search

qzhu2017 / CSP_BO

Crystal Structure Prediction with Bayesian Optimization
0 stars 1 forks source link

Bayesian Optimization with GULP Results #33

Open yanxon opened 3 years ago

yanxon commented 3 years ago

@qzhu2017

Here is the BO result from 100 BO runs for predicting the ground state of Si. It takes about 445.6 s per 1 BO run. Figure_1

It looks encouraging as the previous study shows that the distribution looks like a Poisson distribution: medium

I thought we have the results for random run. But, I didn't have it save apparently since the changed. I'm running the random right now, and will update the graph after it's done.

This is the comparison between optimization with GULP after optimization with GPR vs optimization with GULP without prior GPR optimization. hist

qzhu2017 commented 3 years ago

@yanxon Actually, I just realized that this comparison won't be useful. We are using a surrogate model for BO, while the random search uses the actually model. Anyway, it is good to know that the BO works.

I suggest you

yanxon commented 3 years ago

@qzhu2017

Btw, you asked my the other day, why after the GPR optimization, the final energy is higher than the initial energy. I believe when I use FIRE, I set up the optimization step to stop at 20 steps. Sometimes, GPR can give non-sensical energy during the optimization. If the energy at the 20th step is not sensical, it will return that energy.

yanxon commented 3 years ago

@yanxon Actually, I just realized that this comparison won't be useful. We are using a surrogate model for BO, while the random search uses the actually model. Anyway, it is good to know that the BO works.

I suggest you

  • [ ] figure out the reason why sometimes the code stops

I believe after adding the Lennard Jones term, the code works from the beginning to end (100x BO run).

  • [ ] switch to the vasp interface now.

I will move on to this now.