ACE1 Software Paper

[cortner]

This is a META-issue collecting what are the key priorities to complete for the publication of the software paper. I will start filling in details here and ask everybody I CC to closely follow discussions here. If there are specific issues on your mind that you'd like included here, it may be better to discuss on Slack and post below only when a discussion has settled and there is a clear TODO item. I will keep editing this post with updates.

0. Paper writing

1. Documentation

2. Names of user-facing functions and parameters

3. Tutorials

4. Default Heuristics

5. Code Code and Performance

CC @tjjarvinen @WillBaldwin0 @gelzinyte @wcwitt @casv2

[cortner]

0. Paper

• Write appendix that sketches out fast evaluation with reference to new kernel. (CO)
• Fix how we present the code / pseudocode
• JSON interface

1. Documentation

• rewrite "intro to ACE models" once the manuscript is done and base it on that. -> moved to another issue
• need a page in the docs on how to prepare data to work with ACE1pack Documentation on Preparing Data #95 -> moved to another issue
• The ASE and LAMMMPs tutorials should be extended and self-contained.
• liberally add links into the documentation, e.g. every time we mention ACE1 or ACE1pack etc.

[cortner]

2. Names of user-facing functions and parameters

• should the default key for forces be "forces" instead of "force"? @wcwitt. WCW edit: @gelzinyte and I made this change.
• The smoothness prior is at the moment impossible to construct for a user, we should introduce the function smoothness_prior, already done for the ACE1Model in WIP: New Affine Model Interface  #92 , but should also introduce this for the ACE basis and the combined basis; done in ACE1pack, TODO move it to ACE1? Elena edit: Also rename ACE1pack's regularizer.laplacian_regularizer to be consistent with ACE1Model.
• Closely related to the previous item: should the prior be passed to the solver or should it be passed to acefit!?
• Consider Elena's suggestions for clearer keyword argument, maxdeg2, rcut2; cf Verbose parameter names for acemodel #99
• ACEfit Bayesian solver names and default parameters; Clarify naming of Bayesian solvers ACEfit.jl#41
• P vs Gamma for the regulariser. Probably P for prior, but let's please never call it a preconditioner again
• species -> elements (requested by Gabor)
• N -> bodyorder or should it be order? (requested by Gabor)

[cortner]

3. Tutorials

• Fix the output of ACE1pack.linear_errors; cf. Redesign linear_errors #98
• what is the default solver and default solver parameters that we should use in the tutorials? @wcwitt
• need expanded examples (for dummies!) for ASE and LAMMPs usage, Working Examples for ASE and LAMMPS #94 ; those tutorials should be fully self-contained and need to be revised (broken links etc...)
• add a tutorial on running a simple MD simulation with Molly.jl

[cortner]

4. Default Heuristics

• add multi-degree and wL choices to the acebasis and acemodel interfaces - CO, done for ACE1.jl, todo for ACE1pack.jl cf WIP: Backporting simplified multi-degree interface #104
• When using BLR then the solver should automagically generate a committee and attach it to the potential. Or, alternatively, this can be done in acefit! instead of BLR? What's the right approach? cf BLR and Committees #106
• BLR should be the default solver and if no solver is specified it should work reasonably well with default parameters. To be tested on a few benchmarks?
• maxdeg_pair = maxdeg; i.e. hide the 2-body from the user until they dig deeper into the documentation

[cortner]

5. Core Code and Performance

• CI needs to be fixed
• consider moving to new kernels asap
• Make ACE1 potentials compatible with Molly.jl - cc @tjjarvinen
• Add unit test for acemodel (- EG edit)

Random stuff

• export smoothness_prior

[wcwitt]

Thanks for this.

[cortner]

I see you edited your message. I guess it's ok to discuss here as well, but let's make sure all ACTUAL action points remain in the posts above.