Help with interfacing xtb with LBFGS

[retinavr]

Hi

I am currently working on a project where I want to implement molecular optimization using xTB into a VR platform. This has to be done in C++, but i am very very new at C++. I can see you have used xTB with LBFGS optimizer. I was looking into doing the same. Can you help me writing a simple code that only takes xyz coordinates (array or 2d array) and amount of steps (int) for optimization as input and outputs the optimized coordinates (array or 2d array) and final energy (double)?

Kind regards

[conradhuebler]

Hey retinavr,

some things to start. Since everything is implemented in "src/capabilities/optimiser/LBFGSppInterface.h" for the LBFGS part and in "src/core/xtbinterface.cpp" for the xTB part, it should be a good start. However, you have to interface quite a bit, to control all parts of xTB (spin, charge, accuracy) and neither LBFGS nor another BFGS-based optimiser I tried, are as powerfull as I wanted them to be.

So, let me be interessed in VR - are you talking about virtual reality? If yes, why not let curcuma work for now as interface between them ( I know, you want to implement some stuff on your own, but is it worth for two developers to keep track of an API that might change) and try to interface curcuma ( I will support that ;-) ) and then improve the curcuma codebase on the other hand? - If you were interested, I would be happy and provide you with some appropriate starting points.

Cheers and best regards
Conrad

[retinavr]

Hi Conrad

So, for my purpose, i am not interested in spin and charge for now. The purpose is to provide a fast optimization scheme (5 step optimization within 0.015 seconds for around 200 atoms) for approximate structure optimization.
The platform I am building is indeed a Virtual Reality platform. So far I have successfully created a molecular builder and implemented Openbabel for relaxation. I am interested in xTB because it is more quantum chemically correct compared to the forcefields in Openbabel. I would definitely not mind interfacing with curcuma. For me it is important to create a function in c++ which takes structures as input and atoms to freeze/constrain and output relaxed structures as well as energy. I would really appreciate any help and will of course document your contribution to the project. Another purpose of mine is also to learn c++ while creating this interface.

FYI. I am developing on windows.

Kind regards

[conradhuebler]

That sounds pretty cool, but there are some things, that have to be dealt with in the first place:

• I don't know if xtb can be compiled on windows platform.
• On linux systems, using gcc and gfortran, the binaries are much slower than using icc and ifort.
• I did not get LBFGS to do constrained optimisation together with xtb. It seems a bit more complicated, and I am not trained in mathematics ;-).
• Furthermore, LBFGS is not very nice to be used if there are custom convergency criteria, like the change in strucure is smaller than an RMSD of a value, or energy convergency.

Actually, the best thing were to have an own BFGS like optimisation, however I currently have no time to do such a big stuff. A better way would even to adopt the ANCopt scheme from xtb, but it is not published yet, so the only way is to work through the source code.

So, I will today or tomorrow try to implement a "library"-version of curcuma, which then can be included in any C project and a sample main.cpp file, that just uses it. I think, that is a good start for you to have first look into the problem. And then, let see how it goes from there.

Cheers,
Conrad

[retinavr]

I got my hands on a precompiled build of xTB that works on windows. So far it has been working fine also when using the C API.
If constraints dont work, its not a big issue and nor is there any need for custom convergence criteria as the relaxation only needs to take a few steps per optimization.
That sounds really awesome if you could do that. Your help is much appreciated!!
Shall we continue over mail? Or is it okay to do it here?

Kind regards

[conradhuebler]

I would say, the https://github.com/conradhuebler/curcuma/discussions section is the best part. I will close the Issue with this comment and you can start a topic regarding the problem, and we will continue there.
Cheers,
Conrad

PS: Moved it to Discussion, was easy.

[retinavr]

My bad. Just saw you had it moved to discussion. Feel free to remove any of the threads :)

[conradhuebler]

Removed the other one.

[conradhuebler]

So, I have just added some "documentation" into a new branch:

https://github.com/conradhuebler/curcuma/tree/interface

src/helpers/interface.cpp is the file to start with. Have a look and feel free to ask.

[retinavr]

Thank you Conrad. You are a champ. I will have a look and probably return with lots of questions :)

[conradhuebler]

If you try to compile, easiest would be linux with gcc and gfortran, you can run just optinterface (name of the binary) and see, what it does. Btw, if you would try to make curcuma with xtb be easily compileable on windows platform, I would be happy. :-)