config.toml

[global]
root = '.'                   # Root directory
random_seed = 3407

[train]
cutoff = 4.0                 # Cutoff radius of machine learning potential
val_ratio = 0.1              # The ratio of validation points in the provided dataset
num_interactions = 3         # Number of message-passing layers
node_size = 128              # Node feature size
output_dir = 'model_output'  # Model output path
dataset = '/content/drive/My Drive/my-path-to/tutorial-video/4waters.traj'  # Dataset for training
max_steps = 2000             # Maximum steps for training
device = 'cuda'              # Use GPU
batch_size = 32              # Batch size for training
initial_lr = 0.001           # Initial learning rate
forces_weight = 0.98         # Ratio of force loss to total loss
log_interval = 200            # Evaluate model every 200 steps
normalization = false        # Normalize energy in the dataset
atomwise_normalization = false  # Normalize atomic energy, scale the output atomic energy to the same level.
stop_patience = 1000         # When test loss is larger than training loss for p times, training stops.
plateau_scheduler = true     # Use ReduceLROnPlateau scheduler to decrease lr when learning curve plateaus
random_seed = 3407           # Random seed ensures the reproducibility of experiments

[train.ensemble]
108_node_3_layer = {node_size = 108, num_interactions = 3}
112_node_3_layer = {node_size = 112, num_interactions = 3}
#116_node_3_layer = {node_size = 116, num_interactions = 3}  # Commented out
#124_node_3_layer = {node_size = 124, num_interactions = 3}  # Commented out

[train.resource]
tmax = '10h'                 # Time limit for each job
cores = 8                    # Cores on the node

[MD.runs.water]
init_traj = '/content/drive/MyDrive/my-path-to/tutorial-video/4waters.traj'  # Initial configuration for running MD
time_step = 0.1              # Time step for MD
temperature = 300            # Temperature for MD
device = 'cuda'              # Use GPU
start_indice = 1             # Select initial configuration
max_steps = 3000             # Maximum MD steps
min_steps = 150              # Minimum MD steps
dump_step = 5                # Dump a structure for every 5 steps
friction = 0.01              # Friction coefficient
print_step = 2               # Print MD info for every 2 steps
num_uncertain = 100          # If 100 uncertain structures are collected, the simulation stops
random_seed = 3407           # Reproducibility

[MD.resource]
#tmax = '10h'                # Commented out
cores = 1                    # Cores on the node

[select.runs]
water = {'method' = 'MD', 'train_set' = '/content/drive/My Drive/my-path-to/tutorial-video/4waters.traj', 'kernel' = 'full-g', 'selection' = 'lcmd_greedy', 'n_random_features' = 500, 'batch_size' = 10, 'device' = 'cuda', 'random_seed' = 3407}

[select.resource]
tmax = '10h'                 # Time limit for each job
cores = 1                    # Cores on the node

[labeling.runs]
method = 'GPAW'

[labeling.runs.water]
nupdown = 48

[labeling.resource]
tmax = '10h'                 # Time limit for each job
cores = 1                    # Cores on the node