MACE

Welcome to the MACE repository!

MACE, a Machine-learning Approach to Chemistry Emulation, by Maes et al. (2024), is a surrogate model for chemical kinetics. It is developed in the contexts of circumstellar envelopes (CSEs) of asymptotic giant branch (AGB) stars, i.e. evolved low-mass stars.

During development, the chemical models of Maes et al. (2023) are used. In this paper you can also find more details about the astrochemical environment used.

MACE is implemented in Python and is trained using PyTorch, together with torchode (Lienen & Gunnemann, 2022).

---

Table of content

• Installation
• What is MACE?
• How to use?
• Example case
• Contributions
• Contact
• Acknowledgements

---

Notes on installation

• MACE is currently not available as a package on pypi. There is a package named mace, but it is not this one.
• To use MACE, please clone the repo and install the required packages, see requirements.txt:

git clone https://github.com/silkemaes/MACE.git

---

What is MACE?

MACE offers a surrogate model that emulates the evolution of chemical abundances over time in a dynamical physical environment. As the name states, it makes use of machine-learning techniques. More specifically, combining an autoencoder (blue) and a trainable ordinary differential equation (ODE) (red) allows to accurately emulate a chemical kinetics model.

Hence, MACE is a framework, an architecture, that can be trained for specific chemical datasets, but before using, should be made compatible with the dataset, see How to use?.

The architecture of MACE is schematically given as

MACE offers a surrogate model that emulates the evolution of chemical abundances over time in a dynamical physical environment. As the name states, it makes use of machine-learning techniques. More specifically, combining an autoencoder (blue) and a trainable ordinary differential equation (ODE) (red) allows to accurately emulate a chemical kinetics model.

In formula, MACE is stated as

$$ {\hat{\boldsymbol{n}}}(t) = \mathcal{D}\Big( G \big( \mathcal{E} ({\boldsymbol{n}}, {\boldsymbol{p}}),t \big) \Big). $$

Here, ${\hat{\boldsymbol{n}}}(t)$ are the predicted chemical abundances at a time $t$ later dan the initial state ${\boldsymbol{n_0}}$. $\mathcal{E}$ and $\mathcal{D}$ represent the autoecoder, with the encoder and decoder, respectively. The autoencoder maps the chemical space ${\boldsymbol{n_0}}$ together with the physical space ${\boldsymbol{p}}$ to a lower dimensional representation $\boldsymbol{z}$, called the latent space. The function $G$ describes the evolution in latent space such that $\boldsymbol{z}(\Delta t) = G(\boldsymbol{z}, \Delta t)=\int_0^{\Delta t} g(\boldsymbol{z}){\rm d}t$.

For more details, check out our paper: Maes et al. (2024).

---

How to use?

The script routine.py gives the flow of training & storing a MACE architecture, and immediately applies to the specified test dataset once training is finished. As such, it returns an averaged error on the MACE model compared to the classical model. More info on the training routine can be found in the paper.

An annotated notebook of the routine can be found in here.

The script routine.py takes an input file with the needed (hyper)parameter setup. An example of such an input file can be found in input/.

python routine.py example

Disclaimer:

In order to train MACE with a certain chemical dataset, the Dataset class should be made compatible with that data. Currently, the script src/mace/CSE_0D/dataset.py works only for the specific dataset used here, i.e. models from Maes et al. (2023), using the Rate22-CSE code.

---

Example case

This repository contains a trained MACE model as a test case, see model/20240604_160152.

The code for loading a trained MACE model can be found in the script src/mace/load.py, testing in src/mace/test.py. An annotated notebook can be found in the documentation.

---

Contact

If any comments or issues come up, please contact me via email, or set up a GitHub issue.

---

Developers & Contributions

Developers:

• Silke Maes
• Frederik De Ceuster

Scientific & technical advisors:

• Marie Van de Sande
• Leen Decin

Contributors:

• Steven Rieder

Feel free to contribute to MACE via pull requests and discussions!

---

Acknowledgements

The MACE architecture is free to use. Please cite our paper Maes et al. (2024).