Hierarchical Neural ODE over multiple pedons (Pillar 2 × Pillar 5)

Fits a covariate-conditioned Neural ODE \(dy/dz = f_\theta(z, y, \mathbf{x})\) jointly across every pedon in a long-format data frame, plus a sibling MLP \(y_0(\mathbf{x})\) that predicts the surface value from covariates. The two nets share the optimiser so training sees all pedons at once.

Usage

piml_hierarchical_fit(
  pedons,
  id_col,
  depth_col,
  value_col,
  covariate_cols,
  hidden = c(32L, 16L),
  y0_hidden = c(16L),
  n_steps = 4L,
  epochs = 500L,
  lr = 0.01,
  seed = NULL,
  verbose = FALSE
)

Arguments

pedons

Data frame in long form (one row per horizon) with columns for pedon id, depth, value, and one or more covariates.

id_col, depth_col, value_col

Character, column names.

covariate_cols

Character vector, names of covariate columns. Assumed constant within a pedon — the first row of each pedon is used.

hidden

Integer vector, MLP widths for the ODE vector field \(f_\theta\).

y0_hidden

Integer vector, MLP widths for the surface-value head \(y_0(\mathbf{x})\).

n_steps

Integer, RK4 steps between successive observation depths.

epochs, lr

Integer and numeric — Adam hyperparameters.

seed, verbose

As elsewhere in edaphos.

Value

A edaphos_piml_hierarchical object (S3).

Details

Once fitted, piml_hierarchical_predict() returns a profile for any new location given its covariates — no horizon measurement required. Pair with al_physics_gate_piml_hierarchical() to use that per-location envelope as a rejection filter inside the Active Learning loop.