[BUG]: Running a simulation using a calibrated configuration gave an error or blank results

[liunelson]

@sabinala

We have a relatively simple model and calibrated it successfully, producing a configuration.

And we were able to simulate it with dopri5:

However, when we try to run a simulation again using the calibrated configuration, we get:

• an error underflow in dt 0.0 when using dopri5

• no error with rk4 or euler (step size = 0.1) but the returned results is mostly blank except the single-valued parameters
  

This happens in this workflow:
https://app.staging.terarium.ai/projects/419b04dd-3499-49d1-9f89-ca357a4affb7/workflow/0ba33700-72ed-4a62-8a03-a830f297e399

[liunelson]

I'll try to send the calibrated configuration files to test outside of Terarium.

[sabinala]

Hi @liunelson, I don't have access to the workflow so to clarify, you're saying that the model was calibrated and simulated successfully from the calibrate operator, but when you try to use the calibrated parameter distributions as a model configuration and simulate that, you are getting an error with dopri5?

A few questions/ideas:

• How does this work exactly? Are you assuming that after calibration the parameter value is independent and normally distributed with the given mean and standard deviation (where these values are calculated from the simulated samples)? And this is what is input into MIRA as (something like):

'parameter_name': Parameter(name='parameter_name', value=sympy.Float(calibrated_mean), units=per_day_units(),
               distribution=Distribution(type='Normal',
                                              parameters={'mean': calibrated_mean,
                                                          'standard_deviation': calibrated_std})) ?

• If so, I'm not surprised you're getting some issues when you try to simulate as (unlike with the uniform distributions) there is really no upper or lower bound on the parameter values in this case, beta could be huge or even negative. Not sure exactly how you'd remedy this without some user-imposed bounds. Moreover, this is related (an approximation) to how PyCIEMSS samples from a calibrated model using inferred_parameters, but not the same.
• The mean beta value is 0.3794, that seems large, but maybe ok given the large population. I'd be curious to see what the plots look like for the infectious population.
• As for the solvers, using euler will get you a solution, but it can be extremely inaccurate, especially with a large step size of 0.1. In the screenshots where you're using rk4, it looks like you're using a step size of 1.0. Again, this will get you an answer but it will be super inaccurate. These fixed step size solvers keep "plodding along" even when they get nonsense, or infinite value answers, so it' possible/likely that the blank spaces are actually NaN values - you're just not getting an error message like with dopri5 since no next step needs to be calculated.
• In conclusion, I recommend simulating calibrated models from the calibrate operator where the inferred_parameter guide is used instead of feeding an approximation of this into the simulate operator.

[sabinala]

Closing because we've figured out what the issue was, see above comment/discussion