PaperlayerView exact revision on arXiv

Source-linked AI summary

OceanMind: A multi-agent AI system for ocean diagnosis

Fan Zhang, Weicong Cheng, Yuheng Chen, Hiuseut Kung, Ying Zhang, Aixi Han, Quanjia Zhong, Can Yang, Jianping Gan

arXiv:2610.03780v1physics.ao-phcs.AI

TL;DR

Comprehensive 3D ocean data are difficult to turn into quantitative evidence because analysis demands substantial effort and expertise. OceanMind couples LLM agents with ocean states through skill-guided workflows and execution-informed replanning. It reports strong benchmark results and demonstrates reproduced diagnostics, hypothesis validation, and support for environmental decisions.

  • Problem

    Extracting quantitative evidence from evolving 3D ocean states requires substantial analytical effort, and comprehensive data alone do not provide scientific understanding or evidence.

    Source: Introduction · pp. 1–2

  • Method

    OceanMind coordinates specialized LLM agents across four analytical stages, using reusable ocean-specific skills and execution feedback to construct, run, and repair workflows.

    Source: Introduction · p. 2, Discussion and summary · p. 15

  • Results

    93.8% pass rate for statistical diagnostics and 75.0% for interpretation, compared with at most 64.6% and 25.0%, respectively, for ReAct baselines.

    Source: OceanMind outperforms competing AI… · p. 7

  • Takeaways & Limitations

    OceanMind reproduced published diagnostics, exposed evidence chains for hypothesis validation, and connected ocean hazards to spatially explicit management recommendations.

    Source: Discussion and summary · p. 15

    Conclusions depend on the quality and coverage of ocean state estimates, and LLM misinterpretation risks remain despite safeguards and the need for expert review.

    Source: Discussion and summary · p. 15

Abstract

from arXiv · show

Time-dependent, three-dimensional (3D) oceanic multi-variables define coherent states of the evolving ocean to facilitate ocean diagnosis and advance ocean science to better inform environmental and hazard management. However, extracting quantitative evidence from these variables requires substantial and complex analytical effort. We introduce OceanMind, a multi-agent AI system that directly couples large language models (LLMs) with comprehensive time-dependent 3D ocean states for swift and effective diagnosis. OceanMind organizes the analytical process into four coordinated complexity stages: Query Routing, Skill-based Planning, Tool Execution, and Evidence-based Summary Generation. Specialized agents interpret user requests, construct and execute multi-step computational workflows, and synthesize quantitative evidence. To ensure reliable workflow construction, 63 reusable ocean-specific analysis skills serve as procedural manuals that guide the LLM agent in selecting data, conducting diagnostics, and applying analytical tools. With reflection and replanning mechanisms that use execution feedback to repair invalid plans, OceanMind ensures reliable analysis workflows across diverse needs. On a benchmark of 240 computational-workflow queries spanning the four stages, OceanMind outperformed general ReAct agents with the same registered tool pool, achieving relative improvements of 41.2% in effectiveness and 21.5% in efficiency. Beyond the benchmark, OceanMind reproduced published oceanic diagnostics, validated hypotheses, and supported environmental decision-making over global oceans. Overall, OceanMind advances LLMs by integrating them with time-dependent 3D ocean analysis, enabling scientific interpretation and enhancing formulation of environmental policies based on quantitative ocean evidence.

Introduction

OceanMind addresses the difficulty of turning comprehensive, evolving 3D ocean data into reliable evidence, a task that demands substantial analytical expertise and effort. It couples LLM agents with ocean-state data and skill-guided workflows, then evaluates workflow reliability and ocean-diagnosis applications.

Introduction

  • Ocean-state analysis requires coherent views across space, depth, and time, but using diverse tools entails computational skills and substantial effort, limiting access for non-specialists.Experts must integrate data processing, visualization, hypothesis testing, and interpretation into coherent workflows.Source: Introduction · pp. 1–2
  • General and domain-adapted LLMs cannot directly analyze ocean states, while OceanAI lacks broad 3D physical-biogeochemical coverage and coupled-variable analysis.LLMs may also produce plausible but inaccurate ocean descriptions; linking variables and processes is described as critical for coherent diagnosis and forecasting.Source: Introduction · p. 2
  • OceanMind coordinates four stages—Query Routing, Skill-based Planning, Tool Execution, and Evidence-based Summary Generation—to turn requests into quantitative ocean evidence.Specialized agents interpret intent, construct and execute workflows, and synthesize results transparently.Source: Introduction · p. 2
  • OceanMind uses 63 reusable ocean-specific analytical skills to guide data selection, diagnostics, and compatible tool use in complex workflows.Reflection and replanning use execution feedback to repair failed operations, invalid intermediate results, or unresolved dependencies.Source: Introduction · p. 2
  • On 240 CMOMS queries, OceanMind achieved a 91.2% pass rate and 90.5% effective step efficiency across four complexity levels.Pass rates were 100%, 96%, 83%, and 91% from L1 to L4; ReAct had access to the same registered tool pool.Source: Introduction · p. 3

Results

The Results section examines OceanMind’s analyses of time-dependent 3D ocean states, covering benchmark performance, reproduced diagnostics, hypothesis testing, policy support, and analyses across regions and CMEMS products. It also presents quantitative tropical Pacific diagnostics produced through a transparent workflow accessible in the GUI.

Results

  • OceanMind achieved a 91.2% pass rate and 90.5% effective step efficiency, versus 64.6% and 74.5% for ReAct(Qwen), relative improvements of 41.2% and 21.5%, respectively.Across Claude Sonnet 4.5, Qwen3 Max, and DeepSeek V4 Pro, pass rates/effective step efficiencies were 91.2%/90.5%, 88.8%/89.1%, and 91.2%/89.4%, respectively. OceanMind’s pass rates on L3/L4 tasks were 83.3%/90.6%, 85.7%/75.0%, and 84.5%/84.4% with those backends, versus 63.1%/28.1% for the strongest ReAct baseline.Source: OceanMind outperforms competing AI… · pp. 5, 7, Figure 2 · p. 6

OceanMind reproduces published ocean diagnostics with independent observational verification

  • OceanMind reproduced Gan et al.’s connected China Seas circulation diagnostics and found that generated fields followed independent CTD and satellite-observed spatial patterns.Nineteen bottom CTD observations followed the major gradients in the generated January 2018 fields; CMOMS and GHRSST MUR SST climatologies also showed the large-scale winter gradient and summer warming. The reproduced diagnostics linked circulation fields to variable exchange pathways, including a Luzon Strait mean transport of 1.245 Sv.Source: OceanMind reproduces published ocean… · pp. 7, 9, Figure 3 · p. 8

OceanMind enables traceable validation of scientific hypotheses

  • An 11-step analysis found Coastal Fresh Surface Water in 34.5% of East China Sea bloom-hotspot tiles versus 13.0% of the regional background, supporting—but not proving—the hypothesis.Warm Salty Surface Water accounted for 7.0% of hotspot tiles versus 42.3% of the background. ChatGPT offered a literature-based explanation without quantitative analysis of the fields, while OceanAI returned a server error in this run.Source: OceanMind enables traceable validation… · pp. 10–11, Figure 4 · p. 10

OceanMind supports evidence-based policy making

  • For Yellow Sea management, OceanMind recommended targeted oxygen monitoring and seasonal operational adjustments, treating stratification as context rather than evidence of increasing risk.From 2020 to 2022, area-weighted bottom oxygen declined at −9.15 mmol m−3 yr−1 (p < 0.001), and a hotspot had up to 368 hypoxic days. The squared buoyancy frequency anomaly trend was not significant (p = 0.209); recommendations included monitoring near the hotspot and reviewing stocking, feeding, and harvest timing during warm-season risk windows.Source: OceanMind supports evidence-based… · pp. 11, 13, Figure 5 · p. 12

OceanMind can generalize across ocean products and regions

  • OceanMind applied the same natural-language workflow to CMEMS products in the Gulf of Mexico and tropical Pacific, extending diagnostics beyond the China Seas.The Gulf analysis included circulation, heatwaves, and hypoxia-like events; its source data had inconsistent metadata, and the relaxed 150 mmol m−3 threshold identified 415 events. In the tropical Pacific, the Niño 3.4 anomaly reached +3.09°C on 1 December 2015, while SST trends were positive in the western Pacific warm pool, Niño 4, and Niño 3.Source: OceanMind can generalize across ocean… · p. 13
  • The two Niño-region trends had R² < 0.02; the first EOF mode explained 82.6% of EI Niño 3.4 anomaly variance, and the second explained 9.9%.The first-mode principal component peaked during the 2015–2016 EI Niño, while the second mode exhibited a west–east dipole.Source: OceanMind can generalize across ocean… · p. 14
  • OceanMind applied across ocean products and regions, with Gulf of Mexico full-depth and tropical Pacific surface-layer diagnostics obtained through a transparent data-to-interpretation sequence in the GUI.The demonstrations included Gulf of Mexico current, eddy, marine heatwave, and bottom-oxygen diagnostics, plus tropical Pacific SST trends and EOF modes.Source: Figure 6 · p. 14, OceanMind can generalize across ocean… · p. 15

Discussion and summary

OceanMind couples specialized agents and ocean-state data to produce transparent, inspectable analyses. Its tests cover benchmark performance, scientific hypothesis validation, management support, and limitations.

Discussion and summary

  • OceanMind couples LLMs with time-dependent 3D ocean states to diagnose conditions, validate hypotheses, and support evidence-based policymaking.Its GUI exposes analysis plans, execution progress, and intermediate results.Source: Discussion and summary · p. 15
  • OceanMind uses 63 reusable ocean skills and execution-feedback replanning to compose workflows, particularly for complex L3 and L4 tasks.A deterministic ToolOrchestrator executes workflows from a pool of 125 registered tools.Source: Discussion and summary · p. 15
  • The test cases reproduced circulation diagnostics, exposed evidence chains in hypothesis tests, and connected marine hazards to spatially explicit management recommendations.Global workflows transferred to other regions after data were converted to compatible variables, coordinates, time ranges, and depth dimensions.Source: Discussion and summary · p. 15
  • Conclusions are bounded by ocean-state estimate quality and coverage, possible LLM misinterpretation or overstatement, and the system's supported variables.The safeguards do not replace expert review.Source: Discussion and summary · p. 15
  • Future development priorities include stronger uncertainty handling, broader data coverage, and tighter coupling with numerical ocean modelling.The paper proposes integrating ensemble estimates, satellite products, and field observations to assess uncertainty and distinguish robust signals from product-specific artifacts.Source: Discussion and summary · p. 15
  • The authors position OceanMind as a foundation for reusable analysis of ocean state estimates supporting scientific discovery and environmental decision-making.The conclusion frames this as a move from text-based assistance toward executable and traceable analysis.Source: Discussion and summary · p. 16

Methods

OceanMind converts natural-language ocean questions into diagnostic workflows that execute over gridded state data and produce evidence-grounded scientific answers. Its architecture combines specialized agents, registered tools, a GUI, and partition-aware computation.

Architecture of OceanMind

  • The four stages route requests, plan skill-based workflows, execute computations with validation and revision, and synthesize scientific answers from computed evidence.Query Routing determines whether the active dataset is needed; summary generation integrates completed steps and visual evidence.Source: Architecture of OceanMind · p. 16
  • Six LLM agents and a deterministic ToolOrchestrator divide routing, planning, tool execution, evidence synthesis, and multi-turn memory functions.A CodingAgent can generate a restricted Python step when no registered tool covers a required diagnostic operation.Source: Architecture of OceanMind · p. 16
  • The study configuration contains 63 reusable ocean skills and 125 registered tools, which guide workflow construction and support data access, diagnostics, visualization, and evidence assembly.Skills describe applicable scenarios, inputs, parameters, permitted tools, and execution strategy.Source: Architecture of OceanMind · p. 16
  • The GUI lets users inspect plans, execution progress, and intermediate results, while partition-aware parallel computation streams progress during large workflows.OceanMind keeps loaded subsets as lazy partitioned arrays when possible and merges partition outputs into normalized results.Source: Architecture of OceanMind · p. 16

Query Routing

Query Routing decides whether a request needs the active ocean dataset or can be handled directly. It uses query and context information to select the appropriate response pathway.

Query Routing

  • Query Routing determines whether the active ocean state estimate is required for a request.It considers the query, dataset context, interface inputs, and recent conversation context.Source: Query Routing · p. 17
  • Dataset computation, inspection, and metadata requests are directed to the dataset-analysis pathway.Source: Query Routing · p. 17
  • General ocean-science questions, editing requests, and other requests not requiring the active dataset go to the AnswerAgent for a direct answer or requested external search.Source: Query Routing · p. 17

Skill-based planning

Skill-based planning translates an analytical request into a traceable workflow by resolving data scope, selecting relevant skills, and organizing their operations. Planning is constrained by dataset context and tool contracts.

Skill-based planning

  • OceanMind's 63 skills document scientific purpose, applicable and excluded scenarios, evidence requirements, inputs, compatible tools, and expected outputs.The library spans diagnostics including time series, profiles, event detection, transport, EOF analysis, and environmental health assessment.Source: Skill-based planning · p. 17
  • The PlannerAgent resolves variables, region, time window, season, and vertical selection, then chooses the skill or minimal skill set needed to generate evidence.It organizes operations into a traceable workflow with named intermediate and final artifacts.Source: Skill-based planning · p. 17
  • If a required input cannot be inferred from the query, interface, dataset metadata, or conversation, the PlannerAgent requests clarification rather than inventing a parameter.Planning is constrained by skill workflows, dataset metadata, and registered tool contracts.Source: Skill-based planning · p. 17

Tool execution

Tool execution runs planned workflows against ocean-state estimates while validating tool calls and passing structured results between steps. When registered tools do not cover an analysis, a constrained CodingAgent can add a validated Python step.

Tool execution

  • Tool names, arguments, and references are checked against contracts, with scalar and array parameters normalized where possible and invalid parameters reported before execution.These checks prevent unsupported arguments or broken references from being silently propagated.Source: Tool execution · p. 17
  • The ToolOrchestrator resolves dependencies, calls Python tools, normalizes and caches outputs, and wraps ocean fields and derived datasets as metadata-bearing artifacts.It also summarizes structured outputs for the interface and synthesis while retaining full numerical objects for downstream tools.Source: Tool execution · p. 17
  • When registered tools do not cover an analysis, a CodingAgent designs a restricted Python step from specified inputs, schemas, output type, and diagnostic operation.Generated code must use approved scientific libraries, define run(inputs, params), and return a normalized result matching the output contract; execution validates and stores it as a standard artifact.Source: Tool execution · p. 17
  • Workflows pass evidence forward as structured artifacts, such as loading bottom oxygen, detecting hypoxia, mapping oxygen-deficit burden, and extracting regional statistics for assessment.Dynamic workflows can also use current-derived vorticity or transport diagnostics in mechanism ranking or evidence synthesis.Source: Tool execution · pp. 17–18

Evidence-based summary generation

OceanMind grounds synthesis in compact evidence packets and completed computations. Its synthesizer distinguishes computed evidence from indirect support and data gaps, while framing policy statements as scientific decision support.

Evidence-based summary generation

  • Evidence packets expose interpretation-relevant quantities, including spatial extrema, trend slopes, event counts, burden values, and mechanism scores, rather than full numerical arrays.This keeps synthesis grounded in completed computations while limiting the raw data passed to the language model.Source: Evidence-based summary generation · p. 18
  • The ResultSynthesizer combines the query, workflow, completed step records, result summaries, and optional evidence packets into schema-validated outputs such as findings, limitations, evidence tables, and policy guidance.For integrated environmental-health and suitability questions, an assessment profile distinguishes computed evidence, indirect support, and data gaps.Source: Evidence-based summary generation · p. 18
  • Descriptive claims report computed quantities, mechanism claims connect only executed diagnostics, and policy statements are framed as scientific decision support rather than legal, economic, or operational prescriptions.The design preserves a transparent link between conclusions and the computed evidence chain.Source: Evidence-based summary generation · p. 18

Reflection and replanning

OceanMind uses execution feedback to repair workflows when planned evidence chains fail, while bounding retries and preserving valid intermediate work. If no valid repair is available, it stops rather than produce an unsupported interpretation.

Reflection and replanning

  • Runtime tool failures, invalid CodingAgent artifacts, and unresolved downstream dependencies trigger feedback containing the failed step, error, previous plan, and completed artifacts.The PlannerAgent repairs the workflow only while replanning budget remains and the original scientific intent can be preserved.Source: Reflection and replanning · p. 18
  • Replanning is bounded, repaired workflows undergo the initial plan's validation, and valid intermediate artifacts remain available to avoid repeating successful computation.If no valid repair is available, OceanMind returns an explicit failure message instead of an unsupported interpretation.Source: Reflection and replanning · p. 18

Graphical user interface (GUI) and streamed interaction

OceanMind presents routes, workflow progress, evidence, and synthesis through a web application, with streamed updates that let users inspect intermediate results. A MemoryAgent retrieves relevant context before planning and synthesis and records reusable information after completed turns.

Graphical user interface (GUI) and streamed interaction

  • The web application displays the route, planned workflow, step progress, result cards, visual evidence, source cards, and final synthesis, alongside manual visualization controls for variable, region, time range, and vertical mode.Agent queries use the same data and rendering infrastructure as manual visualization after workflow execution.Source: Graphical user interface (GUI) and… · p. 18
  • Before planning and synthesis, the MemoryAgent selects relevant conversation, project, or reference memory, then records reusable definitions, preferences, artifacts, and workflow patterns after completed turns when appropriate.Explicit constraints in the current query take precedence over memory.Source: Graphical user interface (GUI) and… · pp. 18–19
  • Streaming sends routing, memory, planning, execution, replanning, and result events; linked result cards with maps or charts can appear before final synthesis is complete.Users can inspect intermediate evidence while interpretation is being generated, and completed conversations can be exported as reports.Source: Graphical user interface (GUI) and… · p. 19

Benchmark and evaluation protocol

The benchmark used 240 ocean-analysis queries spanning four complexity levels, with shared tools and scoring rules across system configurations. Its protocol assessed task completion, diagnostic fidelity, and tool-step efficiency.

Benchmark and evaluation protocol

  • 240 queries covered four complexity levels, from direct extraction and simple statistics to multi-step diagnostics and interpretation for policy questions.The suite contained 24 L1, 100 L2, 84 L3, and 32 L4 queries, each paired with a canonical gold tool chain.Source: Benchmark and evaluation protocol · p. 19
  • All configurations received the same 240 queries, 125 registered tools, and scoring protocol, with at most two execution-triggered repair or retry attempts.The main comparison used Claude Sonnet 4.5 for OceanMind and ReAct(Claude), and Qwen3 Max for ReAct(Qwen).Source: Benchmark and evaluation protocol · p. 19
  • A run passed only if it produced the requested result, completed a successful answer-bearing tool path, preserved query specifications, and delivered diagnostics sufficient for the skill.OceanMind traces also had to include selector and planner stages.Source: Benchmark and evaluation protocol · p. 19
  • Effective step efficiency compared executed core tool calls with the gold-chain reference count and penalized both missing and unnecessary steps.First-pass rate counted passing runs completed without a retry.Source: Benchmark and evaluation protocol · p. 19
  • Qualitative comparisons used identical query text and one recorded run per system; web responses were collected on 3 August 2026 without manual rewriting.Sampling controls were unavailable in the web interfaces and remained at provider defaults.Source: Benchmark and evaluation protocol · p. 19

Data availability

The study used CMOMS and CMEMS ocean-state estimates. CMEMS subsets supported applications in the Gulf of Mexico and northern Caribbean and the tropical Pacific.

Data availability

  • CMOMS is a validated, three-dimensional coupled physical-biogeochemical model for the China Seas and adjacent western Pacific.It was developed by the Ocean Dynamics and Modeling Program at HKUST and used in previous China Seas studies.Source: Data availability · p. 20
  • CMEMS data were publicly available through the Copernicus Marine Data Store and extracted using the Copernicus Marine Toolbox.The CMOMS data were available through the Ocean Dynamics and Modeling Program website.Source: Data availability · pp. 19–20
  • The Gulf of Mexico and northern Caribbean application used CMEMS temperature, salinity, currents, chlorophyll-a, and dissolved oxygen fields from April 2025 to March 2026.The subset covered 98.68°W–76.98°W, 16.35°N–32.02°N, and depths from 0.494 to 5727.917 m.Source: Data availability · p. 20
  • The tropical Pacific SST application used CMEMS temperature data over 120°E–80°W and 10°S–10°N from 2005 through 2025.The product catalogue records and identifiers were accessed on 4 August 2026.Source: Data availability · p. 20

Related papers

Loading 2610.03780v1…