How to use multiple calculators in the loss function?

[yonatank93]

Hello,

I have a test case where I need to use multiple calculators when training an empirical potential. Specifically, I have a dataset consisting diamond silicon configurations with various lattice constants. For each lattice constant, I have configurations with and without perturbation. My thought is to create a calculator for the unperturbed configurations, where I only compute the energy (due to symmetry, the forces will always be zero), and another calculator for the perturbed configurations that compute both the energy and forces. I was thinking of doing this because it might speed up the calculation, since I don't need to compute forces for some configurations.

I noticed that there is _WrapperCalculator for this case, however I had several issues when using this. First, when I used _WrapperCalculator, Loss created a neural network loss function instead. If I change this line with

if isinstance(calculator, (Calculator, _WrapperCalculator)):

so that I can get the physics motivated loss function, the next error I got is related to line 266,

AttributeError: '_WrapperCalculator' object has no attribute 'use_energy'

Do you have any suggestions about this? A work around for this case that I can think of is to use the weights. However, this alternative might not work in general, e.g., if we also want to fit elastic properties that requires a separate calculator.

[mjwen]

Hi @yonatank93, yes, _WrapperCalculator is one way to do it. It was disabled (thus the leading underscore in its name) at certain point because the lack of testing. I am not surprised it is not working now. Did a quick fix #95 to make it work. However, it is not well tested, so please use with caution. Would be happy to discuss how to make it more formal if you are interested.

[mjwen]

Forget to mention that there was an example script to use it. Updated as well. https://github.com/openkim/kliff/blob/master/examples/legacy/multiple_calculators.py

[yonatank93]

Thanks for this quick fix. This works for me, at least with some optimizers. If I use scipy.optimize.least_squares-based optimizer.

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/nobackup/scratch/usr/yonatank/practice/KLIFF/bootstrap/SW_multicalculators.py in <module>
    110
    111 # Train the model
--> 112 loss. Minimize(method="lm")
    113 model.echo_opt_params()

/nobackup/scratch/usr/yonatank/modules/kliff/kliff/loss.py in minimize(self, method, **kwargs)
    308             kwargs: extra keyword arguments that can be used by the scipy optimizer
    309         """
--> 310         kwargs = self._adjust_kwargs(method, **kwargs)
    311
    312         logger.info(f"Start minimization using method: {method}.")

/nobackup/scratch/usr/yonatank/modules/kliff/kliff/loss.py in _adjust_kwargs(self, method, **kwargs)
    335
    336             # adjust bounds
--> 337             if self.calculator.has_opt_params_bounds():
    338                 if method in ["trf", "dogbox"]:
    339                     bounds = self.calculator.get_opt_params_bounds()

AttributeError: '_WrapperCalculator' object has no attribute 'has_opt_params_bounds'

This is not a pressing matter, at least for my calculations. I still don't find a need to train the potential to other quantities beside energy and forces. And again for energy and forces calculation, I can still get around it with the weights. However, this might be something to consider in the future.