This repository contains the code for using a conditional neural field as an emulator for instellar chemistry.
This repository makes use of PyTorch and other Python packages. You can have them installed in your system, but we recommend to follow these instructions to have a working environment with everything needed for running the code.
Install Miniconda in your system. Go to https://docs.conda.io/projects/miniconda/en/latest/ and download the executable file for your system. For instance, for a typical Linux system, you should do:
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
or
curl -lO https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
Now install Miniconda
chmod +x Miniconda3-latest-Linux-x86_64.sh
./Miniconda3-latest-Linux-x86_64.sh
and follow the instructions for selecting the directory.
Now create an environment and install the packages:
conda create -n chemistry python=3.10
conda activate chemistry
conda install -c conda-forge numpy tqdm
Now install PyTorch. Go to https://pytorch.org/ and select from the properties of your system from the matrix (Linux/Mac/Windows, CPU/GPU, conda/pip). If you have an NVIDIA GPU card in your system, you can take advantage of the acceleration that it provides (around a factor 10 faster). For instance, in a system with an NVIDIA GPU card with CUDA drivers on version 12.1, you can install PyTorch using:
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
or
pip3 install torch torchvision torchaudio
if using pip for the installation of packages.
In a CPU-only system:
conda install pytorch torchvision torchaudio cpuonly -c pytorch
or
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
if using pip for the installation of packages.
An example of how to run the code is shown in example.py, that we reproduce here:
# Define the number of random models
n_models = 64
batch_size = 32
nh = 10.0**np.random.uniform(np.log10(1e4), np.log10(1e7), size=n_models)
T = np.random.uniform(10.0, 80.0, size=n_models)
crir = 10.0**np.random.uniform(np.log10(1e-17), np.log10(1e-15), size=n_models)
sulfur = 10.0**np.random.uniform(np.log10(1.5e-5), np.log10(7.5e-8), size=n_models)
uv_flux = 10.0**np.random.uniform(np.log10(0.1), np.log10(1e4), size=n_models)
t = np.logspace(0, 7, 64)
net = emulator.ChemistryEmulator(gpu=0, batch_size=32)
abundance = net.evaluate(t, T, nh, crir, sulfur, uv_flux, batch_size=32, species=None)
We choose 64 models with random properties inside the range of validity. Select the output times and call the emulator. The batch_size defines the number of models that will be computed in parallel. This depends on the amount of memory you have but it can be a large number. The keyword species indicates a list of the indices of which species you
want to compute, from the list of 192 species that can be found in list_molecules.txt. If abstent or
set to None, all species are computed.
You can download the weights of the model and the training/validation data from
https://cloud.iac.es/index.php/s/eZAT4ocJFkPyn3R
Put the weight files in the working directory and run the code.