feature - 'DFT smart grouping'.

[williamdavie]

Looking to add a feature to group similar configurations for DFT calculations. In doing so, the CHGCAR file from the most similar DFT calculation can be used as a starting point for the next loop, significantly improving convergence.

Suggested algorithm structure:

Consider 8 DFT calculations on a periodic solid which relaxes to BCC.

When DataCollection.write_QM_submission_scripts() is called with for example npartitions = 2.

Of the 8 configurations, we choose the 2 closest structures BCC (not altering the lattice parameter) as these are the most likely to converge. Leading to tracks of calculations:

---config1---
---config2---

All 6 remaining configurations are then assigned iteratively, finding the closest config (Using EMD for now, mlipts.similarity.emd) at the end of each chain. E.g.

---config1---
---config2---config3---

where of all the configurations, config3 is closer than anything else to config2, and closer to config2 than anything to config1. Since calculations are looped sequentially, CHGCAR can be copied from prior to the next.

Then we can include into main class with something like DataCollection.write_QM_submission_scripts(smart_convergence=True)

[williamdavie]

Turns out this is a form of greedy clustering https://en.wikipedia.org/wiki/Greedy_algorithm

[williamdavie]

Key changes included the addition of smart_group_calcs() in the new path mlipts.similarity.group

To receive the desired output a scoring system is added to score access both the EMD and a normalized difference in cell dimensions.

$S_{ij} = \alpha{}|\boldsymbol{a}_i - \boldsymbol{a}_j| + (1-\alpha{}) \text{EMD}(i,j)$

This is implemented as

configs[i].cell - configs[end_index].cell / cell_norm
score = 0.9 * cell_diff_score + 0.1 * emd

To ensure CHGCAR is copied and used correctly edits were made in mlipts.data_collection, in particular setting ICHARG appropriately and editing the submission script setup i.e. write_run_calculation_scripts()

Note

Many tests required to optimise both convergence in vasp (choosing a good value of $\alpha$) and leaving with opportunities optimise and clean up code, generated as sub issues.