I’m a self-taught student of Molecular Modeling who is just trying to learn by doing. I have a math and coding background and know the fundamentals on quantum-mechanics , but have limited prior experience with chemistry.
As a self-assigned exercise, I’m trying to build a Python framework to accurately model interatomic distances for many (all?) of the diatomic molecules.
From past failed experiments, i can now see that wise Force Field choice is a fundamental aspect of this kind of effort.
Does anyone have any advice for me on how to think about which Force Fields I should choose for this? From my naive perspective, it feels plausible that for the ~150 dimers; it’d be prudent to at least choose three different force-fields for the cases of: (i) non-polar covalent, (ii) polar-covalent, and (iii) ionic bond.
Amongst the 100s of models out there and dozens of packages, this is a tough matter to see clearly.
I am having some initial successful results with MMFF94 and UFF with non-polar covalently bonded dimers, but am still trying to figure out the other cases.
Curious to know if anyone has advice for me here. If anyone has survey material (papers, blogs) on this that’d provide a guide; I would be very grateful.
LAMMPS contains a wide range of different interatomic potentials, because practitioners of MD have many different motivations. Sometimes they are looking for a model that is a simple as possible, while still capturing some basic physical or chemical behavior. Other times, they need a model that describes the atomic motion of a specific chemical system as accurately as possible, perhaps by comparing to quantum electronic structure calculations. If you just want to learn about molecular dynamics, the best place to start is the Lennard-Jones (LJ) potential. It was originally derived as a classical approximation to the self-interactions of noble gasses, but it also qualitatively captures many of the properties of typical solids and liquids too. The properties of Lennard-Jonesium has been described in many papers and the LJ dimer has also been studied extensively.
You have given yourself a very tough assignment and classical models are on average not going to be very good for that. You may end up having to use a different parameterization for each that was parameterized from quantum calculations. Even with quantum calculations, getting accurate results without extra information that is specific to the individual pair of atoms like the spin state (n.b. ground state of oxygen is a triplett state!) can be difficult.
Please also note that a key property for classical models is that they are computationally much cheaper (and have better scaling with size) for larger systems. So looking at dimers is not taking advantage of the biggest benefit of classical models.
@akohlmey – quite helpful. thank you. Your remarks are helping refine my intuition on how to choose models.
I’m now making my way through a UFF-based force-field modeling of inter-molecular distance for all of the dimers and finding some areas where there is decent accuracy and others where there are serious errors. Prediction errors are common here even with the “prior” of a SMILES specification of the dimer which I guess is an extra boost to accuracy given that it fills in some details on the ansatz of the simulation.
If instead wanted to pivot this approach (since now I have the framework in code) towards using ab-initio methods, do you have a suggestion on which class of quantum methods (and which class of basis sets) would be most pertinent to a “re-run” of these efforts?
It depends on the level of accuracy you want. Some semi-empirical method might do or some hybrid DFT method or you may need to do coupled-cluster calculations. You would need to check the literature. There should be plenty of studies of diatomics.