Machine-learning approaches to improve telescope efficiency and adaptive-optics instrument strategy at the Large Binocular Telescope

Abstract

Efficient operation of a large optical telescope requires the telescope, the active optics, and the adaptive optics to begin each acquisition close to their physically correct state in order to reduce setup time. At the Large Binocular Telescope (LBT), the acquisition sequence must place the guide star on the active-optics sensor or Pyramid Wavefront Sensor (PyWFS), update the residual pointing-model coefficients, tune the active-secondary collimation, close the active-optics or AO loop, and, for AO observations, maintain appropriate loop gains as seeing evolves. We present a unified machine-learning framework, built independently for two projects; pointing/collimation and AO loop-gain optimization but sharing a common Random Forest (RF) point-estimation and Quantile Regression Forest (QRF) uncertainty-quantification architecture under one observatory-operations program. For telescope pointing and collimation, we fuse LBT environmental telemetry with per-side pointing terms (azimuth collimation and elevation index error) and hexapod telemetry stored in the Distributed Monitoring System (DMS) on a common 60 s grid over 2023–2026, and train independent RF/QRF regressors for the pointing coefficients azimuth collimation (CA) and elevation index error (IE), and for the five hexapod rigid-body terms (X, Y, Z, Rx, Ry), across both LBT sides and the Front/Central Bent Gregorian foci (LUCI and LBTI instruments). The mean RMSE ratio against a training-mean baseline is 0.94 (rolling-origin) and 0.87 (date k-fold); a held-out test day (2025-12-19) recovers CA to 9.5–11.6′′ and IE to 1.6–7.7′′ RMSE with 90% prediction intervals (PI90) covering 97–100% of measurements, and a strict future hold-out ten days beyond the training cut-off reproduces these numbers (RMSE 9.6′′ CA, 10.3′′ IE, PICP 0.99). For the SOUL (Single-conjugated adaptive Optics Upgrade for LBT) AO system, we predict the tip-tilt (TT) and two high-order (HO1, HO2) integrator loop gains, whose optimal values depend on WFS frame rate, atmospheric seeing, wind, and elevation, and fall into distinct spatial-frequency regimes; below 400 Hz frame rate, HO2 is governed by a deterministic 0.10 floor rule rather than a fitted relationship. Grouped 5-fold cross-validation by observing night (1,883 DX / 2,378 SX sweeps) gives R2(MEC) of 0.23/0.16 (TT), 0.37/0.15 (HO1), and a combined 0.50/0.45 (HO2, DX/SX) once the floor rule is folded in, with 90% prediction-interval coverage of 0.78–0.84. A genuine held-out night (2026-04-27), re-scored against each sweep’s real measured frame rate across the night’s five distinct frequency blocks, gives positive R2 on all six side×target channels (0.33–0.71). Side-local swing-arm wind telemetry is a significant, independent predictor of HO2 specifically, on both sides, complementing rather than replacing the exterior-wind feature already in production use. The side-resolved RF/QRF models are deployed as a confidence-gated bootstrap for the optimize-gain sequence, not as a replacement for operator judgment or the on-sky sweep. Both models are packaged as lightweight, independently deployable command-line and HTTP inference tools intended to bootstrap operator decisions with calibrated confidence rather than replace them.