Berlin 2024 – wissenschaftliches Programm

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

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 2: Machine Learning & Physics

AKPIK 2.3: Vortrag

Mittwoch, 20. März 2024, 15:30–15:45, MAR 0.002

Combining genetic algorithm and compressed sensing for features and operators selection in symbolic regression — •Aliaksei Mazheika¹, Sergey V. Levchenko², and Luca M. Ghiringhelli^3,4 — ¹Technische Universitaet Berlin, DE — ²Moscow, RU — ³The NOMAD Laboratory at the Fritz Haber Institute and Humboldt University, Berlin, DE — ⁴Friedrich-Alexander University, Erlangen, DE

The symbolic inference method SISSO (Sure-Independence Screening and Sparsifying Operator) has recently found a broad application in materials science. It performs regression or classification by adopting compressed sensing for the selection of an optimized subset of features and mathematical operators out of a given set of candidates. However, SISSO becomes computationally unpractical when the set of candidate features and operators exceeds the size of few tens. Here we combine SISSO with a genetic algorithm (GA) for the global search of the optimal subset of features and operators. We test GA-SISSO for the search of predictive models of perovskites lattice parameters, and demonstrate that our method efficiently finds more accurate models than the original SISSO. GA-SISSO was also applied for the search of the model for prediction of CO₂ adsorption energies on semiconductor oxides. The model learned by GA-SISSO has much higher accuracy compared to previously discussed models based on the O 2p-band center. The statistical analysis of contributions of all features to the learned models shows that, besides the O 2p-band center, the electrostatic potential above adsorption sites and surface formation energies are key features.

Keywords: symbolic regression; perovskites; CO2; adsorption