Physics-Informed Coherent-Basis Networks with Near-Unitary Mixing: Toward Calibrated-Uncertainty Inference of Exoplanet Transmission Spectra

Yun Juei Yen

Keywords: near-unitary mixing, coherent basis expansion, uncertainty calibration, systematics-robust inference

Abstract:

Recovering exoplanet transmission spectra with reliable uncertainty from ARIEL-like observations is challenged by strong systematics such as pointing jitter, PSF variations, and thermal drifts. Generic deep networks lack physical priors and often yield poorly calibrated uncertainties.
We introduce a physics-aligned \emph{coherent-basis front end} for exoplanet transmission spectroscopy that differs from conventional end-to-end CNN/Transformer pipelines in two key ways. First, instead of learning arbitrary features, each time step is projected onto instrument- and systematics-aware Hermite/Radial/Angular bases to form complex coherent states that explicitly encode PSF/jitter and common-mode structure. Second, channel interactions are performed by a near-unitary mixing layer parameterized via a Cayley transform, preserving geometric energy while remaining fully vectorizable; states are propagated as a stable two-channel real representation $[\Re\psi,\Im\psi]$ to avoid brittle phase differentiation. This front end plugs into a lightweight temporal head (e.g., TCN/SSM) and a heteroscedastic output layer to produce per-wavelength predictions with calibrated uncertainties. The resulting architecture offers (i) sample-efficient inductive bias aligned with the optics and pointing physics, (ii) robustness to slowly varying systematics through conservative (near-unitary) mixing, and (iii) interpretability via basis responses and mixing spectra. Compared with generic deep models, the approach targets where physics matters—feature construction and mixing—yielding a compact, uncertainty-aware pipeline for high-noise spectral inversion.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: true

Submission Guidelines: true

Anonymous Url: true

No Acknowledgement Section: true

Submission Number: 25160