Metal cluster possible?

[marcelmbn]

What's the best way of generating larger solid state structures?

• test generation of metal clusters (just try & error?)

Possible solutions

1. possibly attachment of small molecules
2. generate large structure and "pull" it together
3. systematic generation of a cube-like structure (with systematic distances/angles/...)

Further thoughts

• increase D4 contribution in xTB methods to enforce more compact structures

[jonathan-schoeps]

Update on possible solution number 2

code: The code works for now but it needs minor improvements to the algorithm. https://github.com/jonathan-schoeps/MindlessGen/tree/dev/contract_coordinates
program: It is currently possible to create molecules with up to 35 completely random atoms. For structures such as "ternary oxide", molecules with 80 atoms have been found.
Method: The gfn-1 xTB level works well with the structures while gfn-2 xTB does not really converge on the structures found.

[marcelmbn]

Update on possible solution number 2

code: The code works for now but it needs minor improvements to the algorithm. program: It is currently possible to create molecules with up to 35 completely random atoms. For structures such as "ternary oxide", molecules with 80 atoms have been found. Method: The gfn-1 xTB level works well with the structures while gfn-1 xTB does not really converge on the structures found.

Please label with which code (e.g., https://github.com/jonathan-schoeps/MindlessGen/tree/dev/contract_coordinates) this was conducted or is tested currently.

[jonathan-schoeps]

Idea for ternary structures

---

• Use oxides (ground is mainly oxide).
• Do not use Thallium (not really interesting) use more like indium (semi-conductors) or similar important.
• Gfn-1 on its own not really interesting more important with post process on.
• Further advantages with insights into the structure with small molecules attached to the sides of the core structure.

[marcelmbn]

Perovskite structures as potential targets

• Perovskites have a lot of applications (see, e.g., here: https://www.nature.com/articles/s43246-022-00325-4)
• Could also be modelled as amorphous structures (e.g., https://www.nature.com/articles/s41467-020-16561-6)

[jonathan-schoeps]

Update on possible solution number 2

Code

The newest version of the code (https://github.com/jonathan-schoeps/MindlessGen/blob/dev/contract_coordinates/src/mindlessgen/molecules/generate_molecule.py) works fine.

Proof of concept

Here are some examples for the calculated structures while using the new version of the code:

Structure with 43 atoms:

Optimized with GFN1-xTB:

mlm_C1H18O2Li1F3P2Zn1Ga4Se2Br2Kr1Mo1Ag1Te2I1Cs1_4cd696.zip

A first try to converge the geometry of this structure with Orca, but it did not converge after 25 optimization cycles:

orca_43_atoms.zip

Structure with 48 atoms:

Optimized with GFN1-xTB:

mlm_H25N3O2Be1B1F1Al2S2Cl1Ar1Ca1Br2Sb2Cs1Hf1W1Ir1_2e138c.zip

Ternary Oxides:

The tested structure contains 80 atoms with the sumformula NaInO2:

Optimized with GFN1-xTB:

mlm_O40Na20In20_42de87.zip

[marcelmbn]

All in all, great results! 🚀

In the following, some detailed feedback on the specific points:

Structure with 43 atoms:

Optimized with GFN1-xTB:

mlm_C1H18O2Li1F3P2Zn1Ga4Se2Br2Kr1Mo1Ag1Te2I1Cs1_4cd696.zip

A first try to converge this structure with Orca, but did not converge after 25 optimization cycles:

Regarding the wording in this context: Usually, people think about SCF convergence when talking about trying to converge this with ORCA, so make it clearer if you mean SCF or geometry convergence.

orca_43_atoms.zip

Structure with 48 atoms:

Optimized with GFN1-xTB:

mlm_H25N3O2Be1B1F1Al2S2Cl1Ar1Ca1Br2Sb2Cs1Hf1W1Ir1_2e138c.zip

Regarding both, the 43 and 48 atoms structure: We should make use of the new feature in mindlessgen.toml:

# > Scaling factor for the employed van der Waals radii. Options: <float>
scale_vdw_radii = 1.3333

By using maybe 1.1 or 1.2, we're reducing these "semi-fragments" that seem to be present in the structures (some atoms or structural motifs might not be connected via covalent bonds; IMHO the distances are sometimes a bit too large). Maybe we should also reduce the default from 1.3333 to 1.2.

Ternary Oxides:

The tested structure contains 80 atoms with the sumformula NaInO2:

Optimized with GFN1-xTB:

mlm_O40Na20In20_42de87.zip

The ternary oxide looks great! Have you tried converging this with PBE (SCF convergence)? Have you already tested generating even larger structures (maybe 120 atoms)?