Solvent masking is off when SFcalculator is not provided with an experimental dataset

[alisiafadini]

I think the solvent masking is not estimated correctly if only a PDB file is provided (i.e. no experimental data) to SFcalculator.
This is observable in the real space electron density (e.g if loaded in Coot).

I believe the issue is that the scales are not initialized to good enough values when data is not provided.

In _set_scales, currently:

kiso=1.0,
kmask=0.35,
uaniso=[0.01, 0.01, 0.01, 1e-4, 1e-4, 1e-4]

But at the very least kmask should be resolution dependent with some average values. The following seems to work well and gives good real space density for a few test cases:

        kiso=1.0,
        kmask=0.00,
        uaniso=[1e-4, 1e-4, 1e-4, 1e-4, 1e-4, 1e-4]

with added resolution dependence

       self.kmasks = []

        for bin_i in np.sort(np.unique(self.bins)):
            index_i = self.bins == bin_i
            s_squared = 1 / np.mean(self.dHasu[index_i])
            kmask = 0.85 * np.exp(-250 * s_squared / 4)
            kmask_tensor = (
                torch.tensor(kmask).to(self.atom_pos_frac).requires_grad_(requires_grad)
            )
            self.kmasks.append(kmask_tensor)

I use the kmask expression:
kmask = ksol * exp( -Bsol * s**2 /4) with ksol = 0.85 (midway between 0.75-0.95)
and Bsol = 250 (midway between 150-350) from Glykos & Kokkinidis, Acta Crystallogr D Biol Crystallogr. 2000 Aug;56(Pt 8):1070-2

These are probably not the most representative values, but they seem to do the trick. Thoughts?

[minhuanli]

Sorry for the late reply. Before we were thinking the solvent mask is used as a modeling for better fitting the experimental observables, so it was not clear about how to apply a solvent mask without experiments.

But the values you provided all make sense. Also I feel this could be used to implement a new method to generate synthetic crystallography observables when only PDB is provided. I am working on the implementation and will create a PR in the torch version. Thanks for the suggestions!