active learning lot of duplicate configurations
Dear developer,
I found that in an active learning case, an updated training set should be used to retrain the potential. I have started with two different kinds of initial training sets and started active learning with mlip-3 and mlip-2 as well. When my calculations were done, I found a lot of duplicate structures were added to the training set during active learning. I am just worried about how we can avoid such situations. Do we need to test extrapolation_control:threshold_break=10 value for each new system? In addition, I have attached my lammps data file for job submission and all the rest of the data files as well. In your provided examples files, you have always started with one single configuration. How does mlip generate new structures? how much it will be reliable for simulations if we start? I will be thankful for your kind response. One more short question, using mlip-3 with lammps when we use select=true why does it show to turn off the configuration selection method?