A Hierarchical Multi-Agent System for Autonomous Discovery in Geoscientific Data Archives

Dmitrii Pantiukhin; Ivan Kuznetsov; Boris Shapkin; Antonia Anna Jost; Thomas Jung; Nikolay Koldunov

cs.AIFeb 24, 2026

A Hierarchical Multi-Agent System for Autonomous Discovery in Geoscientific Data Archives

Authors: Dmitrii Pantiukhin, Ivan Kuznetsov, Boris Shapkin, Antonia Anna Jost, Thomas Jung, Nikolay Koldunov

Organizations: Alfred Wegener Institute for Polar and Marine Research, Bremerhaven, Germany

Abstract

The rapid accumulation of Earth science data has created a significant scalability challenge; while repositories like PANGAEA host vast collections of datasets, citation metrics indicate that a substantial portion remains underutilized, limiting data reusability. Here we present PANGAEA-GPT, a hierarchical multi-agent framework designed for autonomous data discovery and analysis. Unlike standard Large Language Model (LLM) wrappers, our architecture implements a centralized Supervisor-Worker topology with strict data-type-aware routing, sandboxed deterministic code execution, and self-correction via execution feedback, enabling agents to diagnose and resolve runtime errors. Through use-case scenarios spanning physical oceanography and ecology, we demonstrate the system's capacity to execute complex, multi-step workflows with minimal human intervention. This framework provides a methodology for querying and analyzing heterogeneous repository data through coordinated agent workflows.

Figures & tables

Figure 1: Conceptual framework of the PANGAEA-GPT Multi-Agent System (MAS). The system uses a two-phase, hierarchical architecture. First, a Search Agent discovers relevant datasets from a user’s natural language request. A Supervisor Agent then delegates analysis and visualization tasks to a team of specialist agents working within a secure sandbox environment. Finally, a Writer Agent synthesizes the results into a cohesive report.

Panel of Figure 2 — Figure 2: Exploratory analysis of microplastic distribution in the Weddell Sea (Supplementary Note 1).

Figure 3: Validation scatter plot comparing in-situ temperature observations (0–500 m) from ten HAUSGARTEN moorings with co-located values from the Copernicus GLORYS12V1 reanalysis. The Oceanographer Agent retrieved daily thetao fields, performed 4D nearest-grid-point matchup ( $N=135{,}678$ ), and calculated statistics (Bias $+0.35\,^{\circ}$ C, RMSE $1.09\,^{\circ}$ C, $r=0.31$ ). Color scale indicates observation depth.

Panel of Figure 4 — Figure 4: ERA5/MOSAiC Lagrangian validation and wind regime characterization.

Figure 5: Bio-physical coupling generated by the Oceanographer Agent. The system augmented the PANGAEA biological dataset by retrieving and co-locating 4D hydrographic data ( $\theta$ / $S$ ) from the Copernicus Marine Service GLORYS12V1 reanalysis. The T-S diagram highlights the co-occurrence of Aglantha digitale with distinct water masses ( $\log_{10}$ -abundance).

Figure 6: Grouped bar chart comparing Baseline (Elastic Search), Simple LLM, and Agentic Search architectures across five semantic metrics using 100 geoscientific queries. Error bars indicate standard deviation. The Agentic Search method outperforms traditional approaches in all categories.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

#	Actor	Action	Content / Reasoning
1	User	Prompt	“Retrieve Copernicus Marine surface currents ( uo , vo ) for [78 ∘ S–60 ∘ S, 57 ∘ W–12 ∘ E]. Calculate mean current speed averaged over austral summers 2018–2019. Create a regional map with current speed heatmap, streamlines, and sampling stations sized by microplastic abundance.”
2	Supervisor	Planning	Decomposes into two sub-tasks: (i) data retrieval and processing $\rightarrow$ Oceanographer Agent; (ii) cartographic visualization $\rightarrow$ Visualization Agent.
3	Oceanographer	Temporal inference	Autonomously parses the Date/Time column of the observational dataset and determines the sampling campaigns span January–February 2018 and January–February 2019. Constructs two separate API requests with matching temporal bounds.
4	Oceanographer	API call ( fail )	First Copernicus Marine request specifies minimum_depth=0.0 . Service rejects the request: depth axis begins at 0.494 m.
5	Oceanographer	Self-correction	Parses the error message, extracts the minimum valid depth coordinate (0.494 m), and resubmits. Both requests (2018, 2019) succeed. Downloads daily $u$ , $v$ fields for the specified domain.
6	Oceanographer	Code execution	Computes temporal mean of daily velocity fields across both summers using xarray . Derives current speed as $\sqrt{u^{2}+v^{2}}$ . Saves result as NetCDF ( mean_surface_currents_JF_2018_2019_Weddell.nc ).

Appendix

Table S1: Annotated execution log for Scenario 1 (Weddell Sea microplastic exploration). Bold text indicates autonomous agent decisions. The workflow comprises three user prompts (Steps 1, 9, 11) and eleven autonomous agent actions, including one API error-recovery episode (Steps 4–5) and one reflexive visualization refinement cycle (Step 8).

Depth bin	$N$	Bias ( ∘ C)	RMSE ( ∘ C)	$r$
50–100 m	7,738	$+0.20$	0.97	$-0.05$
100–200 m	47,119	$+0.42$	1.06	$+0.11$
200–500 m	80,821	$+0.32$	1.12	$+0.42$

Appendix

Table S2: Depth-binned validation statistics for GLORYS12V1 potential temperature against HAUSGARTEN mooring observations (0–500 m, 2006–2012).

#	Actor	Action	Content / Reasoning
1	User	Prompt	“Generate a plot visualizing the distribution of instrument depths across all loaded mooring datasets.”
2	Supervisor	Planning	Decomposes into two sub-tasks: (i) multi-dataset ingestion and depth extraction $\rightarrow$ Visualization Agent (with inline data wrangling); (ii) exploratory depth-distribution figure $\rightarrow$ Visualization Agent.
3	Visualization	Data profiling	Loads all 10 CSV files and inspects column names. Identifies Depth water as the depth column across all datasets. Notes that three datasets (FEVI21, FEVI22, FEVI23) contain an additional Sal column (14 columns vs. 13). Records total row counts: 340,022 records across 10 moorings.
4	Visualization	Data engineering	Computes per-instrument unique depths to avoid time-repeated values: groups by Gear ID within each event and takes the median of Depth water . Produces a consolidated table of $N=35$ unique instrument depths across 10 moorings, spanning 81–2,737 m.
5	Visualization	Code execution	Generates a two-panel figure: (top) overall histogram with KDE of the 35 unique instrument depths; (bottom) violin + box + strip plot faceted by mooring event, ordered by median depth. Annotates per-mooring instrument counts ( $n=3$ – $5$ ).
6	Visualization	Reflection loop	Reflexive quality control scores the initial figure at 7/10 : “Axis labels slightly small; histogram x-axis extends into negative values; font sizes could be increased for readability.”

Appendix

Table S3: Annotated execution log for Scenario 2 (HAUSGARTEN mooring validation). Bold text indicates autonomous agent decisions. The workflow comprises three user prompts (Steps 1, 9, 11) and sixteen autonomous agent actions, including one error-recovery episode (Steps 15–16, datetime type mismatch) and two reflexive visualization quality control cycles (Steps 6–8, 19).

#	Actor	Action	Content / Reasoning
1	User	Prompt	“Validate ERA5 reanalysis against the ship’s measurements while minimizing data download size. Calculate the exact spatial bounding box and time range from the ship’s data. Retrieve ERA5 2m_temperature only for that specific region. Interpolate the grid to the ship’s hourly positions, convert to Celsius. Create a time-series plot comparing in-situ air temperature (column TTT ) in blue against ERA5 temperature in red.”
2	Supervisor	Planning	Decomposes into three sub-tasks: (i) spatiotemporal bounds extraction and hourly resampling $\rightarrow$ DataFrame Agent; (ii) ERA5 cloud-native retrieval and Lagrangian interpolation $\rightarrow$ Oceanographer Agent; (iii) time-series visualization $\rightarrow$ Visualization Agent.
3	DataFrame	Temporal inference	Parses the Date/Time column and computes spatiotemporal bounds: bounding box [18.99 ∘ E–138.14 ∘ E, 69.68 ∘ N–86.60 ∘ N], time range 2019-09-20 to 2019-12-13. Resamples to hourly resolution (last observation per hour), producing $N=2{,}013$ hourly positions. Saves hourly_ship_table.csv .
4	Oceanographer	API call	Submits ERA5 cloud-native retrieval request for 2m_temperature with bounding box [18.99 ∘ E–138.14 ∘ E, 69.68 ∘ N–86.60 ∘ N] and temporal range matching the cruise. Receives Zarr store: 2,013 time steps $\times$ 68 latitude $\times$ 477 longitude ( ${\sim}$ 261 MB).
5	Oceanographer	Code execution	Interpolates ERA5 2m_temperature field to $N=2{,}013$ hourly ship positions using trilinear interpolation ( xarray.interp , method= linear ). Converts from Kelvin to Celsius.
6	Oceanographer	Quality check (fail)	Detects 83 of 2,013 interpolated values returned NaN . Diagnoses the cause: ship positions near the bounding-box edge fall outside the ERA5 grid domain (ERA5 grid starts at 69.75 ∘ N / 19.00 ∘ E, but ship positions extend to 69.68 ∘ N / 18.99 ∘ E).

Appendix

Table S4: Annotated execution log for Scenario 3 (ERA5/MOSAiC Lagrangian validation). Bold text indicates autonomous agent decisions. The workflow comprises two user prompts (Steps 1, 11) and fourteen autonomous agent actions, including two error-recovery episodes (Steps 6–8, bounding-box edge effect; Steps 13–14, bin-label mismatch) and two reflexive visualization quality control cycles (Steps 10, 16).

#	Actor	Action	Content / Reasoning
Phase 1: Geographical transect division
1	User	Prompt	“Divide the sampling stations into two geographical transects based on longitude to separate the Norwegian Sea leg from the Irminger Sea leg. Create two new dataframes: ‘Eastern Transect’ and ‘Western Transect’ for stations, use $-20^{\circ}$ longitude as a border. Visualize these transects at the end.”
2	DataFrame	Data engineering	Identifies Longitude and Event columns. Splits at $-20^{\circ}$ E: Eastern Transect (60 rows, 18 unique stations) and Western Transect (34 rows, 11 unique stations).
3	DataFrame	Code execution	Generates lon/lat scatter plot with cruise track, dashed border at $-20^{\circ}$ , and “Irminger Sea leg” / “Norwegian Sea leg” annotations. Saves transects_map_lonlat_border_-20.png .
4	Visualization	Code execution	Parses Date/Time , splits transects, sorts by time. Regenerates the transect map with connected cruise-track lines, styled scatter points, and text annotations.
5	Visualization	Reflection loop	Reflexive quality control scores the transect map at 8/10 : “Axis labels clear, colors distinguishable, no overlapping elements.” Accepted without further correction.

Appendix

Table S5: Annotated execution log for Scenario 4 (G.O. Sars jellyfish transect analysis with Copernicus validation). Bold text indicates autonomous agent decisions. The workflow comprises three user prompts (Steps 1, 6, 20) and twenty-three autonomous agent actions, including two error-recovery episodes (Steps 7–8, divide-by-zero in log; Steps 11–19, extended reflexive visualization quality control with five iterations, score progression: 3

\rightarrow

\rightarrow

\rightarrow

\rightarrow

9/10), and two single-pass reflection cycles (Steps 5, 26).

Scenario	Errors	Error Types	VQC Cycles	Recovery
1: Weddell Sea	1	API depth parameter validation	1	1/1
2: HAUSGARTEN	1	Datetime type mismatch	2	1/1
3: MOSAiC/ERA5	2	Bounding-box edge effect, bin-label mismatch	2	2/2
4: Jellyfish T-S	1	Divide-by-zero in $\log$	6	1/1
Total	5		11	5/5 (100%)

Appendix

Table S6: Summary of autonomous error recovery across all validation scenarios. “Errors” counts distinct runtime failures resolved via kernel traceback introspection. “VQC Cycles” counts the total number of reflexive visual quality-control (VQC) cycles performed, including both correction iterations (score below acceptance threshold) and single-pass evaluations. All errors were resolved without human intervention.

Query intent	Baseline (Elasticsearch)	Simple LLM	Agentic Search (ours)
Q065: Weddell Sea, Winter 2013 (temporal)	Failure: Returns IceBird Winter 2019/2023 altimetry. Wrong year, wrong parameter.	Partial: Returns CTD profiles but from Summer cruises (ANT-XXIX/2, PS82).	Success: Returns ANT-XXIX/6 (AWECS) , the specific winter cruise (Jun–Aug 2013).
Q059: Satellite altimetry (source)	Failure: Returns lake water quality and biological transects.	Validation trap: Returns in-situ mooring data “used to validate” satellites.	Success: Returns Level-3/4 gridded SSHA from CryoSat-2, Envisat, TOPEX/Poseidon.
Q080: Polar glacier GPR (spatial)	Geo-drift: Returns GPR from Alpine glaciers (Austria/Italy). Matches “Glacier+Radar” but ignores “Polar”.	Mixed: Returns Antarctic data but includes Alpine mass balance stakes.	Success: Strictly polar. All 30 results from Antarctica or Greenland. Expands to RES (Radio-Echo Sounding).
Q096: Coastal erosion rates (intent)	Failure: Returns vegetation maps and oceanographic bottle data.	Intent miss: Returns raw aerial imagery subsets (18/30 results from one campaign).	Success: Returns derived shoreline retreat rates (Yukon, Laptev, Baltic, Portugal).

Appendix

Table S7: Head-to-head comparison of the top-ranked search result for four representative complex queries, illustrating qualitative failure modes in baseline methods that are resolved by the agentic architecture. Each query tests a distinct category of semantic reasoning: temporal constraint resolution, source vs. validation data disambiguation, geographic filtering, and product-level intent recognition.

Earth Science

See all 23 papers

SciDataSailor: Deep Scientific Data Exploring
Jul 29, 2026Jiyong Rao, Yicheng Qiu, Chi Zhang +2Scientific QA AI Agent Benchmarks
Shanghai Artificial Intelligence Laboratory · Tongji University
CLIMATEAGENT: Multi-Agent Orchestration for Complex Climate Data Science Workflows
Nov 25, 2025Chenyue Li, Hyeonjae Kim, Wen Deng +4AI for Science Multi-Agent Orchestration
1The Hong Kong University of Science and Technology
SciDER: Scientific Data-centric End-to-end Researcher
Mar 2, 2026Ke Lin, Owais Aijaz, Yilin Lu +3AI Agents for Scientific Discovery Scientific Workflow Automation
Data Science, College of William & Mary, Williamsburg, USA · Natural Language Processing, MBZUAI, Abu Dhabi, UAE · Computer Science, University of Minnesota, Minneapolis, USA +1

Jul 29, 2026J/K move · Enter open · S save

SciDataSailor: Deep Scientific Data Exploring

Jiyong Rao, Yicheng Qiu, Chi Zhang, Chunfeng Song+1

Shanghai Artificial Intelligence Laboratory · Tongji University

Scientific QA AI Agent Benchmarks LLM Agent Trajectory Synthesis Synthetic Data Generation for Language Models Data Analysis Agents

Open paper AskPDF arXiv

Scientific datasets are commonly organized as hierarchical repositories containing heterogeneous and interdependent files, making their inspection, integration, and analysis labor-intensive and reliant on domain expertise. Although large language model (LLM) agents have advanced substantially in planning, reasoning, and tool use, existing research has largely overlooked their ability to interact with real scientific data assets through executable environments. We introduce Deep Scientific Data Exploration, an agentic task paradigm in which agents navigate repositories, interpret heterogeneous files and schemas, execute analyses, integrate cross-file evidence, and produce conclusions grounded in executed observations. To operationalize this paradigm, we present SciDataSailor, a framework for synthesizing tool-interactive trajectories by balancing broad exploration with targeted exploitation. SciDataSailor instantiates trajectory synthesis as Monte Carlo Tree Search (MCTS) with four task-specific mechanisms: difficulty-stratified exploration seeds, dual-feedback first-play urgency, hierarchical strategy-to-tool action generation, and entropy-guided branching. Using this framework, we construct SciDataSailor-SFT-2K for supervised fine-tuning and SciDataSailor-Bench for evaluation, with the latter comprising 627 meta-information summarization tasks and 586 scientific question-answering tasks across 27 datasets spanning the life, earth, and physical sciences.

Abstract

Earth Science

Explore similar work

SciDataSailor: Deep Scientific Data Exploring

CLIMATEAGENT: Multi-Agent Orchestration for Complex Climate Data Science Workflows

SciDER: Scientific Data-centric End-to-end Researcher