geeXHP is a Python package for generating synthetic reflected-light spectra of Earth-like exoplanets and preparing machine-learning datasets for atmospheric and planetary retrieval studies.
This repository supports the work presented in:
Barbosa, S. G. A., Estrela, R., da Silva Filho, P. C. F., Mugnai, L. V., & de Freitas, D. B. (2026). Towards the Habitable Worlds Observatory: Retrieval of Reflection Spectra from Evolving Earth Analogues using 1D CNNs. RASTI preprint.
The package was developed to generate mission-relevant synthetic spectra for future direct-imaging studies of terrestrial exoplanets, especially in the context of the Habitable Worlds Observatory (HWO). The workflow combines physically motivated Earth-analogue atmospheric scenarios with NASA's Planetary Spectrum Generator (PSG) and machine-learning utilities for 1D CNN retrieval experiments.
The paper uses geeXHP to build a dataset of 108,246 synthetic reflected-light spectra spanning Archean, Proterozoic, and Modern Earth atmospheres under LUVOIR-B and HabEx/SS noise models. The corresponding CNN framework retrieves six atmospheric mixing ratios (CH4, CO2, H2O, N2, O2, O3) and four planetary parameters: radius, surface gravity, surface temperature, and surface pressure.
- Generate synthetic reflected-light spectra with PSG.
- Sample Earth-like planets across Archean, Proterozoic, and Modern atmospheric compositions.
- Apply habitability and atmospheric-retention filters for physically plausible planet configurations.
- Produce spectra for LUVOIR-B and HabEx/SS instrument configurations.
- Prepare datasets for CNN-based atmospheric and planetary retrieval.
- Convert generated data into machine-learning-ready formats.
- Visualize noiseless and noisy spectra across instrument channels.
Clone the repository and install it in editable mode:
git clone https://github.com/SarahBarbosa/geexhp.git
cd geexhp
git lfs pull
pip install -e .This repository uses Git LFS for large data/model artifacts used by the desktop app and examples, including the test TFRecord and trained Keras
models. If git lfs pull is not available, install Git LFS first and run git lfs install.
For development tools:
pip install -e .[dev]geeXHP requires access to PSG. For large-scale generation, the paper used a locally hosted PSG instance running inside a Docker container.
from geexhp import datagen
dg = datagen.DataGen(stage="modern")
dg.generator(
start=0,
end=8,
random_atm=False,
output_file="modern_0-8",
instruments="LUVOIR",
)This saves the generated dataset to data/modern_0-8.parquet.
To generate spectra for other geological stages:
from geexhp import datagen
proterozoic = datagen.DataGen(stage="proterozoic")
archean = datagen.DataGen(stage="archean")Instrument options include:
"all": all available instrument channels."LUVOIR": LUVOIR-B channels."SS": HabEx/SS channels.- Specific channels such as
"B-UV","B-Vis","B-NIR","SS-UV","SS-Vis", and"SS-NIR".
This repository includes an offline PySide6 desktop app for exploring the trained LUVOIR-B and HabEx/SS retrieval models, inspecting reflected-light spectra, visualizing Integrated-Gradients sensitivity maps, comparing retrievals, and running custom pasted spectra through the models. The app also includes an interactive neural-network walkthrough that shows a selected spectrum moving through normalization, convolutional blocks, attention, latent embeddings, and the final retrieval heads, with hover/click inspection of live tensors.
The app lives in desktop_app/. For installation, double-click launcher, environment setup, Git LFS requirements, and usage details, see desktop_app/README.md.
Additional usage notes are available in docs/:
docs/1-psg_installation_steps.rst: PSG installation notes.docs/2-how_to_use_geexhp.rst: data-generation and visualization guide.
The trained CNN models for the HabEx/SS and LUVOIR-B configurations, together with the corresponding training, validation, and test datasets, are available on Zenodo: https://doi.org/10.5281/zenodo.15648637
We gratefully acknowledge the financial support from the Brazilian agency CAPES (grant No. 88887.622098/2021-00), as well as the STELLAR TEAM at the Federal University of Ceará for our discussions and collaborative support. Part of the research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. Special thanks to Geronimo Villanueva for his assistance with setting up the PSG, and to Yui Kawashima for providing data on the Proterozoic Earth.
This project is distributed under the BSD 2-Clause License.