Cross-scale reservoir computing (CS-RC) is a hierarchical architecture for forecasting high-resolution spatio-temporal systems. It passes non-local, coarse-grained dynamics from lower-resolution reservoir layers to finer ones, combining long-range context with local spatial detail.
Alboré, N., Di Antonio, G., Coccetti, F., & Gabrielli, A. (2026).
Cross-scale reservoir computing for large spatio-temporal forecasting and
modeling. Neurocomputing, 692, 133849.
doi:10.1016/j.neucom.2026.133849.
See CITATION.cff for machine-readable paper citation metadata.
We propose a new reservoir computing method for forecasting high-resolution spatiotemporal datasets. While many machine learning methods suffer from limited interpretability and computational intractability in high-dimensional settings, reservoir computing offers inherent advantages through its lightweight linear training. Our hierarchical architecture combines multi-resolution inputs from coarser to finer layers, enabling the capture of both local dynamics and long-range dependencies. Applied to Sea Surface Temperature data, it outperforms standard parallel reservoir models in long-term forecasting, demonstrating the effectiveness of cross-layer coupling in improving predictive accuracy. We find that the optimal network dynamics in each layer become increasingly linear, revealing the slow modes propagated to subsequent layers. Finally, we evaluate the method on a chaotic system, where a strongly coupled two-layer configuration achieves accurate forecasts and offers favorable accuracy-cost tradeoffs compared with established reservoir baselines.
-
Clone the repository
git clone https://github.com/nicoalbo0/CrossScaleRC.git cd CrossScaleRC -
Install JuliaUse Julia 1.11+.
-
Activate the project and install dependencies
julia --project=.
In the Julia REPL:
using Pkg Pkg.instantiate()
Or from the shell:
julia --project=. -e 'using Pkg; Pkg.instantiate()' -
Run the package
using CrossScaleRCOr run a main script from the project root, e.g.:
julia --project=. main_single_layer.jl
The KS data used in the single- and multi-layer RC experiments can be generated with the MATLAB code from the parallelized-reservoir-computing repository:
- Repository: jdppthk/parallelized-reservoir-computing
- Script: KSBasicSingleReservoir/generate_data.m
That script integrates the Kuramoto–Sivashinsky equation (with an optional multiplicative asymmetry term) and saves the solution. Export or convert the output to CSV and place the file under data/kuramoto/ with the naming convention expected by load_data (e.g. Q128_L44_mu0.01_ks_data.csv for Q=128, L=44, μ=0.01). See src/data/loading.jl and the main scripts for the exact path and naming.
To use sea surface temperature data, follow the scripts in the scripts/ folder in this order:
scripts/download_sst_data.py(Python)Downloads daily SST from the Copernicus CDS (satellite sea surface temperature ensemble product) for the chosen years/months/days. Requires a CDS API key andcdsapi. Writes files underdata/sst/YYYY/MM/DD/.scripts/main_regrid.jl(Julia)Reads the raw daily files and regrids them to the desired resolution (e.g. 2° or 18°), applying scale and offset. Saves regridded fields as JLD2 files (e.g.sst_regridded_<resolution>.jld2per day).scripts/main_produce_sst_data.jl(Julia) Concatenates all daily regridded JLD2 files into a single 3D array (lon × lat × time) and saves it asdata/sst/sst_final_<resolution>.jld2. The package’sload_data(resolutions_vec; ...)then uses thesesst_final_*(and buildssst_clean_*if needed).
Run the Python script from the project root so that data/sst/ is created in the right place; run the Julia scripts with Pkg.activate(".") or from the repo root with --project=. so that CrossScaleRC and paths resolve correctly.
These are the main entry points for running experiments.
| File | Description |
|---|---|
| main_single_layer.jl | Single-layer reservoir computing on 1D KS data. Loads KS data, optionally regrids, builds spatial blocks, trains one ESN per block with local + neighbor inputs, and runs closed-loop prediction. Good starting point for the block-wise RC pipeline. |
| main_multi_layer.jl | Two-layer (coarse → fine) RC on 1D KS data. Coarse layer is run first; its predictions are upscaled and fed as “layer” input to the fine layer. Usesrun_multi_layer and 1D blocks from make_blocks(..., overlap_mode=...). |
| main_deep_rc.jl | Deep Echo State Network (DeepESN) on KS data. Multi-layer reservoir with quadratic readout features. Trains withDeepESN_train! and runs closed-loop test with DeepESN_test_closed_loop. No spatial blocking; single global deep reservoir. |
| main_ngrc.jl | Next-Generation RC (NG-RC) on KS data. Uses polynomial features on a delay embedding and ridge regression for the one-step map, then closed-loop prediction. No reservoir; lightweight and fast. |
| main_sst.jl | Two-layer RC onsea surface temperature data. Loads multi-resolution SST via load_data([18.0, 6.0]; ...), builds 2D blocks with make_blocks(data, grids, mixing), flattens to matrices, and runs run_multi_layer (coarse + fine) with 2D block structure. |
All main scripts set up data, hyperparameters, and (where applicable) blocks, then call the appropriate high-level API (run_single_layer, run_multi_layer, DeepESN_*, nextgen_closedloop). Tuning and grid-search are handled by the run_tuning_*.jl scripts.
