0 Assigned ID: 7588 ----- Attempting Submission 1 (Thu Apr 16 19:51:59 GMT 2026) ----- The Nancy Grace Roman Space Telescope will serendipitously observe thousands of solar system objects during its Core Community Surveys, but the default pipeline will flag moving-target signatures as cosmic rays and remove them. Recovering these streaks is essential for planetary defense: Roman can detect NEOs as small as 20 m, reducing orbital uncertainties by 95-99%. Community ML pipelines (ZStreak, Tails) are being developed but trained exclusively on ground-based data, and no ML-ready database of real space-based streak detections exists. We propose STASH (Space Telescope Archival Streak Harvester): an open-source pipeline that extracts ML-ready asteroid streak data from the HST archive for Roman moving-object detection. STASH will harvest all streaks from the Hubble Asteroid Hunter project, construct a labeled training set including cosmic rays and artifacts, and demonstrate that HST-trained classifiers outperform ground-only models. A key innovation uses 3x3 binned WFC3/UVIS data (0.12 arcsec/pix) as a Roman-proxy test set, matching Roman WFI to within 9%. Deliverables include open-source software, a MAST HLSP catalog, pre-trained models, and astrometry submitted to the Minor Planet Center. PI Sasli is a postdoctoral researcher specializing in deep-learning methods for astronomical image classification. Sasli leads the AppleCiDEr project, a multimodal classification framework that combines image cutouts with contextual metadata and has demonstrated state-of-the-art performance for transient classification in ZTF data. Sasli has expertise in exploring and benchmarking diverse deep-learning architectures---including convolutional neural networks, vision transformers, and multimodal networks---for astronomical applications. Co-I Coughlin is an expert in alert stream software development and time-domain astronomy infrastructure. Coughlin leads the development of SkyPortal, a web-based target and observation manager deployed for ZTF and planned for LSST, LS4, ULTRASAT, and UVEX, and is co-lead of BOOM/BABAMUL, one of the approved LSST community brokers. Coughlin's group has extensive experience processing large astronomical survey datasets in real time, including filtering ~10^6 alerts per night from ZTF and classifying >370 million light curves. The Tails deep-learning system for detecting cometary and asteroidal morphologies has been integrated into SkyPortal and runs in production on ZTF data. Co-I Bolin is an expert in near-Earth object discovery and characterization using wide-field surveys. Bolin has extensive experience with ZTF's NEO detection systems, including the ZStreak fast-moving object pipeline and ZMODE tracklet detection engine, and has led numerous NEO discovery and characterization campaigns using ZTF data, including the recent characterization of Earth-crossing asteroid 2024~YR_4. Bolin's work includes the discovery and characterization of interstellar objects, fast-rotating NEOs, and minimoons, with objects regularly reported to the Minor Planet Center. Solar System Astronomy 0 false false false true true true false SUPPORTED