Re: CCL:New linear scaling method

Dear CCLers,
 John McKelvey has asked me a question regarding the local minimum
 problem. I believe the problem is very important to be discussed on the
 list. Therefore I post my answer to the list.
 ----- Original Message -----
 From: <jmmckel -AatT- attglobal.net>
 > Victor,
 >
 > How would the problem of local minima be handled if the program were to do
 a
 > geometry optimizatopn on 10**5 atoms?
 >
 > Regards!
 >
 > John McKelvey
 Dear John,
 Perhaps you agree, that the local minimum problem is a separate one to
 the linear scaling solution of the diagonalization of Fock matrix. LocalSCF
 method itself resolves just the diagonalization problem. The same do other
 respected linear scaling methods, e.g. MOZYME, D&C, CG-DMS.
 However I agree that the local minimum problem has to be addressed if one
 is serious about protein modeling. This requires molecular dynamics
 averaging
 applied after some preliminary structure refinement done by geometry
 optimization.
 Taking the present program code one can do full geometry optimization
 for 100,000+ real protein on a single-CPU PC. I bet QM MD for about
 1,000 atoms protein is a feasible job for a 10 CPUs Linux cluster nowadays.
 Although this has to be implemented first.
 Full QM MD modeling of proteins is not a long future. The present LocalSCF
 code
 works just on 1 CPU. Being parallelized it could utilize about 1000 CPUs
 quite
 effectively. This unusual level of parallelization is not just my
 imagination.
 This capability comes from extremely  high localization level of the
 LocalSCF
 method. As the LocalSCF preprint shows, LMOs expanded just on 30 atomic
 centers are enough to get 0.001 RMS error on atomic charges. This level of
 localization is extremely favorable for parallelization.
 We already made some tests on 1 CPU machine applying D&C technique on
 the top of LALM, which is a similar to the LocalSCF engine. 1,000,000 atoms
 irregular linear peptide built from random combination of 14 different amino
 acids
 was divided on 300 fragments. Calculating each fragment sequentially on the
 one CPU gave the energy converged in 7 hours. Of course, the convergence
 criterion was quite loose but this was enough to show a feasibility of the
 D&C / LocalSCF concept.
 I believe the publishing of the LocalSCF paper makes a valuable contribution
 for
 the wide dissemination of the idea. As I mentioned in my previous posting
 the
 preprint is available from the Chemistry Preprint Server by URL
 http://preprint.chemweb.com/physchem/0304005
 The server requires registration, but the registration is free.
 Hope, I answered your question.
 With kind regards,
 Victor
 --
 Victor Anisimov
 FQS Poland