Chemulator: Fast, accurate thermochemistry for dynamical models through emulation

TL;DR

Chemulator is a neural network-based emulator that provides fast and accurate thermochemical predictions in dynamical models, significantly reducing computational costs while maintaining acceptable accuracy.

Contribution

This work introduces Chemulator, a neural network emulator trained on UCLCHEM outputs, enabling rapid thermochemical calculations with minimal accuracy loss in dynamical simulations.

Findings

Chemulator achieves an MSE of 0.0002 for temperature predictions.

It is approximately 50,000 times faster than the full chemical model.

The emulator remains stable over 1000 iterations with low error accumulation.

Abstract

Chemical modelling serves two purposes in dynamical models: accounting for the effect of microphysics on the dynamics and providing observable signatures. Ideally, the former must be done as part of the hydrodynamic simulation but this comes with a prohibitive computational cost which leads to many simplifications being used in practice. To produce a statistical emulator that replicates a full chemical model capable of solving the temperature and abundances of a gas through time. This emulator should suffer only a minor loss of accuracy over including a full chemical solver in a dynamical model but would have a fraction of the computational cost. The gas-grain chemical code UCLCHEM was updated to include heating and cooling processes and a large dataset of model outputs from possible starting conditions was produced. A neural network was then trained to map directly from inputs to outputs. Chemulator replicates the outputs of UCLCHEM with an overall mean squared error (MSE) of 0.0002 for a single time step of 1000 yr and is shown to be stable over 1000 iterations with an MSE of 0.003 the log scaled temperature after one time step and 0.006 after 1000 time steps. Chemulator was found to be approximately 50,000 times faster than the time dependent model it emulates but can introduce a significant error to some models.