ML4GASTLI — A Neural-Network Surrogate for GASTLI

A fast deep-learning emulator of the GASTLI gas-giant interior model: predict a planet's radius, entropy and thermal state from its bulk parameters in milliseconds, instead of running the full physical model.

GASTLI (the GAS gianT modeL for Interiors, Acuña et al. 2024) computes interior-structure models for gas giants — mass–radius relations, thermal evolution and composition retrievals — by solving the coupled interior/atmosphere physics. That accuracy is expensive: large parameter sweeps and retrievals require many thousands of evaluations.

ML4GASTLI is a Deep Residual Neural Network (ResNet) trained on GASTLI outputs that reproduces its key predictions almost instantly, so it can stand in for the full model wherever speed matters (e.g. MCMC retrievals, population synthesis, interactive exploration).

Authors

Developed by Kunal Bhatia at the Max Planck Institute for Astronomy (MPIA), Heidelberg — ran the GASTLI simulations, generated the training set, and built and trained the surrogate network.

Acknowledgments

Conceived and supervised by Lorena Acuña (MPIA), author of the GASTLI model that this surrogate emulates and that the training data was generated from. Please cite GASTLI if you use this work (see Citing).

What it predicts

Given six planetary parameters, the model returns three outputs:

• Inputs: mass, core mass fraction, envelope metallicity, core water fraction, surface temperature, surface pressure
• Outputs: planetary radius, specific entropy, thermal parameter (f_s)

Parameter ranges (training domain)

Parameter	Range	Unit
mass_MEarth	0.1 – 600	Earth masses
CMF	0.0 – 1.0	fraction
Zenv	0.0 – 1.0	fraction
Zwater_core	0.0 – 0.5	fraction
Tsurf_K	700 – 6000	Kelvin
Psurf_bar	1 – 1000	bar

Inputs outside this domain extrapolate and may be unreliable.

Installation

conda create -n ml4gastli python=3.10 -y
conda activate ml4gastli
pip install -r requirements.txt

The trained model and scalers are included in models/ (final_model.keras, x_scaler.joblib, y_scaler.joblib), so no extra downloads are needed.

Usage

Edit the parameters at the top of predict.py:

INPUTS = {
    'mass_MEarth': 100.0,
    'CMF': 0.10,
    'Zenv': 0.10,
    'Zwater_core': 0.0,
    'Tsurf_K': 1000.0,
    'Psurf_bar': 100.0,
}

Then run:

python predict.py

Output

The model returns:

• radius_Rearth — planetary radius in Earth radii
• entropy_SI — specific entropy in J kg⁻¹ K⁻¹
• f_s_SI — thermal parameter in J K⁻¹

Internally the network predicts log10(radius) and a signed-log transform of f_s; gastli_core.py converts these back to physical units automatically.

Repository layout

ML4GASTLI/
├── predict.py        # entry point: set INPUTS, run a prediction
├── gastli_core.py    # GastliModel: loads the network + scalers, runs inference
├── models/           # trained ResNet (.keras) + input/output scalers (.joblib)
├── requirements.txt
├── LICENSE
└── README.md

Citing

If you use this surrogate, please cite the underlying GASTLI model:

Acuña, L., et al. (2024). GASTLI: An open-source coupled interior–atmosphere model to unveil gas giant composition. Astronomy & Astrophysics, 688, A60. doi:10.1051/0004-6361/202450559

• GASTLI code: https://github.com/lorenaacuna/GASTLI
• GASTLI docs: https://gastli.readthedocs.io

License

Released under the BSD 3-Clause License, matching GASTLI. © 2025 Kunal Bhatia, Max Planck Institute for Astronomy (MPIA).