Skip to contents

Fit a DeepONet for depth-profile operators

Source: R/pilar8_neural_operators.R
no_deeponet_fit.Rd

Fit a DeepONet for depth-profile operators

Usage

no_deeponet_fit(
  depths,
  targets,
  covariates,
  branch_hidden = 16L,
  trunk_hidden = 16L,
  output_dim = 8L,
  epochs = 300L,
  lr = 0.02,
  seed = NULL,
  backend = c("r", "torch"),
  device = c("cpu", "mps", "cuda")
)

Arguments

depths

Numeric vector of depths (length n_depths).

targets

Matrix of targets, shape (n_obs, n_depths).

covariates

Matrix of per-site summary covariates, shape (n_obs, p) – NOT a depth-dependent trajectory; each site is represented by a vector of static covariates. This is the canonical DeepONet setup where the branch input is a fixed- length vector.

branch_hidden, trunk_hidden

Integer hidden sizes for the branch and trunk MLPs.

output_dim

Integer; dimension of the inner-product space (p in the notes above).

epochs, lr

Training hyperparameters.

seed

RNG seed.

backend

"r" (default) or "torch" (full autograd).

device

"cpu", "mps", or "cuda" when backend = "torch".