained on JET D pulses, and JET T/DT data for the transfer learning. We use model interpretability to guide our choice of transfer learning strategy. Analysis of model behavior post transfer learning reveals that sparsity and multicollinearity in the T/DT data lead to severe overfitting when fine-tuning the weights of the pre-trained neural network-based models. Therefore, we instead use a more robust and simple output calibration approach to facilitate the impact of isotope composition. This yields models with R2 between 0.66-0.87, performing significantly better than uncalibrated models though not matching the original D-only performance. The scaling coefficients obtained qualitatively agree with previous research, namely that the pedestal density scales positively with increased isotope mass, while pedestal temperature exhibits a weak negative scaling with isotope mass. This work highlights the importance of understanding model behavior in transfer learning to ensure reasonable functional mappings, which is particularly relevant for fusion research where sparse, multicollinear data are encountered.