ddmo: A Compatible PySINDy Engine for Feature Optimization in Sparse Dynamical-System Discovery
- 1Fritz-Haber-Institut of the Max Planck Society, Faradayweg 4–6, Berlin, Germany, 14195
- 2University of Colorado Boulder
Abstract
Sparse equation discovery depends on the candidate functions, the treatment of noisy data, and the sparse regression procedure. These components are commonly configured independently, although their interactions directly affect model recovery. Our data-driven model optimizer ddmo uses PySINDy as its default sparse-regression backend and adds a modular, multiobjective optimization layer around it. Within one configurable search, users can vary parameters of prescribed candidate functions, regression settings, and the test functions and windows used to construct weak-form equations. For each configuration, the package builds the regression problem, estimates model coefficients with the selected backend, and evaluates the resulting model against objectives that include predictive accuracy and sparsity. This workflow makes it possible to explore how these modeling choices interact without writing a separate optimization pipeline for each study. We demonstrate the framework using the nonlinear Goodwin oscillator under increasing levels of measurement noise. The optimized weak formulation improves noise robustness relative to the strong formulation and supports accurate recovery of sparse dynamics. Joint library optimization explores competing nonlinear function families and estimates their parameters. The package provides a unified optimization layer for investigating how library design, weak formulation, and regularization jointly affect sparse equation discovery.
Related
- Machine-Learning-Compatible CALPHAD-type Optimization from Phase Equilibria by Auto-differentiation
- Feature tracking in physics-informed neural networks via joint optimization of nonlinear deformation manifolds: application to shocks
- Optimization of fluid mixing by reinforcement learning using limit cycles of a dynamical system
- QTUNER: A Feature-Guided Adaptive Framework for Quantum-Compiler Optimization Pass Tuning
- Discovery of connectivity-trainability trade-off of IQP circuits for Hamiltonian optimization
BibTeX
@misc{lombardi2026ddmo,
title = {{ddmo: A Compatible PySINDy Engine for Feature Optimization in Sparse Dynamical-System Discovery}},
author = {Juan Manuel Lombardi and Maryke Kouyate and Susanne Fürst and Rainey Lyons and Christian Kunkel and Karsten Reuter and Christoph Scheurer and Gianmarco Ducci},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010479/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010479/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - ddmo: A Compatible PySINDy Engine for Feature Optimization in Sparse Dynamical-System Discovery
AU - Juan Manuel Lombardi
AU - Maryke Kouyate
AU - Susanne Fürst
AU - Rainey Lyons
AU - Christian Kunkel
AU - Karsten Reuter
AU - Christoph Scheurer
AU - Gianmarco Ducci
PY - 2026
DA - 2026/10/14
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010479/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010479/v1
AB - Sparse equation discovery depends on the candidate functions, the treatment of noisy data, and the sparse regression procedure. These components are commonly configured independently, although their interactions directly affect model recovery. Our data-driven model optimizer ddmo uses PySINDy as its default sparse-regression backend and adds a modular, multiobjective optimization layer around it. Within one configurable search, users can vary parameters of prescribed candidate functions, regression settings, and the test functions and windows used to construct weak-form equations. For each configuration, the package builds the regression problem, estimates model coefficients with the selected backend, and evaluates the resulting model against objectives that include predictive accuracy and sparsity. This workflow makes it possible to explore how these modeling choices interact without writing a separate optimization pipeline for each study. We demonstrate the framework using the nonlinear Goodwin oscillator under increasing levels of measurement noise. The optimized weak formulation improves noise robustness relative to the strong formulation and supports accurate recovery of sparse dynamics. Joint library optimization explores competing nonlinear function families and estimates their parameters. The package provides a unified optimization layer for investigating how library design, weak formulation, and regularization jointly affect sparse equation discovery.
ER -