amirhajibabaei / autoforce Goto Github PK

ActiveCalculator
In on-the-fly MD, if the initial atomic forces are zero and the model is empty/immature, model may remain blind in the following steps. The solution is to set covdiff to a finite value (~0.1).

parallel variance calculation in atoms-distributed regime is missing.

In calculation of the gradients of spherical harmonics, there is a division by sin(theta) in lines:
....
Y_theta = cos_theta * self.l_float * Y / sin_theta
Y_theta[1:, 1:] -= _r * Y[:-1, :-1] * self.coef / sin_theta
....
When sin_theta is 0, Y_theta becomes nan.
At the moment nan_to_num is applied when the gradient is returned which simply replaces nan with zeros.
But in calculations of the gradients of SOAP inconsistencies exist with autograd when xyz=[0,0,1].
The correct workaround may be to add a small positive number where sin_theta is zero.
But a beter solution maybe possible.

Define a addition kernel class, where two (or more) kernels are given at init.
func, leftgrad, rightgrad, gradgrad, etc for the product kernel can be deduced from argument kernels.

Theoretically
Stress=stress1 + stress2,
where
stress1 = (sum over all atoms F.T@r) / volume
where r is coordinates and F forces on atom.
stress2 is related to derivatives wrt cell.
This works fine in ParametricCalculator (test using calc.calculate_numerical_stress(atoms)).
But in AutoForceCalculator (machine learned) we have to multiply stress1 by -1 in order to get the right numerical stresses.
Everything seemingly works just fine but I still can't explain the -1 multiplyer.

Using sparse methods, calculating the full Gram matrix is bypassed.
Similarity calculations is only calculated between potentially many large (data) systems
and a few small inducing systems which reduces the computational complexity considerably.
Only the diagonal elements of the full gram matrix are needed;
either for calculation of the variance or the "trace" term in variational ELBO.
At the moment, covariance of all forces with each other (in one system) is calculated and
then the diagonal elements is extracted.
Instead the diagonal elements should be calculated directly, inside the similarity kernel.

When I chose noise=1e-6 (too small) in kernel, the predicted energies by the posterior potential contained a large shift from the actual energies.
I used shift because the predictions were perfectly correlated with the data (R2~1).
This could be related to the jitters added to the diagonal of the Gram matrix.
But other possibilities also exist.
First, we need to find out why this happens.
Second, the program should issue an error or a warning when this happens.

It looks like active calculator does not accept None type.

At the moment, when Local objects are sampled from data as inducing geometries, they include i, j indices which belong to the structure that they are embedded in.
Converting these to atoms, saving them in a traj, and reloading them causes these indices to be changed.
i, j indices are relevant via the "bothways" keyword in Local.select: if bothways=False, generally neighbors with j>i are returned.
Therefore in kernels such as PairKernel where bothways=False is frequently used, it is possible that an empty array is returned even if there are relevant atoms in the local environment.
Plus the behavior might change simply by saving and reloading locs.
This issue needs to be fixed.

Define a product kernel class, where two (or more) kernels are given at init.
func, leftgrad, rightgrad, gradgrad, etc for the product kernel can be deduced from argument kernels.

In conditional adding of data or references to a model, sometimes the order at which data and references are added becomes important.
For instance if an atoms object is added first, addition of its locals to references becomes less likely.
Thereof, in training by MD, situations raise when multiple data are added consecutively, but no references are added.
Usually this is accompanied by sharp discontinuities in the energy time series, every time a data is inserted.
One might refer to this as an stressed model.
This stress usually is eventually released when a few references are successfully added to the model.
What is the best order of adding data and references to a model, to avoid this stressed phases?