qlroms

Quantized-local reduced-order models (qlROMs). The state space is partitioned into charts by k-means, each chart gets its own local POD basis, and an exact shared atlas stitches the charts into one global coordinate system. On that geometry, this package advances the local dynamics with qlGalerkin: an intrusive projection of the governing equations, one quadratic reduced system $\dot{\bm{a}} = \bm{b}_k + \mathbf{A}_k\bm{a} + \mathsf{B}_k(\bm{a},\bm{a})$ per chart, plus the transition maps that carry coordinates across chart boundaries. The QLModel adapter wraps a qlROM as a dynamodels Model, so it plugs directly into romda estimators and ntsa analysis.

Install

conda create -n qlroms python=3.11   # >=3.10
conda activate qlroms
pip install qlroms
pip install -e ".[dev]"   # development

Quickstart

import numpy as np
from qlroms import free_run
from qlroms.intrusive_qlroms import ks1d
from qlroms.intrusive_qlroms.build_ks import build_fom, get_full_trajectory

# A case FOM and its (cached) trajectory: the 1-D KS chaotic regime, short window
fom, cfg = build_fom(ks1d, case="chaotic", overrides={"Ntrain": 20_000, "Ntest": 2_000})
traj = get_full_trajectory(Ntot=cfg["i0"] + cfg["Ntrain"] + cfg["Ntest"], model=fom, i0=cfg["i0"])
Xtrain, x0 = traj[:, :cfg["Ntrain"]], traj[:, cfg["Ntrain"]]

# Charts (k-means + local POD + atlas) and one projected (b, A, B) per chart
rom = ks1d.build_local_model(Xtrain, fom, K=10, r=30, save_dir=".")

# Closed-loop forecast through the shared step(apod) interface
X_rec, cluster_path = free_run(rom, x0, n_steps=500)

# Run it as a dynamodels Model (romda estimators, ntsa analysis)
from qlroms.model import QLModel
model = QLModel(rom, obs_idx=np.arange(0, fom.Nx, 8), psi0=x0)

Building a qlROM from a config

One YAML file per model: which test case and window, which family, how many charts and modes, and what to measure once it is fitted.

python -m qlroms.utils.builder configs/ks1d_chaotic.yml   # [--set rom.K=20] [--set case.params='{"Ntrain": 5000}']

case: {model: ks1d chaos}          # a qlroms.utils.config.CASES name, or ks1d/chaotic
rom:  {kind: qlgalerkin, seed: 0, clustering: {kmeans_method: minibatch}}
check: {n_steps: 1000, compare_global: true}

A case is defined in exactly one place -- its own module's TEST_CASES entry, which holds the physics, the windows it is generated with, the $(K, r)$ its ql-ROMs are built at and the timescales its characterization needs. qlroms.utils.config only maps a readable name onto one of those entries (get_case("ks1d chaos") reads it back live), and a config restates none of it: leave rom.K / rom.r or case.params out and the case's own values stand, set them and this run overrides them. Unknown keys fail loudly, dotted --set beats the file, and both the trajectory and the fitted ROM come from the case's cached pipeline. The run prints a JSON summary: representation error, closed-loop mean error and prediction horizon, against the global ($K=1$) ROM of the same family when compare_global is on.

Theory

Derivations (local POD with the mass-weighted inner product, OpInf regression with $\lambda_1$/$\lambda_2$ regularization, ESN training) live in docs/theory/; full documentation is published at https://andreanovoa.github.io/qlroms.

References

If you use this package, please cite the papers the method comes from:

• Colanera, A., & Magri, L. (2025). Quantized local reduced-order modeling in time (ql-ROM). Computer Methods in Applied Mechanics and Engineering, 447, 118393.
• Colanera, A., & Magri, L. (2026). Towards extreme event prediction of turbulent flows with quantized local reduced-order models. Journal of Physics: Conference Series, 3230(1), 012003. IOP Publishing.