More about HKUST
Agentic Systems for Earth Science: From Tool-Augmented Assistants to Self-Evolving Scientific Agents
PhD Qualifying Examination
Title: "Agentic Systems for Earth Science: From Tool-Augmented Assistants to
Self-Evolving Scientific Agents"
by
Mr. Chenyue LI
Abstract:
Earth science is entering a transition from model-centric AI to
capability-centric agent systems. Earlier waves of AI improved local
components of the scientific pipeline—forecasting, retrieval, language
interaction, and multimodal understanding—but they rarely closed the
end-to-end loop of scientific work. This paper reframes current progress
through a four-level capability ladder: Level 1 tool-augmented assistants,
Level 2 autonomous workflow agents, Level 3 research agents, and Level 4
self-evolving scientific agents. We argue that this taxonomy should organize
the field from the outset rather than appear only as a concluding outlook. We
then reconstruct the co-evolution of AI and Earth-science technology, from
physics-first numerical modeling and observation-constrained reanalysis to
deep representation learning, Earth foundation models, reasoning-oriented
LLMs, and emerging agentic systems. A central conceptual distinction is that
Earth foundation models act as world models for spatiotemporal state
evolution, whereas climate or geo LLMs act as language interfaces for
knowledge, synthesis, and planning. Agent systems are the orchestration layer
that connects these components to tools, data, and executable environments.
Reviewing current systems shows that Earth science already has credible Level
1 and early Level 2 examples, while Level 3 and Level 4 remain open
frontiers. The key barriers are physical-model integration, domain skill
infrastructure, memory and tacit knowledge accumulation, and evaluation
frameworks that measure scientific usefulness rather than isolated model
performance. We conclude that reasoning LLMs can serve as the thinking layer,
Earth foundation models and physical simulators can serve as the world-model
layer, and agent systems can provide the control loop for scientific action,
validation, and long-horizon discovery.
Date: Monday, 21 April 2026
Time: 2:00pm - 4:00pm
Venue: Room 2131B
Lift 22
Committee Members: Dr. Binhang Yuan (Supervisor)
Dr. Ling Pan (Chairperson)
Dr. Mengqian Lu (CIVL)