Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

DY: Fachverband Dynamik und Statistische Physik

DY 6: Machine Learning in Dynamics and Statistical Physics I

DY 6.12: Vortrag

Montag, 9. März 2026, 12:30–12:45, HÜL/S186

Discovering Mechanisms and Governing Laws with Sparse Regression — •Gianmarco Ducci, Maryke Kouyate, Juan Manuel Lombardi, Artem Samtsevych, Karsten Reuter, and Christoph Scheurer — FHI Berlin

Interpretable data-driven methods have proven viable for deriving complex vector fields directly from experimental data. Their inherent differential formulation, however, make them vulnerable to noise, which can compromise the sparsity of the inferred models. In order to promote sparsity, a weak formulation can be employed. Then, finding the optimal set of basis functions is a necessary prerequisite, yet a challenging task to determine in advance.

We present the release version of the Data-Driven Model Optimizer ddmo, a symbolic regression tool which provides fine-grained control over the admissible space of candidate terms. Its core contribution lies in the systematic optimization of the library of functions, implemented through two complementary engines: a standard SINDy-based differential formulation and a weak-form variant. Its modular structure further enables the optimization of test functions within the weak formulation. An overview of the software capabilities is provided, alongside with a case study illustrating the reconstruction of effective kinetics from experimental reactor data.

Keywords: Sparse Regression; Regularization; Data-Driven Modelling