Kai Hou Yip

Research profile

Research summary

Kai Hou Yip (King's College London, United Kingdom; ORCID 0000-0002-9616-1524) works at the interface of astrophysics and machine learning, with an even 50/50 split between trustworthy machine learning for astronomy (astroinformatics) and computational astrophysics centred on exoplanet atmospheres and scientific machine learning. This profile is name-resolved from OpenAlex bibliographic data. The record lists an h-index of 10 and an i10-index of 11, 32 works over the last three years, and 9 distinct co-authors; annual output rises from 3 works in 2018 to 10 in 2020, 10 in 2022, a peak of 14 in 2024, and 9 in each of 2025 and 2026, giving a maximum activity score of 100/100. OpenAlex topics are led by stellar, planetary, and galactic studies (58) and astronomy and astrophysical research (56), with astro and planetary science (22), adaptive optics and wavefront sensing (13), and galaxies (8); a focus score of 0/100 reflects this breadth. Evolve has reviewed two 2026 arXiv preprints on which Yip is a middle, non-corresponding author: a graph neural network surrogate accelerating exoplanet disequilibrium chemistry to under one second with 2-3x lower error than a U-Net, and a safety-cage framework bounding the operational range of spectroscopy ML models, claiming error reductions on Ariel-spectra tasks at a 20% coverage cost. Both carry medium repro-readiness and test-set-selection caveats. Frequent co-authors include Ingo P. Waldmann, Nikolaos Nikolaou, Ahmed F. Al-Refaie, Olivia Venot, and Luis F. Simoes.

Research keywords

Highlights

• GNN surrogate runs exoplanet disequilibrium chemistry in under a second

  Evolve's review of this 2026 arXiv preprint (middle author) reports a graph neural network surrogate that speeds exoplanet disequilibrium chemistry to <1s with 2-3x lower error than a U-Net, positioned as a chemical kinetics surrogate for atmospheric retrieval; the verdict calls the work solid but flags test-set model selection inflating metrics.

• Safety cage framework bounds operational range for spectroscopy ML

  In this 2026 arXiv preprint (middle author), a safety-cage architecture using runtime monitoring and operational domain bounding claims to cut errors on Ariel transmission-spectra tasks at a 20% coverage cost; Evolve calls the approach systematic but notes test-set selection and an abstract-body mismatch temper the headline.

• A consistent thread: trustworthy ML for Ariel-era spectra

  Both reviewed papers sit in the 50/50 field split between astroinformatics ML and computational astrophysics, targeting Ariel-relevant exoplanet spectra workflows; both carry medium repro-readiness and the same test-set-selection caveat, defining the current evidence base for this profile.

Timeline

1. 2018Earliest year in the OpenAlex counts-by-year record, with 3 works.
2. 2020Output reaches 10 works, the first double-digit year in the record (matched again in 2022).
3. 2022A second 10-work year consolidates the mid-period research pace.
4. 2024Peak annual output of 14 works, the high point of the counts-by-year record.
5. 2026Two arXiv preprints reviewed by Evolve: the GNN chemical kinetics surrogate and the spectroscopy safety-cage framework, both with Yip as middle author.

Papers in the Evolve catalog

• 2026-09-15

  Graph neural networks for exoplanet atmospheres

• 2026-09-11

  Operational Range Bounding in Spectroscopy: A Safety Cage Framework for Machine Learning Models

Frequent co-authors