Because materials literature search is far more exploratory than simple keyword matching, it would be possible to try a method that logs the sequence in which researchers or agents read papers, figures, captions, and appendices, and uses those trajectories to train retrieval. In topics with varied terminology such as alloy composition, defects, and process conditions, search that follows the research flow may fit better.
At a small scale, one could collect several dozen to several hundred search sessions in an internal literature Q&A system and compare them with conventional BM25/embedding search. If it is possible to track which queries actually helped subsequent experiment design or calculation setting selection, evaluation more useful than simple search accuracy becomes possible.