Integrating DFTK.jl into the active learning workflow

[emmanuellujan]

The purpose of the implementation of this script is to define a basic active learning workflow that integrates the tools we are developing. It needs to be gradually improved. In particular, the function get_dft_data should generate the data for the fitting process. The interatomic potential used in the script is ACE. It would be desirable to integrate DFTK.jl into the active learning workflow by implementing the aforementioned function. Any system that can be used with ACE and aligns with CESMIX objectives is suitable for this instance.

[emmanuellujan]

@mfherbst, when you can, could you take a look at this? I don't have any experience with DFT data generation, is there any code I can copy and paste to implement this function? Thanks :)

[mfherbst]

Sorry @emmanuellujan for being so slow to respond. Sure I'm happy to take a look. What exactly is the purpose of these examples: Something that is a reasonable setup (which in this case means you probably need a cluster to even run it) or something that can be run on a laptop.

Also using HfO2 as an example I'd say is extremely optimistic. I would be more comfortable to go to something like a sodium cluster or bulk system, a bit like we are trying to do with @dallasfoster. Perhaps it's a good idea to move this code over here as an example once we are done figuring out a good computational setup.

[emmanuellujan]

Sorry @emmanuellujan for being so slow to respond.

No worries :)

Sure I'm happy to take a look.

Thanks :)

What exactly is the purpose of these examples: Something that is a reasonable setup (which in this case means you probably need a cluster to even run it) or something that can be run on a laptop.

Short answer: laptop.

In general, because the tools we are developing for cesmix are disseminated on github, and because each tool is focused on a specific domain, it is convenient to have a repo that concentrates our efforts, providing usable and integrated case studies (some of them ready to use on your laptop, and others ready to use on the cluster). Having step-by-step documented and integrated examples will give us more chances to hook new users and also makes it easier to know the status of the project.
For the time being, if we can assemble a "version 0" (with all our Julia stack) that a user can run on a laptop and that is aligned with what we promised DOE I think it will be enough :)

Also using HfO2 as an example I'd say is extremely optimistic. I would be more comfortable to go to something like a sodium cluster or bulk system, a bit like we are trying to do with @dallasfoster. Perhaps it's a good idea to move this code over here as an example once we are done figuring out a good computational setup.

Cool

In case Na's example might take a while, in the meantime, I'd appreciate if you could give me some simple code that can be copied and pasted into that function (get_dft_data)

[emmanuellujan]

I almost forgot, the domain should be a cube, otherwise we cannot run the MD with our current tools.