Supplementary Data for "Self-consistent biased fine-tuning for highly accurate reaction-specific machine-learning interatomic potentials"

This repository contains all relevant data for the manuscript "Self-consistent biased fine-tuning for highly accurate reactio
n-specific machine-learning interatomic potentials".

The fine-tuned MACE MLIPs used for rate constant calculations are stored in the folder: MACE_fine_tuned
There, the MLIPs optimized from structures directly sampled by the analytical PESs are called "analytic", while those optimized by self-consistent biased fine-tuning are called "self".

The folder "input_analytic" contains input files for MACE fine-tunings from direct samplings with the analytical PES. Here the malonaldehyde cheap-h example has been taken.
The training set is generated in the folders "sampling_PES" (along reaction path) and sampling recross (recrossing trajectori
es), here at 150 K, respectively.
The fine-tuning is done in the folder "tuning" (with a training set for all temperatures).
A rate calculation at 150 K is done in the folder "rate_calculation".
The energy deviations are calculated in the folder "benchmark" (here only for 150 K).

The folder "input_self_consistent" contains input files for the self-consistent biased fine-tuning.
Here the CH4+OH normal-e example has been taken (cycle 3).
The training set is generated in the folder "sampling_PES", here at 150 K (no recrossing training needed).
The fine-tuning is done in the folder "tuning" (with the full training set).
A rate calculation at 500 K is done in the folder "rate_calculation".
The energy deviations are calculated in the folder "benchmark" (here only for 300 K).