From owner-chemistry@ccl.net Mon Oct  1 10:23:01 2018
From: "Yury Minenkov yury.minenkov-.-gmail.com" <owner-chemistry_+_server.ccl.net>
To: CCL
Subject: CCL:G: Interaction of QC software components via text files
Message-Id: <-53482-181001045314-21837-ZcCfd+e19qbI/+1hoCwKaA_+_server.ccl.net>
X-Original-From: Yury Minenkov <yury.minenkov..gmail.com>
Content-Transfer-Encoding: 8bit
Content-Type: text/plain; charset="UTF-8"
Date: Mon, 1 Oct 2018 11:52:55 +0300
MIME-Version: 1.0


Sent to CCL by: Yury Minenkov [yury.minenkov:gmail.com]
Dear Grigoriy,

yes, in my opinion, what you are saying is perfectly doable, e.g. in
Python or any other programming language you are comfortable with. In
the simplest approach you need to specify your Cartesian coordinates
as variables to optimize, then you need to write a function which
based on the passed coordinates generates an input file for your
favorite quantum chemistry code to run the energy/gradient
calculation. After that you need to feed both the variables (Cartesian
coordinates) and functions (energy run, gradient run) to an optimizer
of your choice: the simplest would be to use the algorithms available
through https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.minimize.html
. Then - it is done. You can even further calculate the energy second
derivatives e.g. via finite difference (energy (slow) or gradient
based - better) method and calculate the harmonic frequencies. But
please, be aware that any serious quantum chemistry code is written by
professionals and it will be very challenging to "beat" built in
optimizers. Sometimes, slow optimization is rather a consequence of
chemical system at hand (too shallow PES minimum, noisy
energy/gradients, etc). And, yes, FPD or CCSD(T)/CBS
optmization/frequencies can be implemented in this way, but my gut
feeling is that apart from some small systems, these calculations are
going to be quite time-consuming.

Could you be a bit more specific about DFT functionals and neural networks?

With kind regards,
Yury


On Wed, Sep 26, 2018 at 11:24 PM Grigoriy Zhurko
reg_zhurko/achemcraftprog.com <owner-chemistry[a]ccl.net> wrote:
>
>
> Sent to CCL by: Grigoriy Zhurko [reg_zhurko-x-chemcraftprog.com]
>   I suppose it is well-known that the geometry optimization in Gaussian is sometimes accompanied with serious problems, and the users have to invent their own ways of solving them. Besides that, the energy of the final (optimized) geometry computed with Gaussian is often slightly higher than the energy of the penultimate (previous) step. This is usually not a big problem, but this is slightly annoying (unless you use my program Chemcraft) ).
>   My question is, whether it is possible to implement a third-party algorithm of geometry optimization, which implies invoking the Gaussian by a third-party program. I mean that the third-party program generates a Gaussian input file (.gjf) with single point and gradient computation, then Gaussian processes this job and generates the output file, then the program reads and parses this output file, then it predicts the coordinates of the next step and runs Gaussian again, etc. until the energy minimum is reached.
>   If such an approach is possible, I will probably try to implement the following things:
>  - geometry optimization and frequencies computation with CCSD(T)/CBS, or even a composite method like FPD;
>  - performing several computations with different DFT functionals at same geometry, and passing their results into a neural network to obtain a very accurate energy.
>   So, is this possible to implement such interaction of a third-party program with Gaussian (preferably the Windows version), or other QC codes?
> Grigoriy Zhuko>
>