kooplearn: A Scikit-Learn Compatible Library of Algorithms for Evolution Operator Learning

TL;DR

kooplearn is a Scikit-Learn compatible library for evolution operator learning, supporting linear, kernel, and deep learning estimators.

cs.LG 🔴 Advanced 2025-12-25 2 views
Giacomo Turri Grégoire Pacreau Giacomo Meanti Timothée Devergne Daniel Ordonez Erfan Mirzaei Bruno Belucci Karim Lounici Vladimir Kostic Massimiliano Pontil Pietro Novelli
machine learning dynamical systems spectral decomposition evolution operators data-driven models

Key Findings

Methodology

kooplearn implements linear, kernel, and deep learning estimators for learning dynamical operators and their spectral decompositions. It models both discrete-time evolution operators (Koopman/Transfer) and continuous-time infinitesimal generators. By learning these operators, users can analyze dynamical systems via spectral methods, derive data-driven reduced-order models, and forecast future states and observables.

Key Results

  • Experiments show that kooplearn's kernel-based Reduced Rank Regression outperforms traditional methods in approximating the operator's spectrum, especially in overdamped Langevin dynamics.
  • Using the Lorenz 63 dynamical system dataset, kooplearn's kernel model demonstrated significant fit time advantages on 5000 observations.
  • Performance on high-dimensional structured data was validated using the Ordered MNIST dataset.

Significance

kooplearn provides a powerful tool for researchers to analyze complex dynamical systems, particularly where efficient spectral decomposition and dimensionality reduction are required. Its compatibility and open-source nature make it ideal for academia and industry, addressing long-standing challenges in dynamical system modeling and prediction.

Technical Contribution

kooplearn introduces fast kernel estimators and infinitesimal generator models for SDEs, providing new theoretical guarantees. It also incorporates specialized losses for deep representation learning, expanding engineering possibilities in evolution operator learning.

Novelty

kooplearn is the first open-source library to implement kernel-based Reduced Rank Regression, significantly enhancing the accuracy of operator spectrum approximation compared to existing DMD methods.

Limitations

  • Computational costs may be high when handling extremely large datasets, especially in deep learning models.
  • The complexity of certain dynamical systems may lead to reduced prediction accuracy.

Future Work

Future directions include extending kooplearn to support more types of dynamical systems, optimizing computational efficiency, and exploring new deep learning representation methods.

AI Executive Summary

kooplearn is a Scikit-Learn compatible machine learning library focused on evolution operator learning. Existing methods often fall short in modeling and predicting dynamical systems, and kooplearn offers an efficient solution by implementing linear, kernel, and deep learning estimators.

The library supports modeling of both discrete-time and continuous-time dynamical operators, allowing users to analyze dynamical systems via spectral methods, derive data-driven reduced-order models, and forecast future states. Its interface is compliant with the Scikit-Learn API, facilitating integration into existing machine learning and data science workflows.

kooplearn also includes curated benchmark datasets to support experimentation, reproducibility, and fair comparison of learning algorithms. Its open-source nature makes it an ideal choice for research and education, advancing the analysis of dynamical systems.

Deep Analysis

Background

Dynamical systems permeate every scientific discipline, from fluid flows to atomistic motions. Evolution operator learning is a general and principled framework, especially well-suited for interpretability and dimensionality reduction. An evolution operator E characterizes dynamical systems, either stochastic or deterministic, through spectral decomposition expressing dynamics as a linear superposition of modes.

Core Problem

Existing methods for modeling dynamical systems often face challenges in interpretability and computational efficiency. Effectively learning evolution operators and making accurate predictions is a crucial and challenging problem.

Innovation

kooplearn introduces kernel-based Reduced Rank Regression and infinitesimal generator models, providing new theoretical guarantees and engineering possibilities. Its fast kernel estimators significantly enhance the accuracy of operator spectrum approximation, offering notable innovations compared to existing DMD methods.

Methodology

  • �� Implements linear, kernel, and deep learning estimators for learning dynamical operators and their spectral decompositions.
  • �� Supports modeling of discrete-time evolution operators and continuous-time infinitesimal generators.
  • �� Provides a Scikit-Learn compatible interface for easy integration.
  • �� Includes curated benchmark datasets to support experimentation and reproducibility.

Experiments

Experiments utilized datasets such as the Lorenz 63 dynamical system, comparing the performance of kernel DMD and Reduced Rank estimators. Multiple independent runs on an Intel Core i9-9900X CPU validated kooplearn's speed advantages.

Results

Results show that kooplearn's kernel-based Reduced Rank Regression provides more accurate approximation of the operator's spectrum in overdamped Langevin dynamics. Its kernel model demonstrated significant fit time advantages on 5000 observations.

Applications

kooplearn can be used for analyzing and predicting complex dynamical systems, particularly where efficient spectral decomposition and dimensionality reduction are required. Its open-source nature makes it ideal for academia and industry.

Limitations & Outlook

Computational costs may be high when handling extremely large datasets, especially in deep learning models. The complexity of certain dynamical systems may lead to reduced prediction accuracy.

Plain Language Accessible to non-experts

Imagine a factory where kooplearn acts like a smart production line. It analyzes past production data to predict future trends, similar to how factory machines adjust processes for different products. kooplearn adjusts its algorithms for different dynamical systems to achieve optimal prediction results.

ELI14 Explained like you're 14

Hey there! Imagine you're playing a super complex game, and kooplearn is like your game assistant. It helps you analyze every move in the game and predicts what will happen next. Just like you need to make quick decisions in the game, kooplearn helps scientists quickly analyze complex dynamical systems and make smart predictions!

Glossary

Koopman Operator

A linear operator used to describe the evolution of dynamical systems.

Used to analyze deterministic dynamical systems.

Transfer Operator

Used to describe the evolution of stochastic dynamical systems.

Used to analyze stochastic dynamical systems.

Spectral Decomposition

Expressing dynamical systems as a linear superposition of modes.

Used to express the dynamic properties of evolution operators.

Kernel Ridge Regression

A kernel-based regression method for approximating the operator's spectrum.

Used to improve the accuracy of operator spectrum approximation.

Infinitesimal Generator

A differential operator used to describe continuous-time dynamical systems.

Used to model the evolution of continuous-time dynamical systems.

Open Questions Unanswered questions from this research

  • 1 How to maintain computational efficiency on extremely large datasets?
  • 2 How to improve prediction accuracy for complex dynamical systems?

Applications

Immediate Applications

Dynamical System Analysis

Researchers can use kooplearn to analyze the behavior of complex dynamical systems.

Data-Driven Prediction

Businesses can use kooplearn for data-driven future trend prediction.

Long-term Vision

Intelligent Dynamical Systems

kooplearn may become a standard tool for intelligent dynamical system analysis.

Abstract

kooplearn is a machine-learning library that implements linear, kernel, and deep-learning estimators of dynamical operators and their spectral decompositions. kooplearn can model both discrete-time evolution operators (Koopman/Transfer) and continuous-time infinitesimal generators. By learning these operators, users can analyze dynamical systems via spectral methods, derive data-driven reduced-order models, and forecast future states and observables. kooplearn's interface is compliant with the scikit-learn API, facilitating its integration into existing machine learning and data science workflows. Additionally, kooplearn includes curated benchmark datasets to support experimentation, reproducibility, and the fair comparison of learning algorithms. The software is available at https://github.com/Machine-Learning-Dynamical-Systems/kooplearn.

cs.LG