checking the cluster gradient

To solve the scaling wall for QM methods the full biomolecular gradient can be computed in qr.refine using a fragmentation scheme. A fragment consists of a cluster and a buffer region, the latter is obtained from non-covalent interactions density descriptors. Additionally, a surrounding point charge (PC) region (controlled by charge_embedding, charge_cutoff) can be applied. Furthermore, the two_buffer option adds another layer of buffer around a previously constructed fragment.

The convergence of the gradient obtained via fragmentation can be tested using mode=gtest.

The list of available keywords can be found under the cluster scope here

Central variable is called g_mode and indicates what kind of buffer region is applied:

1 = default buffer
2 = default buffer + PC
3 = two buffer
4 = two buffer + PC

two fundemantal modes of operation

Sequence of calculations (recommended): specify two_buffer=1 and/or charge_embedding=1 in the command line. This specifies the largest buffer region up to which calculations are performed.
Direct buffer selection: specify g_mode=1/2/3/4 in the command line to select what kind of buffer region is calculated. Other kinds of buffer regions are not calculated.

output statistics

example output:

    index(g_mode - max_res)
1 - 10   d(angle)  0.914806
1 - 10   d(gnorm)  1.351992
1 - 10   d(max_g)  0.266452
1 - 10   d(min_g)  0.266452
1 - 10   MAD       0.514772

1 - 25   d(angle)  0.000000
1 - 25   d(gnorm)  0.000000
1 - 25   d(max_g)  0.000000
1 - 25   d(min_g)  0.000000
1 - 25   MAD       0.000000

The first number corresponds to the buffer kind (g_mode, see above) and the second to the maximum numbers of residues per cluster. The last gradient is selected as the reference gradient (or the one set by g_ref) and various comparisons are done.

comparisons:

d(angle)  : mean angle deviation between the gradient of all atoms in degree.
d(gnorm)  : difference between the gradient norms
d(max/min): difference between max/min gradient values
MAD       : mean absolute deviation between all gradient components

Example

Running a sequence of gradient calculations with a maximal number of residue per cluster of 10, 15 and 25, using default buffer and point-charges with GFN2-xTB.

qr.refine model.pdb engine_name=xtb mode=gtest g_scan="10 15 25" charge_embedding=1

A total of 6 gradient calculations will be performed. 3 with the default buffer and maxnum_residues_in_cluster=10/15/25 and 3 with the default buffer plus point-charges with the same maxnum_residues_in_cluster settings. At the end the largest gradient (from g_mode=2 and maxnum_residues_in_cluster=25) will be used as reference for a comparison against the other 5 gradients. Deviations against the gradient norm (gnorm) against the max/min value of the gradient (max/min_g) and the mean absolute deviation (MAD) of each gradient component are printed.

This will also produce a file called 2-25.npy, which is the gradient of the largest buffer region. This gradient can be used as reference gradient for another calculation using cluster.g_ref=2-20.npy.