Abstract
Bismuth-doped fiber amplifiers (BDFAs) can extend optical amplification beyond the C + L bands into the low-loss window of silica fibers, but each additional fiber normally requires extensive gain and noise-figure characterisation. Here, we demonstrate physics-guided source selection for few-shot transfer learning in Bi/Ge co-doped BDFAs. Four fibers spanning 5.4–22.9 mol% GeO 2 were characterised over 480 operating points each. Across 12 directed transfer-learning pairs, a two-descriptor feature distance based on GeCl 4 flow during fiber preform fabrication and 1385-nm absorption predicted transfer benefit. Selecting the physically matched source reduced the target-data requirement by up to 7x, and prospective validation on an unseen fifth fiber correctly identified the optimal transfer-learning source. More broadly, these results show that source quality, as captured by simple physical descriptors, can be more important than source-data volume in photonic transfer learning.
Funding Information
  • Engineering and Physical Sciences Research Council (EP/W028786/1)
  • Engineering and Physical Sciences Research Council (EP/T026197/1)
  • Engineering and Physical Sciences Research Council (EP/Z002567/1)

This publication has 15 references indexed in Scilit: