ESN builder

Data

Sites (state size), N_dim

Time step of the data, dt

Upsample (ESN step = dt × upsample)

Training time, t_train

Validation time, t_val

Washout, N_wash (steps)

Layout

one reservoir parallel, shared parallel, independent

Patch size, G

Halo, H

Periodic domain

Inputs and readout

Observed sites (inputs)

Parameters, N_param

Parameters enter the reservoir

Readout reads

[r; 1] [r; u; 1] [r; p; 1]

Input noise, η (training only)

Reservoir

Units per reservoir, N_units

Connections per unit (connect)

Realizations of W_in and W, n_seeds

Seed of the first realization

Input matrix

sparse dense bias dense

Hyperparameters

Tick to search by Bayesian optimization in the range; untick to fix the value.

ρ search in – or fix at

σ_in search log₁₀ in – or fix at

β (tikh) search in or fix at

α (leak) search in – or fix at

Evaluations, N_func_evals

Initial grid per axis, N_grid

Validation

Strategy

RVC SSV WFV KFV

Folds, N_folds

Fold step, val_fold_step

Metric

log nMAE nMSE

Data split and validation folds

washout (open loop) rows of the ridge regression closed-loop validation

Steps: wash / train / val

-

Reservoir state, N_r

-

Readout weights trained

-

Search

-

Closed-loop validation runs

-

Architecture

folded unfolded in time

Equations

The equations that train() and the forecast solve with the current settings.

Code

Run it locally on your record data, of shape (N_t, N_dim).