I have a few questions regarding AdsorbML.
First, when using this method on a new surfaceadsorbate combination, could you please clarify how many total ML relaxations are conducted before being downselected to the best k? And do you think this choice could influence the optimal k?
I’m also a little confused as to the meaning of the Heuristic+Random ratios. Does this percentage refer to the number of heuristic configurations that have been randomly perturbed? For example, when it mentions “heuristic + 80% random DFT” in the paper, does this mean that 80% of the heuristic configurations have been randomly perturbed? I hope I’m not missing something obvious here.
Thanks!
Hi Lance, apologies for the delay on this.

We enumerated all heuristic configurations (on average ~56) and M=100 random configurations. We run ML relaxations on these then select the best k. If you check out the SI in the paper you’ll see an experiment where we explore varying % of random configurations to include and its impact on the results.

Not quite. This paper generates initial configurations via heuristic and random configurations. For random configurations we generate 100 configurations. So when we say “heuristic + 80% random DFT” we are referring to using all the heuristic configurations AND 80% of the random configurations we generated, so in this case 80 configurations. Hope this clarifies it.
– Muhammed