Genetic algorithm trims chemical reaction lists for hot Jupiter atmosphere models

The work

Technical · from the record

The authors built DARWEN, a genetic algorithm that reduces chemical reaction networks used in 1D models of hot Jupiter atmospheres. Starting from a scheme obtained by a local sensitivity analysis and principal component analysis of a 108-species, 1906-reaction network, the algorithm evolved populations of candidate schemes in which each gene switches a reaction on or off, scored by a loss function combining abundance discrepancy against the full network with the number of species. Three schemes were produced for the atmospheres of HD 209458b and HD 189733b: a validation scheme with 47 species and 576 reactions, a low-cost scheme that ran 2.5 times faster than the previously published R20 reduced network and 10 times faster than the full network, and a scheme including 34 photochemical reactions that reached key-molecule deviations of 0.16 and 0.18 with run times of 33 and 43 seconds.

How AI was used

A genetic algorithm was used to search the space of possible reduced chemical networks. Each candidate scheme was encoded as a binary string whose genes mark whether a forward-and-reverse reaction pair is included, giving 958 genes for the full V20 network; the search pool was restricted to reactions tied to the included species and a few additional molecules. The initial population began from a single progenitor derived from a PCA-based reduction of local sensitivity coefficients, which were obtained by re-running the 1D chemical kinetics model to steady state with each rate constant multiplied by 1.1. At each generation, selected progenitors underwent one-point crossover, self-mutation, and mutation and killing rates that add or remove reactions, with an elitism step reintroducing earlier progenitors and a curiosity mechanism rewarding under-explored reactions. Each individual's fitness was computed by running the 1D model and combining the maximum relative deviation from the full network for key species and for major species with a weighted term for the number of molecules, taking the worst value across the two planets when optimising both atmospheres at once. The validation scheme was evolved over 20 generations with a bottleneck step that forced the best scheme to seed a subsequent run; the low-cost scheme was then evolved from it. For the photochemistry case, the algorithm did not alter photochemical reactions directly; the 1D model included them according to which species were present.