Exoplanet Formation Inference Using Conditional Invertible Neural Networks¶

Conference: NeurIPS 2025 arXiv: 2512.05751 Code: None Area: Physics / Planetary Science Keywords: exoplanets, conditional invertible neural networks, Bayesian inference, planet formation, surrogate model

TL;DR¶

A conditional invertible neural network (cINN) trained on 15,777 synthetic planets infers planet formation parameters (disk mass, turbulent \(\alpha\), dust-to-gas ratio) from observables (planet mass, orbital distance), achieving probabilistic parameter retrieval ~10⁶× faster than physical simulations. Multi-planet system data is shown to yield more robust inference than single-planet data.

Background & Motivation¶

Background: Understanding the origins of exoplanets requires tracing formation parameters from observed planetary properties. Direct MCMC approaches are infeasible — a single run of a global planet formation model takes days to months.

Limitations of Prior Work: (a) Physical models are computationally prohibitive for large-scale Bayesian inference; (b) gravitational chaos among planets renders the parameter-to-observable mapping stochastic; (c) data are sparse in the high-dimensional parameter space (disk mass, viscosity, dust-to-gas ratio, inner edge, etc.).

Key Challenge: Accurate probabilistic inference is required, yet physical models cannot be run at scale.

Goal: Train a fast surrogate model on a limited synthetic dataset to enable practical Bayesian inference of planet formation parameters.

Key Insight: cINNs provide exact invertible mappings — the forward pass maps parameters to a standard Gaussian latent space conditioned on observables, while the inverse pass samples the posterior, naturally supporting probabilistic inference.

Core Idea: Use a cINN as a surrogate for the planet formation physical model. Individual planets extracted from multi-planet systems are treated as separate training samples to increase data diversity, enabling millisecond-scale probabilistic parameter inference.

Method¶

Overall Architecture¶

A physical model (dust-to-planet global formation model) generates synthetic planetary data → a cINN is trained to learn an invertible mapping from parameters to latent space conditioned on observables → at inference time, samples drawn from a standard Gaussian are passed through the inverse flow to obtain the posterior distribution.

Key Designs¶

Global Planet Formation Model for Data Generation:
- Function: Generate synthetic planet samples covering the parameter space.
- Mechanism: Tracks the full pipeline of dust grain coagulation, planetesimal formation, protoplanet accretion, and photoevaporation of the gas disk. Comprises 707 single-planet disks and ~15,777 multi-planet systems (up to 100 planets per disk). Four parameters varied in log-space: disk mass fraction (\(10^{-3}\) to \(10^{-0.5}\)), viscosity \(\alpha\) (\(10^{-3.5}\) to \(10^{-2}\)), dust-to-gas ratio (\(10^{-2.4}\) to \(10^{-1}\)), and inner-edge orbital period (1–20 days).
- Design Motivation: Extracting one sample per planet from multi-planet systems provides ~22× more diverse parameter–observable combinations.
cINN Architecture:
- Function: Learn an invertible mapping from parameters to latent space.
- Mechanism: 16 affine coupling blocks with random permutations between blocks; each subnetwork consists of 3 layers × 8 units with ReLU activations. Maps 4D parameters \(\vec{x}\) to a 4D latent space \(\vec{z}\) (unit Gaussian) conditioned on 2D observables \(\vec{c}\) (planet mass, semi-major axis). Loss: \(L = \frac{1}{2}\|f(x;c)\|^2 - \log|\det \frac{\partial f}{\partial x}| + \|\hat{x}-x\|^2\)
- Design Motivation: The invertible mapping guarantees exact posterior sampling without additional MCMC steps.
Multi-Planet vs. Single-Planet Training Strategy:
- Function: Compare the effect of different data organization strategies on inference robustness.
- Mechanism: Multi-planet training extracts individual (parameters, observables) pairs from each planet, increasing training sample diversity. Single-planet training uses only 707 simulations.
- Design Motivation: Single-planet training produces unphysical extrapolations in unsampled regions (overestimated \(\alpha\) at large orbital distances); multi-planet training is more robust.

Loss & Training¶

Combined loss: maximum likelihood (negative log-likelihood under a Gaussian latent space) + reconstruction loss. Adam optimizer (\(\beta_1=\beta_2=0.8\), lr=0.001, decay \(\gamma=0.99\)/epoch); data augmentation with Gaussian noise (\(\sigma=0.01\)).

Key Experimental Results¶

Main Results¶

Training Data	MAP Deviation (σ)	Parameter Space Coverage	Extrapolation Quality
Multi-planet (~15.7k)	0.2 (good)	Excellent	Physically consistent
Single-planet (707)	0.2 (sampled region)	Poor	Unphysical extrapolation
Two-planet systems	Stable	Improved	Good generalization

Parameter Inference Analysis¶

Parameter	Inference Quality	Correlation Pattern
Disk mass \(M_{disk}\)	Good, narrow posterior	Positively correlated with \(\alpha\)
Viscosity \(\alpha\)	Diagonal pattern in mass–distance space	Affects migration and dust properties
Dust-to-gas ratio	Good	Relatively independent
Inner-edge period	Recoverable	Weakly constrained

Key Findings¶

Multi-planet data is essential: Single-planet training yields spuriously narrow posteriors (overconfidence) in unsampled regions; multi-planet training eliminates this artifact.
Chaos does not impede inference: Gravitational chaos among planets does not degrade parameter recovery — chaotic effects are orthogonal to the formation parameter imprint.
~10⁶× inference speedup: Millisecond-scale inference vs. month-scale physical model computation.
Physical causality reflected in \(\alpha\) inference: The diagonal pattern of \(\alpha\) in distance–mass space reflects the dual influence of viscosity on migration and dust evolution.

Highlights & Insights¶

Data diversity > data volume: Extracting individual planets from multi-planet systems provides more uniform parameter space coverage and is more effective than increasing the number of single-planet simulations.
cINNs are naturally suited for parameter inference: Invertibility guarantees exact posterior sampling without MCMC or variational approximations.
Chaos robustness is surprising: Gravitational N-body chaos might be expected to destroy a deterministic parameter-to-observable mapping, yet the posterior remains recoverable.

Limitations & Future Work¶

Data volume remains insufficient for higher-dimensional settings (6D observables, three-planet systems).
Strong dependence on the accuracy of the physical model — violations of model assumptions necessitate retraining.
Single-planet training is unsuitable for real survey data, limiting applicability in the simplest observational scenario.

vs. MCMC / nested sampling: cINNs provide instantaneous posterior sampling, whereas MCMC requires thousands of model evaluations.
vs. Simulation-Based Inference (SBI): cINNs represent one implementation of SBI, with the advantage of exact invertibility.

Rating¶

Novelty: ⭐⭐⭐⭐ First application of cINNs to planet formation parameter inference; the multi-planet data strategy is novel.
Experimental Thoroughness: ⭐⭐⭐ Synthetic data evaluation is comprehensive, but validation on real observational data is absent.
Writing Quality: ⭐⭐⭐⭐ Physical motivation is clear; connections between methodology and physics are well articulated.
Value: ⭐⭐⭐⭐ Provides a practical inference tool for exoplanet population demographics.