PhasesResearchLab/ESPEI

Automated thermodynamic database development within the CALPHAD method using pycalphad

Jan Janssen

@jan-janssen

With 100 steps I already get a similar result, so maybe it is sufficient to restrict the number of monte carlo steps to 100 for the examples.

Jan Janssen

@jan-janssen

@bocklund Any update on updating the weights? I checked the website, but I did not see any updated information, so maybe you can give me a link to the updated weights.

Brandon Bocklund

@bocklund

Hi @jan-janssen, I haven't had time to go back to this yet. Looking through some files and your data, it might be sufficient to just increase the weights for the fcc solubility and the Laves solubility. If you have the vanilla datasets from the ESPEI-datasets repo, try adding weights in the following files:

zpf/CU-MG-ZPF-CUMG2-FCC_A1-LAVES_C15-LIQUID-Bagnoud1978.json set weight to 5.0
zpf/CU-MG-ZPF-FCC_A1-LAVES_C15-LIQUID-Jones1931.json weight to 10.0

If those weights work well, it would be great if you were able to submit a PR to the datasets repo to update those weights and to the ESPEI repo to update the documentation with the new images for the example. I agree with your previous comment that we can limit to the run in the example to maybe 100 or 200 iterations, verified with the lnprob vs. iterations plot

Jan Janssen

@jan-janssen

I updated the weights - the results changed a little bit but I am not sure if it is an improvement or not.

Brandon Bocklund

@bocklund

This looks almost identical to me. Can you share the files you changed?

Jan Janssen

@jan-janssen

I created a pull request with my changes PhasesResearchLab/ESPEI-datasets#17

Semen Kuzovchikov

@spike_k_gitlab

Hello, is there a way to plot activity datasets using ESPEI? I wish to compare them with results of mcmc optimisation, but don't have a clue how it could be done.

Brandon Bocklund

@bocklund

Hey @spike_k_gitlab, you can use pycalphad to plot the activity, along with your experimental data: https://pycalphad.org/docs/latest/examples/PlotActivity.html

Semen Kuzovchikov

@spike_k_gitlab

@bocklund thank you for reply. Also when I try adding experimental Hmix for liquid phase following Cu-Mg example ESPEI throws an error https://pastebin.com/ZW1T9AuL . Any ideas how it could be fixed?

Brandon Bocklund

@bocklund

@spike_k_gitlab can you share the TDB you are trying to fit, and also the version of pycalphad you are using?

Semen Kuzovchikov

@spike_k_gitlab

@bocklund I updated pycalphad to 0.8.3 using conda ~ 2 hours ago. Cannot figure out how to upload files in gitter, is it ok if I upload somewhere else and post link here?

Brandon Bocklund

@bocklund

You should be able to drag and drop the files into Gitter, but if you still have problems, you can email it to me at bjb54@psu.edu

Semen Kuzovchikov

@spike_k_gitlab

Sent you an email, thank you for your help.

Brandon Bocklund

@bocklund

Thanks @spike_k_gitlab. I have a meeting to attend, so I'll try to look at this in an hour or two.

Brandon Bocklund

@bocklund

Hey @spike_k_gitlab, it looks like this is caused by the database having the /- element defined. If you comment out the line starting with ELEMENT /-, that will allow ESPEI to run.

alsi_mur-modified.tdb

Here's the modified database with that line commented out

Semen Kuzovchikov

@spike_k_gitlab

Yes, that worked, thank you so much @bocklund

Semen Kuzovchikov

@spike_k_gitlab

Hello when running espei I keep getting some walkers stuck and not moving at all. I guess that it is not normal. Could it be something with approximate_equilibrium: true or maybe I should enlarge number of walkers or iterations siginifcantly?

Here is my lnprob trace.

Brandon Bocklund

@bocklund

Interesting. That is unusual. What level of output verbosity did you use? Are you able to see the probability in the output?

Semen Kuzovchikov

@spike_k_gitlab

i used verbosity:2 and yes I can see probabilities. I printed everything to stdout (maybe i shouldn't)

Brandon Bocklund

@bocklund

Are you able to scroll back and see what the initial probability for the parameters was?

You should see something like:

TRACE:root:Probability for initial parameters
TRACE:root:Likelihood - 2.44s - Thermochemical: -136.338. ZPF: -2363.344. Activity: -11834.342. Total: -14334.025.
TRACE:root:Proposal - lnprior: 0.0000, lnlike: -14334.0251, lnprob: -14334.0251

Semen Kuzovchikov

@spike_k_gitlab

No, i'm not. Will it be useful to start again for couple of iterations to see it?

Brandon Bocklund

@bocklund

Yeah, you could start again for just one iteration and see the same thing, just be sure not to delete your existing trace and lnprob files

And you database

I'm wondering how the probabilities after the jumps compare to the probabilities of the inital set of parameters. We start off by perturbing the parameters, so it's normal that they start a little worse and converge to (and improve on) your inital set of parameters.

Semen Kuzovchikov

@spike_k_gitlab

TRACE:root:Probability for initial parameters
/home/semikk/anaconda3/lib/python3.7/site-packages/pycalphad/core/lower_convex_hull.py:136: RuntimeWarning: invalid value encountered in double_scalars
  result_array_GM_values[it.multi_index] = new_energy / molesum
TRACE:root:Likelihood - 6.71s - Thermochemical: -1004.646. ZPF: -3424.380. Activity: -14400.540. Total: -18829.566.
TRACE:root:Proposal - lnprior: 0.0000, lnlike: -18829.5658, lnprob: -18829.5658

Brandon Bocklund

@bocklund

And what's the contribution to each of the types of data for chains with very low probability? I'm guessing that it's primarily the activity data that is reduced

Semen Kuzovchikov

@spike_k_gitlab

I guess you're right

TRACE:root:Likelihood - 15.94s - Thermochemical: -966.256. ZPF: -15987.196. Activity: -772364.421. Total: -789317.873.

Brandon Bocklund

@bocklund

The log-likelihood for the activity data in that proposal is about 50x greater than the input, so the set of parameters that correspond to that proposal will be rejected because the probability is much lower (since the overall probability is so much higher than the probability of any of your chains).

Can you find any examples where the Total is close to your -15,500 or -16,000?

For more context: the output will contain the likelihoods for all proposed sets of parameters (accepted and rejected), but the ones you see in the plot of -lnprob are only the accepted parameters that are in the Markov chains

Just based on how large the log-likelihood is there, my guess is that one or more of your parameters have diverged and you are sampling from very unreasonable parameters.

Can you see if the values of the parameters of the chains that have significant jumps in probability are reasonable? Plotting the parameters might be helpful, i.e. https://espei.org/en/latest/recipes.html#visualize-the-trace-of-each-parameter

Semen Kuzovchikov

@spike_k_gitlab

I'm sorry I've scrolled through all my available output and I can't find any of examples with Total close to -16,000. All of them either too large or near -18,000

Brandon Bocklund

@bocklund

That's okay. I'm pretty sure it's the activity data. Another way to check would be to start a new run for 1 iteration using the output database from your 1000 iteration runs as the input database, since the final database that you get is based on the parameter set with the highest probability.

But I think it would be more valuable to check into the parameters and see what their values are

Semen Kuzovchikov

@spike_k_gitlab

Parameter values don't seem unreasonable to me. What I is that starting from nearly 200 iteration some walkers just stop changing alt

Brandon Bocklund

@bocklund

Are these two parameters corresponding to coefficients a and b in an a+bT excess parameter?

Semen Kuzovchikov

@spike_k_gitlab

Yes they are, that's why they so differ in absolute values

Brandon Bocklund

@bocklund

Since the second one is a coefficient that's multiplied by T in the model, that's actually a really big change. A 40x change (-5 to -200) in this parameter would give a decrease in the energy of 40 kJ at 1000 K. Usually a good rule of thumb is that the b coefficient should be roughly 1/3000th of the a coefficient. So for a = -14000, b ~= -5.

Brandon Bocklund

@bocklund

I'd double check your activity data, plotting with the pycalphad link I sent the other day, if possible.

I noticed in the files you sent me before that there are some unusual values, e.g. Si-acr-75cha_1.json has

and

    "values":   [[[
      0.597,
      0.513,
      0.356,
      0.196,
      0.104,
      0.078,
      0.033,
      0.004,
      0.080
      ]]],

For a reference of pure Si, the compositions should follow the trend in the mole fraction of Si. The last two compositions are for X(SI)=0.790 and X(SI)=0.700, which don't match up well with the activities

The ones with pure Al reference state all seem okay, but some other pure Si ones you may want to check into, includuing Si-acr-75cha_2.json and Si-acr-75cha_3.json (maybe the activites are just out of order in this one?)

Semen Kuzovchikov

@spike_k_gitlab

Ok, my bad. Thank you, I should plot and double-check all the input data now, there may be other kind of errors. I'll reply if I encounter this problem after data checks.

Brandon Bocklund

@bocklund

My current hypothesis is that some of the inconsistent data may be contributing to an unrealistically large log-likelihood and that these non-physical data are dominating the likelihood. If that's true, then some of your parameters (like the second one above) may be able to wander parameters space until they find some sets of values that drastically reduce the log-likelihood for the non-physical values of activity, but these are just unlucky guesses and that causes the chains to get stuck there.

I have some alternate hypotheses we can explore if the bad data hypothesis turns out to be wrong or not fix the issue

Where communities thrive

People

Repo info

Activity