AutoPot Workflow

Presently, AutoPot consists of one workflow that can do a pre-selection of training configurations from a large set of training candidates using select, and/or a selection of training configurations during MD simulations using md_select.

After each selection round, selected configurations are automatically calculated with sp_calc and the potential is automatically re-trained using train, taking into account the newly calculated configurations. This process of selecting, calculating, and re-training is repeated until no uncertain configurations are found anymore.

The three stages can be called standalone or in sequence from the orchestrator. They are always called in the order calculate_ts -> select_and_calc_ts -> md_select_and_calc_ts.

The behavior of the orchestrator can be controlled by the following arguments:

usage: motoko orchestrator start [-h] [--detach] [--wait WAIT] --run_name RUN_NAME --potential 08.mtp [--fresh_potential 08.mtp] --training_set ts.cfg [--calculate_ts]
                                 [--atom_types W=0,Ta=1,...] [--select] [--grades [10000 1000 100 ...]] [--training_candidate_set tcs.cfg] [--md_select]
                                 [--md_configurations configurations.xyz]

options:
  -h, --help            show this help message and exit
  --detach, -d          For starting: run in detach/daemon mode with Zdaemon manager
  --wait WAIT           Waiting time between state checks in seconds
  --run_name RUN_NAME, -n RUN_NAME
                        run_name to use for all produced jobs
  --potential 08.mtp, -p 08.mtp
                        interatomic potential to start with
  --fresh_potential 08.mtp, -P 08.mtp
                        untrained interatomic potential to be used for training from scratch after collecting all training configurations
  --training_set ts.cfg, -t ts.cfg
                        initial training set
  --calculate_ts, -c    calculate the initial training set and train the potential on it
  --atom_types W=0,Ta=1,..., -a W=0,Ta=1,...
                        mapping of atomic types to numbers according to convention used in the potential code
  --select, -s          pre-select configurations from the candidate set
  --grades [10000 1000 100 ...], -g [10000 1000 100 ...]
                        Maximum extrapolation grades above which iteratively configurations are selected for training
  --training_candidate_set tcs.cfg, -T tcs.cfg
                        Set with the training candidates
  --md_select, -S       select configurations from MD simulations
  --md_configurations configurations.xyz, -C configurations.xyz
                        file with atomic configurations on which to run the MD simulations. If not given, run MD on each configuration from the training set