Allow reference calibration files to be included, attached to either an Instrument or ExperimentSet?

[jchodera]

It would be useful to be able to add calibration datasets (like water titrations) to use in inference, since these help constrain parameters like sigma

[bas-rustenburg]

Just wanted to share an observation of the current state of the code.

As mentioned in #1, the noise is likely proportional to the filter-time.

Data is already stored when read from a microcal itc file, just not used. I'm actually not sure how to implement it. Would we divide the final answer by the filter time? What if it is not the same per injection, divide by average? Or, do we stick this in at the model level.

[jchodera]

Filter time (the time between writing samples) will be constant over the whole .itc file for MicroCal data, but I suppose it is possible that someone will try to load files with different filter times to use together.

I think it might make sense to use a fundamental baseline noise \sigma where we model the measurement error in an integrated injection heat as (\sigma_0^2 * T_inject), where T_inject is the time between injections N_samples * \tau, where N_samples are collected per injection and \tau is the time between samples. The error in each single measurement (for modeling uncertainty in baselines) would be \sigma_0^2 * \tau.

[jchodera]

Note that I think \sigma_0^2*\tau should be related to the variance in the Gaussian process baseline fit, though I'm not sure if we can use this connection in the current model.

[jchodera]

\sigma_0 would be an intrinsic instrument noise parameter, though I realize it may depend on some instrument settings (like differential power). So we may want to ensure that all experiments in a set analyzed together have the same differential power for now (and maybe also the same filter time).

[bas-rustenburg]

Filter time (the time between writing samples) will be constant over the whole .itc file for MicroCal data

Is it truly a setting that is determined for the entire protocol? I ask, because it is actually specifically written down for each injection in the .inj file and .itc header. Last column.

$ITC
$ 11
$NOT
$ 25
$ 300
$ 1000
$ 5
$ 2
$ADCGainCode:  0
$False,True,True
$ 0.2 , 0.4 , 60 , 1
$ 3 , 6 , 120 , 0.5
$ 3 , 6 , 120 , 0.5
$ 3 , 6 , 120 , 0.5
$ 3 , 6 , 120 , 0.5
$ 3 , 6 , 120 , 0.5
$ 3 , 6 , 120 , 0.5

And it actually seems to be different for the first injection. This is one of the host-guest experiments by the way. We should change that, unless there is a good reason for it to be different.

I think it might make sense to use a fundamental baseline noise \sigma where we model the measurement error in an integrated injection heat as (\sigma_0^2 * T_inject), where T_inject is the time between injections N_samples * \tau, where N_samples are collected per injection and \tau is the time between samples. The error in each single measurement (for modeling uncertainty in baselines) would be \sigma_0^2 * \tau.

That makes sense. I will definitely try to implement something like that at some point in the future.

[jchodera]

And it actually seems to be different for the first injection.

Goodness, I think you're right!

This is one of the host-guest experiments by the way. We should change that, unless there is a good reason for it to be different.

No, we should change this!