QuantumATK Forum

QuantumATK => General Questions and Answers => Topic started by: korandofficial on June 16, 2023, 13:20

Title: Active Learning initiating method for crystalline configuration
Post by: korandofficial on June 16, 2023, 13:20

Hi there,

I want to calculate the potential for a crystalline configuration. Since I received different values for training and testing errors in Batch Learning, I want to apply Active Learning. In this regard I found that initiating Active Learning simulation by calling runOptimizeGeometry for a crystal would be smoother than runMolecularDynamics. Is this hypothesis true?  ???
Also, I would appreciate it if you could let me know that in which version of ATK runOptimizeGeometry for Active Learning is available? I have 2021.06 and I think it doesn't have this function.

Title: Re: Active Learning initiating method for crystalline configuration
Post by: Anders Blom on June 21, 2023, 22:39

You would need 2022.03 but I strongly recommend 2022.12 as many improvements were made.

I am actually not sure what you mean about "initialize" the active learning. You can indeed run active learning on a geometry optimization, but after the batch training the potential should already be capable of an accurate geometry optimization, so this might not really improve the potential much. The point of active learning is to explore regions of phase space that the potential does not yet cover.