Which is the recommended way to optimize "big" systems?

Question

Let suppose the we have a "big" system. By "big" I mean a system that have a certain number of atoms that the software is not capable to treat with the default initial parameters.

Question: Is there a recommended way to proceed in order to get the geometry optimized (or the self-consistence field calculation completed) with production parameters?
(By production parameters I mean good for scientific conclusions)

Some possible examples of actions that came to my mind:

• increase the geometry steps number (straightforward but not always optimal);
• increase the number of self-consistence field steps (straightforward but not always optimal);
• decrease the accuracy and after getting and optimized geometry, increase the accuracy;
• start with a low level basis set, use the optimized geometry as input with a higher basis set;
• use a simple method (like molecular mechanics) to optimize the geometry, then move to the desired method;
• etc.

Answer 1

Suggestions for Molecular-Like Systems (not periodic unit cells)

I'll make some general suggestions for molecular / isolated systems, not crystals. (In the case of periodic systems, some methods aren't available or are different.)

You mentioned increasing the number of self-consistent steps. If your SCF calculation doesn't converge, there are a variety of separate questions / methods to try.

The main problem is that getting to a final equilibrium geometry with a DFT or high-level wavefunction method can be slow. For larger systems, there are many degrees of freedom and often the potential energy surface is fairly flat, especially with dihedral angles.

So the goal is to get close to a converged geometry with a faster method, then to switch eventually to your method of choice.

Years ago, the advice would be to use a force field to "pre-optimize" a system, since these are intended to get quality geometries. I no longer advise this unless you have a specialized force field (e.g., biomolecule force fields work well for biomolecules).

In particular in two benchmark papers, we found that force field methods weren't very good:

• "A sobering assessment of small-molecule force field methods for low energy conformer predictions" Int. J. Quantum Chem. 2018 118 e25512
• "Assessing conformer energies using electronic structure and machine learning methods" Int. J. Quantum Chem. 2021 121 e26381

We found that the MMFF94-optimized geometries had large forces when you switched to semiempirical or DFT methods. We also found that when ranking different conformer poses by energy, the force field had poor correlation with ab initio methods.

Instead, we find much better results with the approximate GFN2 / xtb method, or if possible with methods like ANI-2x, with the caveat that GFN2 works for all elements.

Our most recent workflow is something like this:

1. Generate an initial geometry (e.g. in Avogadro, RDKit, etc.)

2. Optimize the geometry using xtb which is fast and often "forgiving" (e.g., easier to converge SCF).

3. Optimize the geometry using B97-3c or similar GGA method (e.g. r$^2$SCAN-3c looks good too)

4. Optimize the geometry in your favorite method with dispersion correction (e.g., $\omega$B97X-D4 or $\omega$B97M-V for example, triple-zeta basis set).

If step 4 still takes a long time or fails to converge, we'll often optimize with def2-SVP first (often with looser convergence criteria) and then switch to the larger basis set.

It's a process - in part because getting the electrostatic and non-covalent interactions right is hard. Some ML-based methods may help with this in the future, but beyond ANI, they are not widely available.