Fit a physics-informed CNN1D SWRC model

The main user-facing function for training. Given prepared training and (optionally) validation data, it builds the model, creates physics residual sets, runs the training loop with early stopping, and returns a fitted object for prediction and evaluation.

Usage

fit_swrc(
  train_df,
  x_inputs,
  val_df = NULL,
  hidden = c(128L, 64L),
  dropout = 0.1,
  lr = 0.001,
  epochs = 80L,
  batch_size = 256L,
  patience = 5L,
  K = 64L,
  lambdas = norouzi_lambdas("norouzi"),
  S1 = 1500L,
  S2 = 500L,
  S3 = 500L,
  S4 = 1500L,
  pF_lin_min = 5,
  pF_lin_max = 7.6,
  pF0_pos = 6.2,
  pF1_neg = 7.6,
  pF_sat_min = -2,
  pF_sat_max = -0.3,
  wet_split_cm = 4.2,
  w_wet = 1,
  w_dry = 1,
  pf_left = -2,
  pf_right = 7.6,
  seed = 123L,
  verbose = TRUE
)

Arguments

train_df: Data frame for training (output of prepare_swrc_data()).
x_inputs: Character vector of covariate column names.
val_df: Optional validation data frame (same structure as train_df). If NULL, early stopping is skipped.
hidden: Integer vector of length 2: Conv1D filter counts (default c(128L, 64L)).
dropout: Dropout rate (default 0.10).
lr: Learning rate for the Adam optimizer (default 1e-3).
epochs: Maximum number of epochs (default 80).
batch_size: Mini-batch size (default 256).
patience: Early-stopping patience in multiples of 5 epochs (default 5).
K: Number of knot points (default 64L).
lambdas: Named list of loss weights; use norouzi_lambdas() to generate (default: norouzi_lambdas("norouzi")).
S1, S2, S3, S4: Residual set sizes (defaults: 1500, 500, 500, 1500).
pF_lin_min: Lower pF for S1 linearity constraint (default 5.0).
pF_lin_max: Upper pF for S1 linearity constraint (default 7.6).
pF0_pos: pF threshold for S2 (default 6.2).
pF1_neg: pF threshold for S3 (default 7.6).
pF_sat_min: Lower pF for S4 (default -2.0).
pF_sat_max: Upper pF for S4 (default -0.3).
wet_split_cm: Matric head (cm) separating wet/dry end (default 4.2).
w_wet: Sample weight for wet observations (default 1.0).
w_dry: Sample weight for dry observations (default 1.0).
pf_left: Left pF domain boundary (default -2.0).
pf_right: Right pF domain boundary (default 7.6).
seed: Random seed (default 123).
verbose: Logical; print progress (default TRUE).

Value

An S3 object of class swrc_fit, a named list containing:

theta_model: The fitted Keras model.
param_model: The theta_s extractor model.
x_inputs: Covariate names used.
scaler: Fitted min-max scaler.
K: Number of knot points.
dk: Knot spacing.
knot_grid: Knot positions in [0, 1].
pf_left,pf_right: pF domain boundaries.
theta_factor: Unit multiplier for theta.
best_epoch: Epoch at which validation loss was minimised.
lambdas: Loss weights used during training.
history: Data frame of per-epoch training/validation losses.

Examples

if (FALSE) { # \dontrun{
if (reticulate::py_module_available("tensorflow")) {
  df  <- prepare_swrc_data(swrc_example, depth_col = "depth")
  fit <- fit_swrc(df,
                  x_inputs = c("clay", "silt", "bd_gcm3", "soc", "Depth_num"),
                  epochs = 2L, verbose = FALSE)
}
} # }