Isaiah was born and raised in Kona on the island of Hawai’i.
He graduated from Kealakehe high school and is currently
pursuing a Bachelor’s of Science in Computer Science at UH
Hilo. Isaiah is passionate about the synergy between
advancing computational power and space exploration,
specifically how machine learning can be applied to help us
understand our place in the Universe. He aims to pursue a
career at the intersection of AI and astrophysical research,
developing meaningful tools that will help humanity expand
and gain more knowledge. In his free time Isaiah enjoys
surfing, jiujitsu and anything that keeps him active..
Home Island: Hilo, Hawaii Island
High School: Kealakehe High School
Institution when accepted: UH Hilo
Site: Gemini Observatory. Hilo, Hawaii Island
Mentors: Patrick Parks & Winston Wu
Project title: Developing a Machine Learning Telescope Pointing Model for Gemini North
Project Abstract:
Modern large-aperture research telescopes must point with sub-arcsecond precision across the
entire sky under continuously changing conditions. At Gemini North, the existing pointing model
corrects for known geometric and atmospheric effects, yet a systematic residual error remains after
every slew. Operators must image the field, measure the offset, and issue a manual correction before
science observations can begin — a workflow that costs time and limits scheduling efficiency across
every night of operation. This work investigates whether machine learning can predict and remove
this residual, allowing the telescope to land on target on the first attempt without operator
intervention. Using ~34,000 acquisition corrections spanning twelve years of routine operations,
we examine what structure exists in the historical residuals and what is required to exploit it.
Analysis reveals two components: a slow-moving drift evolving over the course of a night, driven by
thermal and mechanical changes in the telescope structure, and a fast, unpredictable scatter tied to
atmospheric seeing and sensor noise. The drift is learnable; the scatter is not. A small causal
transformer trained on per-night slew sequences — treating each correction in the context of those
before it rather than in isolation — is identified as the most promising approach. In shadow-mode
evaluation against historical nights, this model reduces pointing residuals by 29% relative to the
operational baseline. Predictions would feed directly into existing correction slots within the
telescope control system, requiring no changes to hardware or firmware. If residuals can be reduced
below the blind-pointing threshold, manual corrections during normal operations could be
eliminated entirely.