# randomness
seed: 91  # 91
# dataset
generator:
  # options: "bezier_symmetric_double", "bezier_symmetric", "bezier_asymmetric", "bezier_lerp"
  name: "bezier_symmetric_double"  # "bezier_symmetric_double"
  bounds: "saddle"  # options: pillow, dome, saddle
  num_uv: 10  # 10, 16, 23
  size: 10.0
  num_points: 4  # grid points
  lerp_factor: 0.5  # scalar factor in [0, 1] to interpolate between 2 surfaces, only for bezier_lerp
# simulator
fdm:
  load: -0.5  # -0.5, scale of vertical area load
# neural networks
encoder:
  shift: 0.0
  hidden_layer_size: 256
  hidden_layer_num: 3
  activation_fn_name: "elu"
  final_activation_fn_name: "softplus"  # needs softplus to ensure positive output
decoder:
  # If true, the decoder maps (z, boundary conditions) -> x. Otherwise, z -> x.
  include_params_xl: True
  hidden_layer_size: 256
  hidden_layer_num: 3
  activation_fn_name: "elu"
# loss function
loss:
  shape:
    include: True
    weight: 1.0  # weight of the shape error term in the loss function
  residual:  # PINN term
    include: True
    weight: 1.0  # weight of the residual error term in the loss function
# optimization
optimizer:
  name: "adam"
  learning_rate: 3.0e-5  # 3.0e-5 (formfinder),  5.0e-5 (others). Be careful with scientific notation in YAML!
  clip_norm: 0.0
# training
training:
  steps: 10000  # 10000
  batch_size: 64