Nowadays, I am trying to compute the bulk properties of the MgB2. In literature, the bulk parameters a=b=3.06 A, c= 3.52 A. However, the resulting bulk parameters are a=b= 2.88 A (approximately 6 error, acceptable), c= 2.99 A (approx. 15 error). It is quite apparent that the force field that I use is pGFNFF has a poor performance to predict c parameter including Mg-B interactions.
I played with many keywords but nothing has changed.
What can I do to improve ionic interactions between Mg-B?
Hi Sefika
pGFNFF is a universal force field and so it won’t be perfect for all materials, especially if they weren’t included in the training set. If a force field isn’t as accurate as you like then the solution is always that you need to fit your own force field by thinking about the nature of the interactions, choosing a suitable functional form and then fitting against the relevant data.
Regards,
Julian
Generally the result of almost all molecular mechanic FF (including e.g. gaff2 optimized only for organic compounds) are not able to reproduce the reality. I had spend almost 2 years to make it working in GULP automatically (including engine for calculating RESP charges) - and it was giving results not enough good for use real problems.
I suggest you to use machine learning (ML) force field. I have excellent experiences with MACE OFF 23 Large force field inside ASE environment for organic materials. It gives results comparable to DFT + runs in graphic card super fast. It have issues with ionic organic compounds but work on this issue is in progress …
For inorganic compounds specialized MACE parametrization exist as well:
I personally believe the simple MM force field are a dead technology with some sens only for very large systems (proteins) and they will be replaced by ML FF totally soon.
I would give a slightly different take on this. For small bulk unit cells there is indeed little point in running a force field calculation since you can just run a QM calculation very easily (no need for machine learning at all). For larger and more complex systems with dynamics then the situation is different. Machine learning can be valuable, though with the caveat that the results inherit the problems of the QM data used to learn & so if you are using something trained on PBE, for example, then this will have it’s own issues. For example, GGA calculations for multiply charged anions in water can give much worse results than a well-fitted force field. As CCSD(T) quality machine learning models become more common (see for example this preprint) then this won’t be an issue. For systems with strong long-range interactions then it may require an explicit treatment of this in the machine learning due to the short-range cutoffs used. Bottom line is that are still many systems where a force field can be useful, sufficiently accurate and much faster; of course it requires more human intelligence instead of the artificial variety to derive them. If we write off physics-based models and just let machines do the learning then computational chemists will largely be out of a job.
I speak abou the category of the problems I work on: Verification of experimental crystal structure determination. It can be up to now done by meta-GGA DFT (r2SCAN) or for non ionic pure organic molecules by MACE OFF 23 large. Other less computational expensive methods are not able to reproduce expeimental data with enough precision to detect issues with experiment …
Can this MACE be parametrized for the external softwares? such as uff or drieding. Or it should be reparametrized for each structure based on spesific crystals?
MACE is a forcefield itself. UFF an Dreading are another forcefield = not software. MACE is an universal forcefield working with any structure (supported by specific parametrization). So I e.g. run now optimization of 280 000 diffetent experimental crystal structures compatible with the OFF 24 Medium parametrization. And there is no need to parametrize it specificaly for each one, becouse it wors with any non-onic pure organic molecules containing 10 selected elements. There exist other parametrizations for 89 elements … And yes - MACE exist as a plugin to third party software like ASE, Lammps and more …
One last question. For spesific molecules, we generally use different forcefields such as Tip-4,5, Trappe. Especially, for surface-molecule interactions.
Does MACE work for surface-molecule interactions? Do you have a personal experinence?
MACE is a force field, so it can be used for anything = crystals, isolated molecules, surfaces. Check the ASE general tutorials for surface task and select MACE as calculator. I had recently switched from MACE to Fairchem engine + UMA model. It was created by the Meta/Facebook AI people. It works in ASE as well and my test shows better agreement with experiment (atom positions from X-ray).
I have experiences only with crystals an isolated molecules in vacuum. No with surfaces.
