Establish equal-total-cost acceleration in NEB workflows

Determine whether MARC-Field initialization accelerates nudged elastic band calculations when the total electronic-structure cost and optimization schedule are equalized against IDPP initialization.

Background

The NEB benchmark compared MARC-Field and IDPP workflows using the same number of post-construction optimizer steps, but the workflows did not have equal total electronic-structure costs or identical phase schedules. MARC-Field used an additional 11-point DFT scan to center the energy peak, whereas IDPP did not use an analogous scan.

Consequently, the reported improvements establish utility within the predefined workflows but do not resolve whether the learned initializer provides acceleration under a strictly matched computational budget. The paper explicitly identifies equal-total-cost acceleration as untested.

References

Because that workflow also used an 11-point DFT scan to centre the energy peak and a different optimization schedule, the result supports initialization utility; equal-total-cost acceleration remains untested.

— Learning continuous reaction paths for transition-state prediction  (2609.25523 - Yang et al., 22 Sep 2026) in Section 3, Discussion, paragraph beginning “MARC-TS recasts transition-state prediction”