Returns a closure that, when called with a mini-batch, computes the combined data + physics loss and applies gradients.
Usage
make_train_step(
theta_model,
optimizer,
lambda_wet = 1,
lambda_dry = 10,
wet_split_cm = 4.2,
lambda3 = 1,
lambda4 = 1000,
lambda5 = 1000,
lambda6 = 1,
res_tensors,
pf_left = -2,
pf_right = 7.6
)Arguments
- theta_model
Keras model returned by
build_swrc_model()$theta_model.- optimizer
A Keras/TF optimizer (e.g.
tf$keras$optimizers$Adam).- lambda_wet
Weight for wet-end data loss (default
1.0).- lambda_dry
Weight for dry-end data loss (default
10.0).- wet_split_cm
Matric head threshold (cm) separating wet/dry (default
4.2).- lambda3
S1 linearity weight (default
1.0).- lambda4
S2 non-negativity weight (default
1000.0).- lambda5
S3 non-positivity weight (default
1000.0).- lambda6
S4 saturation weight (default
1.0).- res_tensors
Named list of physics residual tensors (output of
residual_to_tensors()).- pf_left
Left boundary of pF domain (default
-2).- pf_right
Right boundary of pF domain (default
7.6).