TNFlow: Amortized Posterior Inference for Trans-Neptunian Object Surface Composition

Agastya Gaur; Cristina M. Dalle Ore; Alessandra Ricca

astro-ph.IMSep 3, 2026

TNFlow: Amortized Posterior Inference for Trans-Neptunian Object Surface Composition

Authors: Agastya Gaur, Cristina M. Dalle Ore, Alessandra Ricca

Organizations: University of Illinois Urbana-Champaign · SETI Institute · Department of Astronomy, University of Illinois Urbana-Champaign, Urbana, IL 61801 · SETI Institute, 339 Bernardo Avenue, Suite 200, Mountain View, CA 94043 ·

Abstract

We present TNFlow, a transformer and normalizing flow architecture for inferring the surface composition of Trans-Neptunian Objects (TNOs) from their reflectance spectra. TNFlow is trained on synthetic spectra generated by the Shkuratov radiative transfer model to act as its inverse. TNFlow takes ${\sim}$ 0.7s to invert one spectrum on a single CPU core, returning a multimodal posterior over simplex-valid compositions and grain sizes. On synthetic spectra, the highest-weight mode achieves a mean total-variation distance of 0.149 from ground truth on the test split, and the model generalizes well to unseen combinations of known components. Qualitative tests on real JWST spectra show blindness or bias towards some materials. We suggest this could be attributed to either simulator fidelity or the training set.

Figures & tables

Metric	Test ( $n{=}5{,}000$ )	OOD ( $n{=}374$ )
Top-mode comp. TV	$0.149\pm 0.125$	$0.156\pm 0.091$
Best-mode comp. TV	$0.121\pm 0.098$	$0.132\pm 0.080$
Top-mode grain (dex)	$0.242\pm 0.212$	$0.279\pm 0.162$
Best-mode grain (dex)	$0.253\pm 0.201$	$0.283\pm 0.161$
Top mode = best (%)	67.7	60.4
Mean mode count	2.50	2.64

Table 1: Per-solution accuracy on the test and group-holdout OOD splits (500 posterior draws per query, joint composition + grain BIC-GMM mode search). Top mode is the highest-weight mode; best mode is the mode closest to the target composition in total-variation distance. All

\pm

values are one standard deviation.

Figure 1: Probability integral transform (PIT) histograms for composition (top) and grain size (bottom), on the test (left) and OOD (right) splits. A uniform histogram (dashed) indicates calibration. Composition intervals are biased in both splits, with a right-skewed ramp indicating systematic underprediction that worsens OOD; grain intervals are well calibrated throughout.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Formula	Bestfit	Gen.	Total	OC Reference
Amorphous carbon (AC)	C	72,991	107,795	180,786	Rouleau and Martin [1991]
Titan tholin	—	57,841	106,883	164,724	Khare et al. [1984]
Olivine	(Mg,Fe) 2 SiO 4	43,140	107,526	150,666	Dorschner et al. [1995]
Triton tholin	—	43,130	95,504	138,634	Khare et al. [1994]
H 2 O	H 2 O	32,928	103,204	136,132	Mastrapa et al. [2008] , Mastrapa et al. [2009]
CH 3 OH (methanol)	CH 3 OH	30,715	96,553	127,268	Hudgins et al. [1993]

Appendix

Table 2: Component library (

K{=}20

) and per-component spectrum counts. “Bestfit” is the pre-existing collection; “Generated” is the CANA/SDOC extension. Optical-constant references are given for the generated extension; bestfit-collection sources are documented in the companion repository.

Figure 2: The three DiSCo-TNOs exemplars, normalized at 0.9 µm. Dotted lines mark the bands discussed in the text.

Object (group)	Mode	$w$	Composition
1998 SN165	1	0.49	AC 0.71, H 2 O 0.22, NaCl 0.07
(bowl)	2	0.39	AC 0.57, H 2 O 0.23, NaCl 0.09, TTh 0.08
	3	0.12	AC 0.39, H 2 O 0.22, Oliv 0.16, NaCl 0.12, TTh 0.07, NH 3 0.02
2007 UK126	1	0.71	H 2 O 0.39, Oliv 0.17, AC 0.14, NH 3 0.11, CO 2 0.10, TTh 0.09
(double-dip)	2	0.20	H 2 O 0.40, Oliv 0.25, NH 3 0.14, AC 0.11, CO 2 0.09
	3	0.09	H 2 O 0.40, AC 0.21, NH 3 0.17, CO 2 0.12, TTh 0.04, NaCl 0.03

Appendix

Table 3: All modes recovered for the three exemplars. Components below 0.01 proportion are omitted.

Figure 3: Composition PIT split by target proportion: trace ( $\leq$ 0.05, top) and non-trace ( $>$ 0.05, bottom), on the test (left) and OOD (right) splits. A uniform histogram (dashed) indicates calibration.

Remote Sensing

Remote sensing uses satellite, aerial, and ground-based observations to measure and map Earth and other planetary surfaces.

See all 528 papers

A Missing Latent, Not a Missing Simulator: Radius-Augmented Inference for Real JWST Retrieval
Sep 30, 2026Angshuman Chakravertty, P V V RajSimulation-Based Inference Bayesian Inverse Problems
SVKM’s NMIMS (Deemed to be University)
Physics-Informed Synthetic Dataset and Denoising TIE-Reconstructed Phase Maps in Transient Flows Using Deep Learning
Apr 12, 2026Krishna Rajput, Vipul Gupta, Sudheesh K. Rajput +1Computational Imaging Image Denoising
Babu Banarasi Das University, 111, Faizabad Rd, Atif Vihar, Lucknow, Uttardhona, Uttar Pradesh 226028, India · Department of Engineering Physics, Koneru Lakshmaiah Education Foundation (KLEF) Vaddeswaram, 522302, Guntur, AP, India · Faculty of Electrical Engineering and Electronics, Kyoto Institute of Technology, Matsugasaki, Sakyo-ku, Kyoto, 606-8585, Japan
CETUS: How Far Do Representations Trained on Earth Transfer to Cassini SAR of Titan?
Oct 6, 2026Kevin LeeRemote Sensing Image Understanding Geospatial Representation Learning
University of California at Los Angeles, Los Angeles, CA · NASA Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA

Sep 30, 2026J/K move · Enter open · S save

A Missing Latent, Not a Missing Simulator: Radius-Augmented Inference for Real JWST Retrieval

Angshuman Chakravertty, P V V Raj

SVKM’s NMIMS (Deemed to be University)

Simulation-Based Inference Bayesian Inverse Problems

Open paper AskPDF arXiv

Amortized simulation-based inference (SBI), which is trained on radiative-transfer simulators, recovers exoplanet atmospheres accurately on synthetic James Webb Space Telescope (JWST) spectra but collapses when it comes to real reduced spectra. The flow posterior collapsed on real WASP-39b (importance-sampling effective sample size (ESS) = 1, best-fit \c{hi}2/N = 301), and such a failure is usually blamed on missing the forward-model physics, but ruling these levers out with nested sampling first (a temperature gradient, SO2 opacity, and a high-fidelity opacity set) leaves the fit unchanged, meaning the collapse is not from the simulator misspecification but instead from a missing latent, the planet radius. To fix this, we build MIRAGE, a radius-augmented flow-matching posterior calibrated against an independent nested-sampling reference with importance sampling and an optimal-transport map, which yields a physical and literature-consistent retrieval of real WASP-39b (with \c{hi}2/N from 301 to 0.06). This same method transfers unchanged across two instruments and three real JWST targets, including one spectrum self-reduced end-to-end from raw Mikulski Archive for Space Telescopes(MAST) data. The lesson is cross-domain, as a latent the simulator encodes but the inference omits can masquerade as misspecification.

Abstract

Remote Sensing

Explore similar work

A Missing Latent, Not a Missing Simulator: Radius-Augmented Inference for Real JWST Retrieval

Physics-Informed Synthetic Dataset and Denoising TIE-Reconstructed Phase Maps in Transient Flows Using Deep Learning

CETUS: How Far Do Representations Trained on Earth Transfer to Cassini SAR of Titan?