Accelerating Radiative Transfer for Planetary Atmospheres by Orders of Magnitude with a Transformer-Based Machine Learning Model

TL;DR

This work tackles the computational bottleneck of radiative transfer in exoplanetary atmospheres by training an encoder-only transformer to emulate bolometric layer fluxes from 1D pressure–temperature profiles. Training data are generated by parameterizing radiative equilibrium for hot-Jupiter-like atmospheres and solving for fluxes with the PICASO forward model, achieving about 1% mean absolute error on unseen profiles. The transformer delivers ~100x speedups in inference compared with traditional radiative transfer and captures non-local atmospheric dependencies, enabling integration into GCMs and faster model exploration. The study also highlights limitations related to the restricted training space (single host star, solar metallicity) and outlines clear future work to extend parameter coverage (clouds, varied chemistry, different stellar types) and to incorporate the emulator into broader climate modeling workflows.

Abstract

Radiative transfer calculations are essential for modeling planetary atmospheres. However, standard methods are computationally demanding and impose accuracy-speed trade-offs. High computational costs force numerical simplifications in large models (e.g., General Circulation Models) that degrade the accuracy of the simulation. Radiative transfer calculations are an ideal candidate for machine learning emulation: fundamentally, it is a well-defined physical mapping from a static atmospheric profile to the resulting fluxes, and high-fidelity training data can be created from first principles calculations. We developed a radiative transfer emulator using an encoder-only transformer neural network architecture, trained on 1D profiles representative of solar-composition hot Jupiter atmospheres. Our emulator reproduced bolometric two-stream layer fluxes with mean test set errors of ~1% compared to the traditional method and achieved speedups of 100x. Emulating radiative transfer with machine learning opens up the possibility for faster and more accurate routines within planetary atmospheric models such as GCMs.

Figures (7)

• Figure 1: The overall structure of the machine learning emulator pipeline presented in this work. First, 1D profiles are created based on randomly sampled parameterizations. Next, these profiles are processed with PICASO in order to solve the radiative transfer equations and determine the net thermal and starlight layer fluxes. Finally, these data are used to train an encoder-only transformer model that can predict radiative transfer fluxes based on an input pressure-temperature profile.
• Figure 2: Example 1D pressure-temperature profiles created using the parameterized relations from line2013. These profiles were generated to be representative of a hot Jupiter atmosphere. We prioritize wide coverage of the parameter space, at the cost of simulating some unphysical atmospheres. The entire training set is composed of 2,000,000 such 1D profiles.
• Figure 3: A diagram showing how data flows through the transformer architecture implemented here. Data padding and masking are left out for simplicity. The model does implement this functionality, but it is not necessary for our constant-length sequences.
• Figure 4: Left: training loss vs. validation loss during model runtime. Right: the learning rate of the model. Near epoch 200, the model began overfitting, and the validation loss plateaued while the training loss continued to decrease, indicating moderate overfitting.
• Figure 5: The layer net fluxes (black) output from PICASO, compared to the values predicted by the transformer emulator (dashed red). Thermal fluxes and associated errors are shown in the top row, and scattered starlight fluxes and associated errors are shown in the bottom row. All profiles shown are drawn from the 15% test set, that the model did not train on. Overall, we found that the emulator is within approximately 1% of the true value.