Machine learning-driven autonomous control of the distributed aperture for the Small Exo-Life Finder (SELF) Telescope

Arteaga-Marrero, Natalia; Padrón-Brito, Auxiliadora; Pérez del Pino, Salvador; Johann Dobat, Max; Sánchez-Chaves, Adrián.; Tamayo-Guzmán, Diego; Iborra Luis, Jordi; Fernández-Izquierdo, Patricia; Lodieu, Nicolas; Kuhn, Jeff
Referencia bibliográfica

Ground-based and Airborne Instrumentation for Astronomy XI

Fecha de publicación:
8
2026
Número de autores
10
Número de autores del IAC
10
Número de citas
0
Número de citas referidas
0
Descripción
The Small Exo-Life Finder (SELF) is a 3.5-meter hybrid-interferometric, ground-based telescope envisioned as a technology demonstrator for the future 35-50-meter Exo-Life Finder (ELF), a next-generation telescope designed to detect potential bio- and technosignatures through direct imaging. SELF employs a distributed aperture architecture consisting of 15 independently controlled primary mirrors, each paired with a corresponding secondary mirror in a one-to-one configuration. The system incorporates numerous optical and mechanical degrees of freedom, making robust machine-learning-driven control essential for maintaining precise alignment to within a small fraction of the wavelength.

A convolutional neural network (CNN) was implemented within a supervised-learning framework to estimate alignment errors directly from focal-plane images. The network was trained to learn the nonlinear mapping between the sub-aperture degrees of freedom and their corresponding focal-plane signatures. The approach was investigated using both experimental measurements acquired with the nanoELF laboratory demonstrator, a simplified four-aperture representation of SELF, and synthetic datasets generated with an HCIPy model of the system. Broadband simulations including atmospheric turbulence were also generated to assess performance under representative observing conditions.

The results demonstrated that tip-tilt perturbations could be estimated from focal-plane images for both experimental and simulated datasets, achieving root-mean-square errors below 16 μrad and 23 μrad, respectively, over a perturbation range of +/-200 μrad. Furthermore, reducing the focal-plane image resolution by half resulted in little degradation in performance while reducing computational requirements. The estimation of piston differences between sub-apertures proved more challenging. Although moderate performance was achieved in simulations, piston differences could not be reliably recovered from the experimental measurements over the full tip-tilt perturbation range. Meaningful piston difference estimation was achieved only when the tip-tilt range was reduced to +/-20 μrad. Additional experiments also revealed a significant domain gap between experimental and simulated datasets when models trained exclusively on simulations were applied to laboratory data.

These results demonstrate the potential of machine-learning-based focal-plane sensing for the alignment of distributed-aperture telescopes and provide a framework for analyzing the sensitivity of focal-plane images to alignment perturbations. The study also highlights the challenges associated with piston estimation and simulation-to-real transfer.