DARWEN: Data-driven Algorithm for Reduction of Wide Exoplanetary Networks

An unbiased approach to accurately reducing chemical networks

¹ Department of Chemistry, KU Leuven, Celestijnenlaan 200F, 3001 Leuven, Belgium
² Institute of Astronomy, KU Leuven, Celestijnenlaan 200D, 3001 Leuven, Belgium
³ Anton Pannekoek Institute for Astronomy, University of Amsterdam, Science Park 904, 1098 XH, Amsterdam, The Netherlands
⁴ Millennium Institute Foundational Research on Data, Vicuña Mackenna 4860, Macul, Santiago, Chile
⁵ Université Paris Cité and Univ Paris Est Creteil, CNRS, LISA, 75013 Paris, France

^★ Corresponding author; arturo.lira@kuleuven.be

Received: 30 August 2024
Accepted: 18 November 2024

Abstract

Context. Exoplanet atmospheric modeling is advancing toward complex coupled circulation-chemistry models, from chemically diverse 1D models to 3D global circulation models (GCMs). These models are crucial for interpreting observations from facilities like JWST and ELT and understanding exoplanet atmospheres. However, maintaining chemical diversity in 1D models and especially in GCMs is computationally expensive, limiting their complexity. Optimizing the number of reactions and species in the simulated atmosphere can address this tradeoff, but there is a lack of transparent and efficient methods for this optimization in the current exoplanet literature.

Aims. We aim to develop a systematic approach for reducing chemical networks in exoplanetary atmospheres, balancing accuracy and computational efficiency. Our method is data-driven, meaning we do not manually add reactions or species. Instead, we test possible reduced chemical networks and select the optimal one based on metrics for accuracy and computational efficiency. Our approach can optimize a network for similar planets simultaneously, can assign weights to prioritize either accuracy or efficiency, and is applicable in the presence of photochemistry.

Methods. We propose an approach based on a sensitivity analysis of a typical 1D chemical kinetics model. Principal component analysis was applied to the obtained sensitivities. To achieve a fast and reliable reduction of chemical networks, we utilized a genetic algorithm (GA), a machine-learning optimization method that mimics natural selection to find solutions by evolving a population of candidate solutions.

Results. We present three distinct schemes tailored for different priorities: accuracy, computational efficiency, and adaptability to photochemistry. These schemes demonstrate improved performance and reduced computational costs. Our work represents the first reduction of a chemical network with photochemistry in exoplanet research.

Conclusions. Our GA-based method offers a versatile and efficient approach to reduce chemical networks in exoplanetary atmospheres, enhancing both accuracy and computational efficiency.