Hsize?

[Thor]

Does the Hsize have anything to do with the time it takes to finish the wu?

I'm just curious because I have seen them wih a size of 245 and 42 for example... right one after another...

Why do they vary that much?


Greets Thor

[graeme]

The HSize has a little bit to do with how long the WU takes to finish. There are three steps to each calculation.

The first step is to find a saddle point, which corresponds to some process by which atoms rearrange on the surface. There is a significant amount of randomness is now long this takes, ranging from a few hundred to a thousand force calls. The force call is the basic unit of computation work within each WU. Each time the force is called, the client evaluates the energy of the atoms, and its gradient (the force).

Once a saddle point is found, the client determines if the process is directly connected to the minimum in which the search started from. To determine this, the client does two minimizations, along the two unstable modes at the saddle. These take about 100 force calls each. If one of these minima corresponds to the initial minima, the saddle corresponds to a valid process available to the system.

Next, the rate of the process is evaluated according to the equation: Rate = Pref * exp(-E/kT), which is the standard Arrhenius form found throughout chemistry. The variable E is the energy required to make the process occur, T is the temperature of the simulation, k is Boltzmann's constant, and Pref is a harmonic prefactor which is a measure of the entropy difference between the saddle point and the initial minimum.

E is known once the saddle is found, and kT are constant. The prefactor is evaluated as a ratio of frequencies at the saddle as compared to the minimum. For a system with 1000 atoms, a full calculation of the prefactor would require 3000 force calls (one for each x,y,z dimension for each atom in the system). Because the processes tend to be local, we only need to calculate the modes in a subspace of atoms which move during the process and their neighbors. So, getting to the question, HSize is the size of the Hessian matrix of second derivatives evaluated to find the prefactor. A total of 2*HSize force calls are required for this step, and then the Hessian matrix is diagonalized to find the frequencies.

So a typical search requires about 400 force calls to find a saddle, 200 to find the associated minima, and 200 to find the prefactor. Larger HSize will increase this last number, but this step does not tend to dominate the overall cost.

[graeme]

Oh, I forgot the last part of the question. HSize can vary because some processes are very local and involve the motion of only one or a few atoms. This will mean that most atoms stay in the same location, and their environment does not change. Other processes involve many atoms. For these processes, the frequency of more atoms contribute to the entropy of the process, and HSize must be larger to include these extra atoms.

[em99010pepe]

I'm getting a lot of HSize=0. Is that normal?

Carlos

[rsbriggs]

I'm not seeing that here, but nearly everything coming through right now shows "Bad Prefs"...

[graeme]

No, this is not normal. The calulation got itself into a state that it's having a hard time getting out of again. I'm not sure why, but most of the searches are returning saddles which do not connect back to the initial state. When this happens the search is deemed not relevant, and additional calculations, such as the prefactor calculation which reports HSize, are skipped. The information needed to get out of the state is slowly being built up, and it should transition to a new (hopefully normal) state in about 4 hours.