MACE: A Machine learning Approach to Chemistry Emulation

TL;DR

This paper introduces mace, a machine learning framework that emulates complex chemical processes in astrophysical simulations, significantly reducing computational costs while maintaining accuracy.

Contribution

The paper presents a novel architecture combining autoencoders and latent ODEs to emulate chemistry in 3D astrophysical environments, outperforming classical methods.

Findings

mace outperforms classical chemistry models by a factor of 26

the architecture achieves sub-linear scaling with simulation particles

successfully reproduces chemical pathways in dynamical environments

Abstract

The chemistry of an astrophysical environment is closely coupled to its dynamics, the latter often found to be complex. Hence, to properly model these environments a 3D context is necessary. However, solving chemical kinetics within a 3D hydro simulation is computationally infeasible for a even a modest parameter study. In order to develop a feasible 3D hydro-chemical simulation, the classical chemical approach needs to be replaced by a faster alternative. We present mace, a Machine learning Approach to Chemistry Emulation, as a proof-of-concept work on emulating chemistry in a dynamical environment. Using the context of AGB outflows, we have developed an architecture that combines the use of an autoencoder (to reduce the dimensionality of the chemical network) and a set of latent ordinary differential equations (that are solved to perform the temporal evolution of the reduced features). Training this architecture with an integrated scheme makes it possible to successfully reproduce a full chemical pathway in a dynamical environment. mace outperforms its classical analogue on average by a factor 26. Furthermore, its efficient implementation in PyTorch results in a sub-linear scaling with respect to the number of hydrodynamical simulation particles.