KASI AI Hub KASI AI Hub

KASI AI Paper Reports 2026-07

1 Executive Summary

KASI's AI adoption stands at just 6.7% (114/1694 papers) and is declining (-0.3 pp/yr, from 8.2% peak in 2023 to 5.4% in 2025), while global astro-AI adoption on arXiv surged from ~5% to ~18% over the same window — a widening 3x gap. KASI's portfolio is concentrated in Tier 1 classical ML (59 papers) with zero Tier 4 foundation-model work, no simulation-based inference (SBI) papers, and only nascent transformer usage (1 paper), leaving it exposed as peers like Flatiron CCA (SBI) and CfA (LLM pipelines) build AI-native infrastructure. Critically, AI papers show no impact premium (median IF 5.36 vs 5.40 non-AI), signaling that current AI use is incremental rather than transformative. With Rubin LSST/SKA demanding AI-native pipelines and Korean national assets (KISTI Neuron GPU cluster, KAIS, K-SSA) available, KASI must urgently pivot from classical ML toward SBI and foundation models via targeted talent hubs already present (Tier 3 leaders Hong, Jeong, Park, Lee).

What we did

We collected 1694 KASI research paper records (2021–2025) from the institute's internal system, each stored as a structured YAML file containing title, abstract, author metadata, journal metrics, and instrument usage. Every paper was then classified by a large language model (Gemini 2.5 Flash Lite) across eight dimensions — AI maturity tier (0–4), centrality, specific techniques, subfield, and confidence score — producing a fully machine-readable enriched dataset. Aggregate statistics, year-over-year trends, departmental breakdowns, and a co-authorship network were computed from the classified data, and a synthesis model (Claude Opus 4.8) generated the strategic narrative presented in this report.

1,694
Total Papers
6.7%
AI Adoption Rate
114
AI Papers
18
Tier 3+ (Innovator)
38.6%
KASI Leads on AI
+-0.3 pp
AI Growth / Year
5.36
AI Median IF

2 Key Findings

  1. 1 AI adoption (6.7%) is stagnant-to-declining, dropping from a 2023 peak of 8.2% to 5.4% in 2025, diverging sharply from the global rise to ~18% by 2024.
  2. 2 The maturity curve is bottom-heavy: 59 Tier 1 (classical ML), 30 Tier 2 (standard DL), only 15 Tier 3 (physics-tailored), and 0 Tier 4 (foundation models).
  3. 3 Zero simulation-based-inference (SBI) papers (first→last: 0→0) despite SBI becoming mainstream globally and being Flatiron CCA's flagship strength.
  4. 4 AI papers deliver no impact advantage (median IF 5.36 vs 5.40 for non-AI) and no collaboration boost (49.1% vs 49.2% intl), indicating AI is applied incrementally.
  5. 5 Technique trends are flat or shrinking: classical-ML 9→7, CNN 8→7, GAN 1→0, normalizing-flow 1→0; only modest growth in gaussian-process (2→4) and transformer (0→1).
  6. 6 AI expertise is heavily concentrated: 태양권관측센터 (27.6%, 29 papers) and 기초천문연구본부 (30.0%, 10 papers) lead by rate, while large departments like 이론천문센터 (238 papers) sit at only 14.3%.
  7. 7 A functional collaboration network exists (176 researchers, 682 edges) with 4 of 5 top hubs at Tier 3, providing a ready nucleus for advanced-AI capability building.

Key Metrics

Metric Value
Total papers (2021–2025)1,694
AI/ML papers114 (6.7%)
Tier 3+ (physics-tailored) papers18 (15.8%)
KASI-led AI papers44 (38.6%)
Year-over-year growth+-0.3 pp/yr
AI median impact factor5.36
Non-AI median impact factor5.4
Coverage period2021–2025

3 Adoption Trends

AI adoption rate by year (2021–2025) — all AI / core AI / supporting AI
AI capability tier distribution by year (Tier 1 = classical ML → Tier 3 = physics-tailored, stacked)
KASI's AI portfolio is bottom-heavy — 59 Tier 1 (classical ML) and 30 Tier 2 (standard DL) dominate, with only 15 Tier 3 (physics-tailored) and zero Tier 4 (foundation models). Compared to peers building AI-native infrastructure (Flatiron SBI, CfA LLMs), KASI is competitive in applied ML but absent from the innovator/pioneer frontier. The presence of four Tier 3 network hubs offers a credible foundation to climb the maturity curve if resourced.

4 Technique Evolution

AI technique mix across years (stacked by technique type)

5 Operational Gap Map

AI application role × capability tier heatmap — paper count per cell
Research subfield × capability tier heatmap

6 Department & Collaboration

AI adoption rate (%) by KASI department — hover for paper counts
International collaboration rate: AI papers vs. non-AI papers
AI papers: 47.4% international collaboration vs. 49.4% for non-AI papers.
KASI authorship role in AI papers: lead author vs. co-author
KASI first-authors 38.6% (44) of its AI papers and co-authors the remaining 61.4%, showing genuine capability ownership but a majority reliance on external-led collaborations. To become an AI thought-leader rather than a contributor, KASI should raise its lead-author share on frontier (Tier 3/4) AI work specifically.

7 Impact Factor & Instruments

Impact factor distribution: AI vs. non-AI papers — whiskers span Q1 − 1.5×IQR to Q3 + 1.5×IQR, hover for statistics
AI papers show effectively no impact premium (median IF 5.36 vs 5.40 non-AI), indicating AI is currently applied incrementally rather than for high-impact discovery; the strategic goal must be AI that measurably lifts scientific impact.
AI paper count by instrument / facility (sorted by AI use, hover for totals)

8 Author Network & Internal Champions

183 researchers · 704 co-authorship edges. Drag nodes to explore, scroll to zoom, hover for details. Color = tier.

Internal AI co-authorship network — node size = connection count

Top AI Research Hubs

Arman Shafieloo (Tier 1, 10 connections) Sungwook E. Hong (Tier 3, 4 connections) Se-Heon Jeong (Tier 3, 8 connections) Eunsu Park (Tier 3, 1 connections) Joongoo Lee (Tier 3, 3 connections) David Parkinson (Tier 2, 8 connections) J. An (Tier 3, 0 connections) Min-Su Shin (Tier 2, 5 connections) Jung Youn Kil (Tier 3, 13 connections) Ji-Hye Baek (Tier 3, 7 connections)

9 SWOT Analysis

Strengths

  • Established Tier 3 physics-tailored expertise via hubs Sungwook E. Hong, Se-Heon Jeong, Eunsu Park, Joongoo Lee, plus Tier 1 leader Arman Shafieloo (cosmology).
  • Dense internal AI network of 176 researchers and 682 co-author edges enabling rapid internal diffusion.
  • Domain leadership pockets: 태양권관측센터 at 27.6% AI rate (29 papers) and 기초천문연구본부 at 30.0% — solar/heliophysics is a defensible AI niche.
  • KASI leads (first-author) on 38.6% of its AI papers (44 papers), showing genuine capability ownership rather than pure participation.

Weaknesses

  • Declining adoption trajectory (-0.3 pp/yr; 8.2%→5.4% 2023→2025) versus global growth to ~18%.
  • Zero Tier 4 foundation-model papers and zero SBI papers — absent from the two fastest-growing global frontiers.
  • No measurable quality dividend from AI: IF 5.36 vs 5.40 non-AI, and intl collaboration essentially identical (49.1% vs 49.2%).
  • AI capacity is siloed; theory-heavy 이론천문센터 (238 papers, only 14.3% AI) and 변광천체그룹 (132 papers, 11.4%) underutilize AI at scale.

Opportunities

  • Leverage KISTI Neuron national GPU cluster to train foundation models and run SBI pipelines without owning hardware.
  • Align with Rubin LSST and SKA AI-native pipeline needs for time-domain and radio survey science (변광천체그룹, transients subfield ready).
  • Join national/regional programs — KAIS, A3 Net, and the K-SSA opportunity — to co-develop AI-native space-situational-awareness and astronomy tooling.
  • Build an SBI/foundation-model partnership with peers (Flatiron CCA on SBI, CfA on LLM pipelines, NAOJ AI observatory programs) to leapfrog the capability gap.

Threats

  • Global peers are entrenching AI-native pipelines (Flatiron SBI, STScI AI task force, MPIA GRAVITY+ ML, CfA LLMs) while KASI adoption stalls at 6.7%.
  • Rubin LSST/SKA data volumes will make non-AI pipelines uncompetitive; KASI risks being a data consumer, not a discovery leader.
  • Talent flight risk: with only 15 Tier 3 and 0 Tier 4 researchers, top AI hubs may be recruited by better-resourced international programs.
  • The 3x and widening gap (6.7% vs 18%) compounds annually, making catch-up progressively more expensive and slower.

10 3–5 Year Strategic Roadmap

The strategic recommendations below were generated by an AI model (Claude Opus 4.8) based on the quantitative analysis above. They are provided as a reference and starting point for discussion — not as institutional policy or expert human judgment.
1

Establish an SBI & Foundation-Model Capability

High priority 0-1 yr

Stand up a dedicated AI-methods unit anchored on existing Tier 3 hubs (Hong, Park, Jeong, Lee) to close the SBI (currently 0 papers) and foundation-model (0 papers) gaps. Secure KISTI Neuron GPU allocation and partner with Flatiron CCA for SBI and CfA for LLM pipelines.

KPIs

  • ≥5 SBI papers within 24 months (from 0)
  • First Tier 4 foundation-model paper published
  • Secured KISTI Neuron allocation with ≥3 active projects
2

Reverse the Adoption Decline via Departmental Diffusion

High priority 1-3 yr

Deploy internal training and embedded ML specialists into large under-adopting departments (이론천문센터 14.3%/238 papers, 변광천체그룹 11.4%/132 papers) to lift institute-wide adoption back above the 2023 peak. Use the 682-edge network to seed AI methods across subfields.

KPIs

  • Institute AI adoption ≥12% by 2027 (from 6.7%)
  • 이론천문센터 AI rate ≥25% (from 14.3%)
  • Tier 2+ papers doubled (from 45 to 90)
3

AI-Native Survey Pipeline Readiness (LSST/SKA/K-SSA)

High priority 1-3 yr

Build production AI pipelines for time-domain (변광천체그룹, transients) and solar/heliophysics (태양권관측센터, 27.6%) tied to Rubin LSST, SKA, and the K-SSA opportunity. Position KASI as Korea's AI-native survey analysis hub via KAIS and A3 Net.

KPIs

  • ≥2 deployed AI-native pipelines for LSST/SKA/K-SSA
  • K-SSA AI role formalized with funding
  • ≥3 A3 Net joint AI publications
4

Convert AI into Impact Advantage

Medium priority 1-3 yr

Target high-impact, AI-enabled discovery science so that AI papers outperform rather than match the baseline (currently IF 5.36 vs 5.40). Prioritize novel physics-tailored methods over incremental classical-ML applications.

KPIs

  • AI-paper median IF exceeds non-AI by ≥0.5
  • AI intl-collaboration rate ≥60% (from 49.1%)
  • Tier 3+ share of AI papers ≥40% (from 13%)

11 Strategic Moonshots

The moonshot ideas below were generated by an AI model (Claude Opus 4.8) based on the quantitative analysis above. They are provided as a reference and starting point for discussion — not as institutional policy or expert human judgment.

KASI Solar Foundation Model (Heliophysics LLM/Transformer)

태양권관측센터's leading 27.6% AI rate (29 papers) and long solar observation archives give KASI a unique domain-data advantage to train the first solar/heliophysics foundation model — a niche no global peer dominates.

Action: Assemble a labeled multi-instrument solar dataset and prototype a transformer-based flare/CME forecasting model on KISTI Neuron within 12 months.

Korean Astro-SBI National Facility

KASI has cosmology strength (Shafieloo) and Tier 3 physics-tailored talent but zero SBI papers; establishing a national SBI service on KISTI Neuron via KAIS would let Korea leapfrog directly to the mainstream frontier Flatiron CCA leads.

Action: Recruit or partner for one senior SBI methodologist and deliver a pilot SBI cosmological-parameter inference pipeline within 18 months.
1,694 papers classified by Gemini-2.5-Flash-Lite  ·  Synthesis by Claude Opus 4.8