Machine-learning inference of the interior structure of low-mass exoplanets

Philipp Baumeister, Sebastiano Padovan, Nicola Tosi, Grégoire Montavon, Nadine Nettelmann, Jasmine MacKenzie, M. Godolt2020 · DOI

TL;DR

This paper introduces a machine learning approach using Mixture Density Networks (MDNs) to infer the internal structure of low-mass exoplanets. By training on a large synthetic dataset, the MDN can rapidly predict the distribution of layer thicknesses (core, mantle, ice, atmosphere) from just mass and radius, significantly speeding up characterization compared to traditional methods. Including the fluid Love number ($k_2$) as an additional input further reduces the ambiguity in these predictions.

What Problem Does This Paper Address?

Characterizing the internal structure of exoplanets is a crucial goal in planetary science. With the rapid discovery of new exoplanets, efficient methods are needed to understand their composition and structure. Traditionally, this involves complex forward modeling and inverse problem-solving, often requiring extensive computational resources and leading to degenerate solutions (multiple interior structures fitting the same observational data). This is particularly true when relying solely on mass and radius, the most common observational constraints. The paper aims to address this by developing a machine learning model that can quickly and accurately infer the probability distribution of interior structures.

Methods

The core of this work is the application of Mixture Density Networks (MDNs), a type of neural network that predicts the parameters of a Gaussian mixture distribution rather than a single output value. This allows the MDN to represent the full probability distribution of possible interior structures for a given input.

The authors built a TATOOINE (Tool for ATmospheres, Outgassing and Optimal INteriors of Exoplanets) model to generate synthetic exoplanets. These planets consist of four distinct layers: an iron-rich core, a silicate mantle, a high-pressure ice shell, and a H/He envelope. The model calculates radial profiles of mass, density, and pressure based on coupled equations for hydrostatic equilibrium and an equation of state.

The MDN was trained on a dataset of approximately 900,000 synthetic planets. The network architecture consists of a feedforward neural network with three hidden layers, each activated by a Rectified Linear Unit (ReLU). The output layer is specialized for MDNs, predicting the means, variances, and weights of Gaussian mixtures for each of the four planetary layers. The loss function used is the average negative log-likelihood, and the Adam optimizer is employed for training.

Two input scenarios were tested:

  1. Case 1: Planetary mass and radius.
  2. Case 2: Planetary mass, radius, and the fluid Love number ($k_2$).

Datasets

The primary dataset used in this study is a newly generated synthetic dataset of 900,000 exoplanets. These synthetic planets were created using the TATOOINE interior structure model, covering a mass range of 0.01 to 25 Earth masses. The dataset includes information on the mass fractions and radius fractions of the four defined interior layers (core, mantle, ice, gas envelope), as well as the fluid Love number ($k_2$).

The trained MDN models are made publicly available through a GitHub repository, allowing the exoplanetary science community to use them.

Key Insights and Ideas

  • MDNs excel at inverse problems: The ability of MDNs to predict probability distributions, rather than single values, makes them well-suited for inverse problems like inferring interior structure from limited observations. This is a significant advantage over standard neural networks.
  • $k_2$ is a powerful constraint: Including the fluid Love number ($k_2$) as an input dramatically reduces the degeneracy of possible interior structures. This observable provides crucial information about the internal mass distribution, which is otherwise hard to disentangle from mass and radius alone.
  • Rapid inference: A trained MDN can infer interior structure distributions in milliseconds, a massive speedup compared to traditional forward modeling and Monte Carlo sampling which can take hours. This enables rapid characterization of newly discovered exoplanets.
  • Layered structure assumptions matter: The accuracy of the MDN's predictions is inherently tied to the assumptions made in the underlying interior structure model used for training. Simplifications, like a zero-temperature H/He envelope or neglecting certain ice phases, can lead to systematic biases in the predicted layer thicknesses.
  • Prior distributions influence results: While MDNs learn from data, the distribution of that data (the prior) can still influence the predicted posterior distributions. However, the paper suggests that as long as sufficient samples exist across the parameter space, the exact shape of the prior is less critical.

Key Results

  • Improved accuracy with $k_2$: The MDN trained with mass, radius, and $k_2$ (Case 2) shows significantly better predictions for all interior layers compared to the MDN trained only on mass and radius (Case 1). The improvement is particularly notable for the mantle layer.
  • Gas envelope well-constrained: Even with only mass and radius, the MDN can infer the relative thickness of the gaseous envelope with good accuracy, consistent with previous findings that planetary radius is a proxy for envelope mass.
  • Core and mantle prediction challenges: The MDN tends to overestimate core thickness for small cores and underestimate large layer thicknesses due to undersampling in the training data. The silicate mantle is the least well-constrained layer, showing broad distributions.
  • Validation on representative planets: The MDN successfully replicates the distributions of possible interior solutions for Earth, Neptune-like planets, Kepler-10b, and GJ 1214b, aligning well with independent Monte Carlo sampling results.
  • Computational efficiency: Training the MDN takes less than 30 minutes on a GPU, and inference is on the order of milliseconds.

Limitations and Future Work

  • Interior model simplifications: The authors acknowledge that their interior model has simplifications, such as a zero-temperature H/He envelope and neglecting liquid/super-ionic water phases. These can lead to overestimations of envelope thickness and underestimations of $k_2$. Future work could involve using more sophisticated, temperature-dependent equations of state and more realistic ice phase models.
  • Prior distribution influence: While the impact of prior distributions was explored, the paper notes that the exact shape of the prior can still influence predictions, especially for under-sampled regions of the parameter space.
  • Generalizability to other models: The current MDN is trained on a specific interior structure model. Future work could involve training MDNs on different interior models or incorporating more complex compositional gradients for gas-rich planets.
  • Incorporating more observables: The paper suggests that as new observational data (e.g., atmospheric compositions from JWST) become available, more complex MDNs trained on larger datasets and incorporating additional observables will be required for even more accurate characterization.

Summary updated

Abstract

<p>We explore the application of machine-learning, based on mixture density neural networks (MDNs), to the interior characterization of low-mass exoplanets up to 25 Earth masses constrained by mass, radius, and fluid Love number k<sub>2</sub>. MDNs are a special subset of neural networks, able to predict the parameters of a Gaussian mixture distribution instead of single output values, which enables them to learn and approximate probability distributions. With a dataset of 900,000 synthetic planets, consisting of an iron-rich core, a silicate mantle, a high-pressure ice shell, and a gaseous H/He envelope, we train an MDN using planetary mass and radius as inputs to the network. We show that the MDN is able to infer the distribution of possible thicknesses of each planetary layer from mass and radius of the planet. This approach obviates the time-consuming task of calculating such distributions with a dedicated set of forward models for each individual planet.</p><p>The fluid Love number k<sub>2</sub> bears constraints on the mass distribution in the planets' interior and will be measured for an increasing number of exoplanets in the future. Adding k<sub>2</sub> as an input to the MDN significantly decreases the degeneracy of possible interior structures.</p>
BibTeX
@misc{baumeister2020machine,
  author = {Philipp Baumeister and Sebastiano Padovan and Nicola Tosi and Grégoire Montavon and Nadine Nettelmann and Jasmine MacKenzie and M. Godolt},
  title = {{Machine-learning inference of the interior structure of low-mass exoplanets}},
  year = {2020},
  doi = {10.5194/egusphere-egu2020-7520},
}
Figure