i’m new to lammps, and have enjoyed learning how to run the example input files. My hobby research is focusing on Graphene HBN heterostructures for twistronics, and how a perpendicularly orientated piece of quartz could be used as an actuator using dc voltage to expand, hopefully shifting the top graphene layer to different resistive states.
Without AI, mainly grok, I don’t think i would have gotten the build setup correct enough to run the interlayer package where graphene hbn examples files are. I have been able to visualize .data files in ovito, and am to the point where a command like OMP_NUM_THREADS=31 lmp -in in.grhBN will run.
Im wondering about how users pair multiple potential files to get a desired outcome? Do researchers end up writing their own potentials from scratch? I have encountered pair_style but have read in the documentation that its often not a sufficient method for concatenating files that are quite different. What potentials would you suggest using to model the piezo properties of quartz? One idea i have is to learn more about how to use quantum esspresso in tandem with lammps by using QM/MM with the fix command. My guess is that LAMMPS could be used to generate an outline of this structure which could then get sent to software thats focused on electrical properties of material. I am looking for help understanding the limits of what lammps can do, along with general discussion of how to use the tools available for modelling this.
It’s very much a post that seems like asking for help on how to do my own research, so apologies if these are the wrong type of questions for this forum.
Yes, this should go into Science Talk, but more importantly, your chances to get assistance on such matters in an online forum are very slim. You need to find a competent collaborator. AI can help you get an input deck that runs to completion, but that can be far away from making sense scientifically. The quality of the advice from AI using LLMs strongly depends on what information is available online. There are tutorials and examples and the LAMMPS documentation, so you can get something that has the correct syntax, but already having inconsistent units is not uncommon. This is where practical experience trumps any AI model since even publications only describe specific successful(!) applications and the experience of what does not lead to success is not considered for publication (how can you prove a negative?). Finding a method to have meaningful AI assistance for LAMMPS input is a topic of ongoing research and - to the best of my personal knowledge - no satisfactory generic solutions exist currently.
This is a tricky question but something that has a long history in the field. At the intermediate/classical scales where LAMMPS is typically used, two main procedures exists: bottom-up parametrization where you fit classical potentials from quantum scale simulations and top-down where you attempt to reproduce physical quantities (phonons density of states, mechanical constants, density, enthalpies of mixing etc.) For the former there are several methods implemented in softwares such as potfit or FitSNAP.[1] For the latter, people usually tried to derive equations of state or go with trial and error from reasonable guesses. Mixing models is also a tricky business that should be done very carefully since models are often parametrized independently. Cross interactions should only occur at interfaces (either QM/MM or boundaries between phases) and have a little or negligible effect on your results. Also it is often not about concatenating files but loading different models for different pair interaction and there are dedicated pair_style commands for this to be done.
In any case I can only go along with @akohlmey advice to look for specialists in MD of your given problem to help you build a reliable model if possible or at least get a better understanding of what is going on than what you can get out of any LLM. Much of scientific knowledge is found in the trials and errors other peoples have made but not published, and there are a lot of recipes and methods that you can only learn from other people guidance.
Note that taking electrostatics and charge transfers into account makes it a hell lot more difficult. ↩︎
Thanks for the advice and your time! There are a few mentors i’ve worked with, however none who are focused on computational material science. I am trying to get somewhere with the help of others, not a chatbot. I’ve found ai useful with getting cmake figured out, but as far as writing code, I am extremely hesitant to use it.
There is a small business research route I may try next to get collaboration. Figuring out whats needed as far as empirical results verses straight theory is a hard question for this problem.
Not being in school currently is partially why I described this as a hobby. Seems like there are plenty of alternative routes to try like offering lammps as a service to those who need more ram than a desktop, and are practicing, that would like persistent memory. Though attempting such moves the goal post away from the modelling.
Thanks for the detailed response! I have looked into both potfit and fitsnap, fun projects to learn! Part of the challenge im running into finding a platform that focus’ on the application of the the hubbard model, and lammps doesnt seem to have that. It becomes more of a quantum espresso question where dft+u and a few other packages exist for it. None that ive encountered for graphene hbn heterostructure, so the same need to tweak an input applies. thanks for the response, a lot to think about.