[mKarpis]

Dear All,
I'm from IT environment so please move this post to the right place if I'm in the wrong now.

I would like to ask you for a help in following: I'm thinking about making as my PhD study a simplified parallel simulation on GPU of molecular mechanics. Basically from atomic bond, inorganic molecules, organic molecules, macromolecules.. I would like to use for that CA (Cellular Automata) model. Idea is to set-up initial rules and a configuration and then 'watch'. I know,. it might look very easy, but as I mentioned I would like to have it as simple as possible. What I miss is some overall information - to find out what conditions/process I need to make. I was in touch little bit with GROMACS, but I would like to design application from scratch to be 100% on GPU + have OpenGL viewer on top of it.

If we simplified it - are we talking about Newtonian laws?

So the questions are:

• do you know about some document/paper (or research group) which would describe 'simplified' model of molecular dynamics?
• do you know about some book which will cover this? (simple...for IT guy, who didn't have chemistry for the last 15 years)

Thank you in advance.

[Enthalpy]

Excellent idea! And welcome here!

The right place to run molecules computations is the graphics processor, sure. If it hasn't been done before, it would be about time.

Not just for dynamics! I wish I could compute single static molecules as well on the graphics processor. Not only for ab initio quantum mechanical computations of orbitals and conformations (for which the graphics processor would fit naturally): even the simplified models (Am1, Pm3...) are uncomfortable on a Core 2 with as few as 15 carbons, and processor speed per task has made zero progress since the Core 2.

In contrast, graphics processors compute faster than a Core, and keep improving.

One difficulty: graphics processors run best on crunch-without-questioning programmes, but molecule conformation does require questioning during the optimisation: has the bond become unlikely long? Is the optimum behind this small energy hill?

As well, the algorithm is way more important for speed than the hardware - as you know. Writing a huge equation with the XYZ of each atom and optimizing them for energy certainly leads to a very slow programme. It would be better to look after molecule parts that depend little on the rest, like a phenyl group having just one link with the molecule, optimize it once and update it only if something new (a collision) happens. For that sake, optimizing bond lengths and angles would be better than atoms positions. As well, a less stupid programme could search if some molecule parts are already well-known: optimizing an alkyl tail or a cyclohexyl again at each molecule makes no sense; even the drawing interface could do it by rules better than now. Then, only multiple rings, steric hindrance, intramolecular interactions, and collisions would make the programme complexity more than linear.

Here again, letting this kind of subtlety run on a graphics processor is a challenge, for which limiting oneself right from the beginning to one programming model like a cellular automata might let miss important gain possibilities.

Well, if this is too simple: I wish there were a programme that computes freezing points... None seems to exist, because it's too difficult. It requires to check all possible ways to stack a bunch of molecules, with many positions and orientations, where some molecules can change their shape to freeze better. Probably inaccessible, even with stiff molecules.

Wish you to succeed, as better software is needed here!

[Enthalpy]

Please take the following with caution, as a good part of it may be wrong...

Academics tend to use Linux and want to insert programme results, including pictures, into Latex papers.
I egoistically wish your programme runs on W2k, and then you have drivers for dX10 graphics cards but only MS' dX9.0c software runs, DotNet 2.0 maximum, OpenGL drivers yes, but PhysX apparently not, and Cuda is unclear.

Molecular software can and does often run on a network of many PC, especially at universities, and using the graphics cards of a network of PC's would be nice as well. You might as well consider that recent supercomputers use graphics processors, again with a network. Meanwhile, Seti@home runs on the graphics processor, and Folding@home maybe as well. Folding@home resembles what you plan, and may inspire you.

Molecular software uses to be split among computation part and drawing+display part, with software of different origins being widely compatible. Following this scheme should improve the acceptance of your contribution.

Discovering the subject:
http://www.fz-juelich.de/nic-series/volume23/allen.pdf
http://en.wikipedia.org/wiki/Molecular_dynamics
http://en.wikipedia.org/wiki/Molecular_modeling_on_GPU
http://www.multiscalelab.org/acemd (MD on GPU)
http://www.biomolecular-modeling.com/Products.html (MD on GPU)
http://codeblue.umich.edu/hoomd-blue/ (MD on GPU)

Mamma mia! From the Pdf from Jülich, it seems that MD is still made with spherical atoms and a force that depends only on the distance, and only on atom pairs... We're still at prehistory here.

Well, this idea was so good that it is already implemented many times.
...switch to the freezing point software I desire, maybe...?
http://www.chemicalforums.com/index.php?topic=46171.0

More knowledgeable people are welcome to correct my frenzy here above!

[mKarpis]

millions of thanks for all your comments!

[Enthalpy]

A similar undertaking would be to simulate galaxies and their collisions.

This is to be checked, but I believe many astronomers are still interested in simulating only the gravitational attraction between point stars and point or diffuse dark matter. In this, galaxies are far simpler than molecules.

As well, gravitation forces decrease slower over distance than forces between atoms do, so the algorithm has to be less subtle: compute over many stars, instead of smartly determining which atoms really interact. This would fit the capabilities of a graphics processor better.

Existing algorithms supposedly replace groups of distant stars by one equivalent mass (and possibly higher moments: dipoles, quadrupoles...) to limit the amount of computation, but such methods are still easy to crunch-without-meditating.

Dark matter is absolutely necessary in any sensible model of a galaxy. It is not if you model only globular clusters, but these are less varied and spectacular.

Should you be interested in such a project, you should check how existing algorithms proceed. Many astronomers program well, so running such a simulation on a graphics card must already exist.
Edit: here a rudimentary example http://sakshiagrawal.com/GPU_Proj_2.htm
Just ask people in this job what your programme must do better to be an improvement.