Closing the performance gap with larger or retrained embeddings

Determine whether larger or retrained learned embeddings can close the performance gap relative to conventional grain and texture statistics on larger experimental campaigns for inferring alloy composition and extrusion-process parameters from optimized metallic microstructure and texture.

Background

The study compares conventional microstructure and texture statistics with a frozen vision embedding and a frozen grain-graph embedding for recovering alloy composition and extrusion parameters from measured magnesium-alloy microstructures and textures. Conventional statistics perform substantially better than the learned embeddings on the 107-condition dataset.

The authors explicitly leave unresolved whether increasing campaign size and using larger or retrained embeddings would improve the learned representations enough to close this performance gap. This question concerns the scalability and transferability of learned representations beyond the small, sparse experimental campaign studied in the paper.