It is advisable to first separate the parts that can be made rule-based, such as materials data cleaning, structure generation, error classification, and exception handling, from the parts that require ambiguous judgment. I recommend assigning the LLM only narrow roles such as candidate selection, exception classification, and parameter recommendation, and then checking those results again with a rule-based verifier.
As a small experiment, modify one existing pipeline and limit the points where the LLM intervenes to 1~2. Then compare success rate, reproducibility, manual intervention time, and ease of recovery after failure; this will reveal the value of this pattern.