Things to Work on for Evolutionary Search App?

[SWeav02]

Hey @jacksund. @scott-materials and I have been discussing what I should work on next. One idea is to continue your work with the evolutionary search app and work out any kinks that remain from when you last worked on it. Afterwards I might explore additional ways we can speed up the process through ML or other ideas, but we both agreed picking up where you left off is a good place for me to start.

Are there any notable features or issues you were planning to work on but didn't get to? Similar to the custom workflow problem in #693? I'll also be looking through the code and messing around with it to get more familiar.

[jacksund]

Nice!

For learning, I wouldn't change anything in the alg and just add some breakpoint()s throughout the code. You can then "step" through the code line-by-line using these options in spyder:

In terms of improvements to focus on, these are (in order) what I would consider the most pressing:

1. switching the default workflow from relaxation.vasp.staged to a relaxation.quantum-espresso.staged one
   • This will make things MUCH easier on first-time users, especially those without linux or HPC experience. The QE + docker setup will be great for windows users just wanting to check out the workflows -- once they see the value in a quick local test, they'll be more willing to put the time in for a full linux+HPC setup
   • I do not have VASP set up on any of my main PCs, and I also think VASP is a pain to set up & work with in general -- so I would appreciate switching over. It'll be easier for me to jump in and help out / debug too
2. optimize similarity checks by doing them within the postgres database with pgvector or a similar extension
   • As this project grows, we'll want more and more searches + workers running at once, and fingerprints (as they are implemented right now) will be the bottleneck & potentially breaking for our database loads.
   • The current implementation involves storing fingerprints in a separate table and downloading the full set of fingerprints to each worker for it to use in validation checks. Similarity checks are cheap and easy -- downloading the full set is slow and very taxing on both the database and the workers
   • The ideal implementation is that the Structure datatable class includes a fingerprint column and a filter_similarity(structure, cutoff) method that calls pgvector directly. This is the setup I use for molecular systems and it scales extremely well
3. allowing custom workflows
   • currently relaxation.vasp.staged has some features that the evolutionary algorithm relies on. We need to identify these and make them as simple as possible, so that (ideally) the only requirement for a custom workflow is that you have a database table that inherits from a StagedWorkflow base class and optionally defines its own write_outputs method
4. allowing custom fitness functions
   • several places in the code assume energy_per_atom is the fitness score and what the search is trying to minimize. This should be adapted to accept some other database column such as electride_character or my_custom_scoring_fxn -- also whether this value needs to be maximized, minimized, or optimized_to_target_value.
   • This is probably lower priority for your work, but I have it here because of my own research
5. benchmarking transformations and adding adaptive steady-states
   • we need better analytics on whether a transformation is being effective -- for adjusting our defaults manually and/or having a mechanism to let the search adjust it automatically. For example, if 100% of the mirror mutations are giving back structures that are lower energy than their parents, then maybe we should increase its steady state relative to some mutation that is producing more higher energy structures
6. a monitoring dashboard
   • the output files are great and fortunately what the mat sci community is used to. But I see a lot of potential in a streamlit dashboard to watch searches progress in real time. I've been using streamlit to prototype interfaces more and more in my work and highly recommend it

I'll stop there for now, but that's a solid list of what I would focus on before jumping to new transformations, ML models, etc.

[SWeav02]

Awesome, thanks for the cohesive list! I'll start learning the code and then get started on these.