1 Introduction↩︎

Trace elements and their isotopes (TEIs) in the ocean serve as vital diagnostic tools for deciphering Earth system dynamics. Among them, lead (Pb) and its isotopes are particularly important, functioning as powerful tracers for ocean circulation pathways and anthropogenic pollution history [1], [2]. Since Patterson’s seminal work [3] revealed global industrial Pb contamination, marine Pb has been extensively monitored to capture the temporal evolution of anthropogenic emissions.

However, the inherent scarcity of high-quality marine Pb data creates a significant bottleneck for global-scale analysis. Obtaining these records requires expensive oceanographic cruises and strict trace-metal-clean protocols [4], [5], making every data point a high-value asset. While a substantial volume of historical in-situ observations has accumulated over decades, the vast majority remain buried within the unstructured text, tables, and figures of academic papers. This phenomenon has created a “data silo” problem, where critical datasets are scattered and effectively unavailable for global-scale synthesis.

Prior works. Existing efforts to bridge this gap have primarily relied on two integration paradigms: manual curation and automated extraction. Manual curation, exemplified by the GEOTRACES program [6] and analogous synthesis efforts across paleobiology [7] and terrestrial biogeochemistry [8], [9], ensures high data fidelity but is labor-intensive and unscalable against the exponential growth of literature. Conversely, automated approaches—spanning discriminative models [10], domain-specific LLMs [11], [12], multimodal document parsing [13], [14], and retrieval-augmented generation (RAG) [15], [16]—offer scalability but often treat scientific extraction as a probabilistic text task. Lacking domain-specific logical constraints, these methods frequently misinterpret fine-grained details—such as isotope notations or unit conversions—resulting in hallucinations that compromise scientific validity.

Challenges. To achieve both scalability and scientific rigor, effective automation requires addressing two fundamental barriers: (1) The Domain Knowledge Gap: General-purpose LLMs lack the specialized knowledge and strict logical constraints required for scientific research. Without domain-specific guidance, these models often hallucinate or fail to distinguish between subtle scientific nuances, rendering their outputs unreliable for high-stakes analysis. (2) Structural Heterogeneity: Scientific records are deeply embedded within unstructured text and diverse table formats. Extracting this fine-grained data requires a sophisticated, systematic workflow capable of navigating these complex structures, rather than simple information retrieval.

Our Work. To address these challenges, we propose Compass, a Knowledge Tree-enhanced LLM Agent framework designed to operationalize an expert-guided adaptation. Instead of relying on black-box fine-tuning, Compass adopts an expert-in-the-loop philosophy. We co-designed a domain-specific Knowledge Tree with marine scientists, encoding rigorous reasoning logic to guide the automated process. Leveraging this structured guidance, Compass decomposes the integration task into hierarchical subtasks across three phases—collection, extraction, and aggregation—with validation checks embedded at each step to ensure that every extracted record satisfies physical constraints. By deploying Compass on over 230,000 relevant open-access papers, we successfully recovered 3,751 previously unincorporated Pb records, demonstrating the framework’s capability to bridge the gap between everyday language models and rigorous scientific demands. The main contributions of this work are summarized as follows:

• We develop Compass, a Knowledge Tree-enhanced LLM Agent framework that hierarchically decomposes complex scientific workflows, enabling accurate extraction and integration of fine-grained data (text and tables) from heterogeneous academic sources.

• We demonstrate that expert-guided knowledge injection via a co-designed Knowledge Tree can effectively bridge the gap between general-purpose LLMs and high-stakes scientific domains without fine-tuning, providing a practical and low-cost alternative to domain-specific model training.

• We deploy Compass to establish the largest integrated marine Pb database to date, recovering 3,751 previously unincorporated records from over 230,000 open-access papers. The integrated data substantially expand coverage in under-sampled regions such as the East China Sea and the Southern Ocean, providing an enriched data foundation for future scientific discoveries.

2 Related Work↩︎

2.1 Scientific Data Integration↩︎

Integrating heterogeneous scientific data is a cornerstone of data-driven discovery. Existing approaches fall into two paradigms. (1) Manual Integration. Research programs such as GEOTRACES [6] and regional syntheses [17] enable systematic integration of marine trace element data through international collaboration. Fan et al. [7] manually synthesize data from thousands of marine fossil species to generate high-resolution biodiversity curves. Similar efforts in terrestrial biogeochemical cycles [8], [9], [18] yield important insights but require substantial human effort, becoming increasingly impractical with the exponential growth of publications. (2) Automated Integration. Text-based extraction methods have evolved from discriminative models like SciBERT [10] to structured extraction with domain-specific and multi-expert language models [19]–[21]. Cross-modality frameworks such as TabPedia [14], MinerU [13], MatViX [22], and Uni-SMART [23] extend beyond text to support joint analysis of tables and figures, with applications spanning document extraction to clinical and materials science data integration [24], [25]. Despite this progress, existing automated methods primarily focus on information extraction from individual documents rather than end-to-end scientific synthesis with physical constraint validation.

2.2 Scientific LLM Agents↩︎

LLM agents leverage advanced comprehension and reasoning to automate scientific workflows. (1) Training-based Agents. Domain-specific LLMs such as K2 [11] for geoscience and OceanGPT [12] for ocean science improve domain knowledge through curated corpora and supervised training. Agent frameworks further employ curriculum learning and reinforcement learning—for example, AutoWebGLM [26] integrates both to improve web navigation for scientific tasks. Multi-agent strategies [27]–[32] enable large-scale coordination across domains but typically demand significant computational resources and domain-specific datasets. (2) Prompting and Retrieval-based Agents. PaperQA [16] employs retrieval-augmented generation (RAG) to perform evidence-based information retrieval across scientific articles. ChemAgent [33] and SciAgents [34] leverage tool-use and ontological knowledge graphs for multi-agent reasoning, while other approaches explore self-updating mechanisms and automated environment generation [35], [36]. However, standard RAG or general agents lack structured domain logic for fine-grained integration in high-stakes scientific fields. Compass addresses this gap through an expert-guided adaptation approach: instead of probability-based retrieval or costly fine-tuning, it embeds structured expert knowledge directly into the agent’s workflow, ensuring scientific validity while maintaining scalability.

3 Background and Problem Formulation↩︎

3.1 Scientific Significance of Marine Lead↩︎

Marine lead (Pb) and its isotopes serve as critical tracers for ocean circulation pathways and anthropogenic pollution history [1], [2]. Pb primarily originates from anthropogenic activities such as leaded gasoline combustion and industrial emissions [37], dispersing via atmospheric transport to even remote ocean regions [38] (Figure 3). Its relatively brief oceanic residence time compared to global mixing timescales enables Pb to retain distinct regional isotopic fingerprints that record transient environmental changes [39], [40], while obtaining high-quality measurements requires expensive oceanographic cruises and strict trace-metal-clean protocols [4], [5].

However, individual studies provide only geographically limited snapshots. A comprehensive integrated database unlocks analyses no single study can achieve: global Pb inventory and distribution estimation, source identification via stable isotope ratios (e.g., \(^{206}\)Pb/\(^{207}\)Pb, \(^{208}\)Pb/\(^{206}\)Pb), and investigation of oceanic responses to the global phase-out of leaded gasoline. Multi-decadal observations, when unified, further support biogeochemical modeling and environmental policy evaluation. Despite this potential, vast historical data remain fragmented across decades of publications. Manual curation programs like GEOTRACES [6] and regional syntheses [17] are fundamentally unscalable, directly motivating the integration framework presented in this work.

3.2 The Data Integration Challenge↩︎

Data Source Categorization. Marine Pb data exists across three categories: structured databases (\(\mathcal{D}_{structured}\)) such as GEOTRACES [6] with well-defined schemas; scattered datasets (\(\mathcal{D}_{scattered}\)) from individual studies with varying conventions; and academic papers (\(\mathcal{D}_{papers}\)), where fine-grained measurements are embedded in unstructured text and diverse table formats.

Technical Challenges. While the general challenges of domain knowledge gap and structural heterogeneity apply broadly, marine Pb data integration presents additional domain-specific difficulties: (1) Data Source Proliferation: Pb measurements scatter across literature spanning over five decades, making manual integration impractical. (2) Fine-grained Semantic Complexity: Subtle distinctions—such as dissolved versus particulate phases, different isotope ratio conventions, and varied analytical methods—are easily confused by general-purpose models. (3) Cross-source Harmonization: Inconsistent units, depth references, and coordinate formats demand systematic normalization.

Problem Formulation. Each source exhibits distinct schema structures, semantic conventions, and quality standards (e.g., varying units, precision levels, and coordinate formats), necessitating a hierarchical pipeline that progressively resolves these heterogeneities. We formalize the data integration objective as constructing a unified dataset \(D^*\) from heterogeneous sources \(\mathcal{D} = \{D_1, D_2, \ldots, D_n\}\) through a hierarchical processing pipeline: \[\begin{align} D^* = \mathcal{A}(\mathcal{E}(\mathcal{C}(\mathcal{D}_{papers})), \mathcal{D}_{structured}, \mathcal{D}_{scattered}), \end{align}\] where \(\mathcal{C}\) denotes the collection function that identifies relevant papers from a large candidate corpus, \(\mathcal{E}\) denotes the extraction function that resolves schema heterogeneity and retrieves structured records from paper tables, and \(\mathcal{A}\) denotes the aggregation function that integrates all sources with standardized units, formats, and quality controls. \(\mathcal{D}_{papers} \subset \mathcal{D}\), \(\mathcal{D}_{structured} \subset \mathcal{D}\), and \(\mathcal{D}_{scattered} \subset \mathcal{D}\) represent the academic paper sources, structured database sources, and scattered dataset sources, respectively. This formulation directly motivates the three-phase architecture of Compass.

4 Compass Framework↩︎

4.1 Expert-Guided Adaptation Paradigm↩︎

Applying LLMs to scientific data integration requires bridging the gap between general-purpose language understanding and rigorous domain reasoning. Existing strategies each have fundamental limitations for this goal. Fine-tuning [11], [12] improves domain knowledge but incurs high computational cost and often degrades the instruction-following capabilities critical for multi-step workflows. Retrieval-Augmented Generation (RAG) [15] injects knowledge at inference time but treats it as unstructured text fragments, lacking mechanisms to enforce logical constraints or sequential validation across complex pipelines. Knowledge graph (KG) and ontology-based approaches [34] organize domain concepts and semantic relationships, but primarily capture declarative knowledge (what is) rather than procedural knowledge (how to act).

Compass introduces an expert-guided adaptation paradigm that encodes domain expertise as a structured, executable Knowledge Tree co-designed with domain scientists. Unlike the above approaches, the Knowledge Tree integrates operational logic—prescribing not only what the agent should know but also how it should reason and validate its outputs—making the workflow interpretable, controllable, and scientifically trustworthy.

4.2 Knowledge Tree: Design and Construction↩︎

Definition (Knowledge Tree). A Knowledge Tree \(T = (N, H, K)\) is a hierarchical structure where \(N\) represents the set of nodes corresponding to specific tasks or subtasks, \(H\) denotes directed edges representing parent-child decomposition relationships, and \(K\) contains expert-provided knowledge constraints associated with each node. Each node \(n_{i,j} \in N\) at level \(L_i\) corresponds to subtask \(t_{i,j}\) and can be decomposed into child nodes at level \(L_{i+1}\), with leaf nodes representing atomic tasks requiring direct execution.

Knowledge Structure. For each node \(n_{i,j}\), the associated knowledge \(K_{i,j} = \{BK, LC, OG, VC\}\) encompasses four complementary dimensions:

• Background Knowledge (\(BK\)): Domain concepts, terminology, and contextual information providing foundational understanding for task execution.

• Logical Constraints (\(LC\)): Explicit rules, conditional dependencies, and logical requirements defining valid execution paths.

• Operational Guidelines (\(OG\)): Step-by-step procedures, decision heuristics, and output format specifications ensuring standardized execution.

• Validation Criteria (\(VC\)): Quality metrics, physical constraint checks, and verification steps enabling automated assessment of output correctness.

This four-dimensional structure ensures that each node carries not only declarative knowledge but also procedural and evaluative guidance, distinguishing our approach from flat structured prompting or static ontologies.

Construction Protocol. The Knowledge Tree is constructed offline by AI researchers in collaboration with domain experts, following a four-step protocol for each hierarchical level \(L_i\):

1. Task Context Initialization: AI researchers present each subtask along with its hierarchical dependencies, providing domain experts with comprehensive understanding of the task scope and requirements.

2. Expert Knowledge Elicitation: Domain experts provide specialized knowledge across the four dimensions (\(BK\), \(LC\), \(OG\), \(VC\)) based on their domain expertise and practical experience.

3. Knowledge Structuring: AI researchers organize expert-provided knowledge into structured knowledge nodes, ensuring consistency for subsequent automated processing.

4. Completeness Validation: Domain experts review the constructed knowledge structure for comprehensive coverage of domain requirements before proceeding to the next level.

This four-step protocol is applied iteratively for each hierarchical level, with successive rounds refining the knowledge structure based on expert feedback and agent execution results on representative samples. The finalized Knowledge Tree is then integrated into Compass to guide agent deployment.

4.3 Marine Pb Knowledge Tree↩︎

We collaborate with marine scientists to construct the Knowledge Tree for marine Pb data integration. Figure 4 shows the resulting structure. The overall task is decomposed into hierarchical subtasks, with Pb-related papers classified into fine-grained categories for accurate identification. Marine Pb data encompasses multiple types—including Pb concentrations (dissolved, particulate), \(^{210}\)Pb activity concentrations, and six Pb isotope ratios—each requiring specific data formats, transformation logic, and validation standards. For each subproblem, marine scientists provide the corresponding \(BK\), \(LC\), \(OG\), and \(VC\) knowledge.

In practice, the Knowledge Tree construction is a one-time effort: the marine Pb tree (approximately 20 nodes) was built in 6–7 hours by two marine scientists in collaboration with AI researchers.

4.4 Agent Deployment↩︎

When LLMs directly process complex data integration tasks, they often produce incomplete execution and logical inconsistencies, struggling to maintain coherent reasoning across interdependent subtasks. Inspired by the analytic hierarchy process [41] and hierarchical decision trees [42], Compass systematically decomposes complex tasks into a three-phase hierarchical pipeline with five specialized components, as shown in Figure 5.

(1) Collection Phase. The paper classification component retrieves candidate papers from academic repositories (e.g., Semantic Scholar [43], AceMap [44]) via keyword search, then applies LLM-powered semantic analysis guided by the Knowledge Tree to identify papers containing target data.

(2) Extraction Phase. The table classification component integrates the PDF parsing tool MinerU [13] to extract tabular content, then leverages LLMs to analyze tables with contextual information (captions, in-text descriptions) to classify and filter target data tables.

(3) Aggregation Phase. Three components handle different integration aspects: the data association component links related elements within papers (e.g., matching Pb measurements with sampling coordinates from separate tables); the cross-source fusion component normalizes heterogeneous table headers into a unified schema; and the data standardization component converts values into consistent units through LLM-generated conversion functions.

For each component, the corresponding Knowledge Tree node \(K_{i,j}\) is translated into a domain-aware prompt that systematically integrates all four knowledge dimensions (\(BK\), \(LC\), \(OG\), \(VC\)); the complete prompts are publicly available in our GitHub repository, ensuring that the LLM operates with domain-specific context, constrained reasoning boundaries, standardized output formats, and built-in quality checks. A record of completed subtasks and their outputs is maintained across hierarchical levels to preserve consistency. While this study focuses on text and tables, the modular architecture is designed to accommodate additional tools as supplementary components in future extensions. Formally, Compass operates as a state transition system where each completed subtask represents a distinct state, with transitions governed by the corresponding Knowledge Tree constraints, ensuring that domain expertise directs every decision point while maintaining systematic progression through the task hierarchy.

4.5 Validation Mechanism↩︎

Scientific data integration demands rigorous quality assurance. Compass implements a multi-layered validation mechanism informed by the Validation Criteria (\(VC\)) in the Knowledge Tree.

Automated Physical Constraint Checks. Every extracted record undergoes value range verification against physically plausible bounds for each data type, unit conversion cross-validation, and geographical outlier filtering against ocean boundary definitions. When validation failures are detected, Compass rolls back the current step and re-processes the failed subtask, preventing error propagation to downstream components.

Expert Manual Validation. Beyond automated checks, we conduct manual validation with marine scientists on a randomly selected subset, providing an independent assessment of end-to-end accuracy. Results are reported in Section 5.2.2.

5 Experiments and Results↩︎

5.1 Pre-Deployment Evaluation↩︎

Before full-scale deployment, we conduct a comprehensive evaluation of Compass to validate its accuracy on marine Pb data integration. We present detailed experimental analyses aimed at addressing the following three research questions (RQs):

• RQ1: Can Compass accurately integrate target marine Pb data, and how does it compare with general-purpose LLMs?

• RQ2: What are the advantages of Compass over domain-specific fine-tuned LLMs in terms of performance and cost?

• RQ3: How does each core component contribute to overall performance?

5.1.1 Experimental Setup↩︎

We define the evaluation task, describe the benchmark dataset, and specify the metrics for evaluation.

Task Overview. To evaluate whether Compass can accurately perform scientific data integration, we structure the evaluation into three core experimental tasks. This stepwise design mirrors the typical workflow used by researchers:

• Paper Classification: Identify whether a given paper belongs to the specific subfield relevant to the data integration task.

• Table Classification: Identify whether a specific table within a paper includes target data for extraction.

• End-to-End Extraction: Extract all target data points from the entire paper collection.

Benchmark Dataset. To evaluate extraction performance, we construct a benchmark dataset distributed across 10 categories (Table 1). “Marine Pb”, “Atmospheric Pb”, and “Terrestrial Pb” are used for the Paper Classification task, as these reflect key environmental compartments relevant to Pb research. Within Marine Pb, we distinguish three target subcategories—Marine Pb conc., Marine \(^{210}\)Pb, and Marine Pb isotope ratios—whose associated tables provide the target data for table classification and end-to-end data extraction tasks. Tables containing target marine Pb data serve as positive examples, while tables from the remaining seven categories act as negative controls to assess classification robustness. We curated a total of 337 tables, among which 63 contain target marine Pb data, yielding 1,397 data points for downstream extraction tasks.

Metrics and Configuration. We employ accuracy (Acc), precision (Prec), recall (Rec), and F1-score for comprehensive quantitative evaluation. All models use greedy decoding (temperature=0) for deterministic and reproducible results. Open-source LLMs are loaded with 4-bit NF4 quantization (BitsAndBytes) with fp16 inference. Maximum generation length is set to 32 tokens for classification and 2,048 tokens for extraction tasks. All experiments are conducted on two NVIDIA RTX 3090 GPUs (24GB VRAM each).

Baselines. As there are no agent-based approaches for scientific data integration, we compare Compass with two categories of baselines. General-purpose LLMs: GPT-4o, Gemini-2.5-pro, Llama-3.1-8B-Instruct, Qwen3-8B, and Qwen2.5-32B-Instruct, representing state-of-the-art proprietary and open-source models. Domain-specific LLMs: K2 [11], a Llama-based model pretrained on geoscience data (limited to 2,048-token context), and OceanGPT [12] (Qwen2-7B-based), fine-tuned on ocean science data. To ensure fairness, all baselines are evaluated using identical, iteratively refined prompts (available in our repository).

5.1.2 Results and Analysis↩︎

Overall Performance (RQ1). As shown in Table [tab:main95result], Compass achieves the highest scores across all tasks, exceeding 90% accuracy in both classification tasks. In end-to-end extraction, Compass (32B) achieves 0.465 F1, outperforming GPT-4o (0.373) and Gemini-2.5-pro (0.404). The advantage is most pronounced in table classification, where Knowledge Tree guidance enables accurate distinction between target Pb tables and scientifically irrelevant ones. While Gemini-2.5-pro achieves higher recall (0.511 vs.), its substantially lower precision (0.334 vs.) reflects a tradeoff unsuitable for scientific database construction, where data integrity is paramount.

Advantages over Fine-tuned Models (RQ2). Domain-specific models perform poorly despite their domain training: K2 fails due to context length limitations, while OceanGPT achieves only 0.009 F1 in extraction. Fine-tuning enhances domain vocabulary but degrades the instruction-following capabilities essential for multi-step integration. Compass preserves these capabilities while injecting expertise through the Knowledge Tree, avoiding costly retraining.

Ablation Study (RQ3). Table [tab:ablation] shows that removing tree structured logic reduces end-to-end F1 from 0.465 to 0.402, while removing knowledge nodes reduces it to 0.381. The tree structure ensures systematic task decomposition, while expert knowledge enables scientifically grounded decisions—their combination yields synergistic improvements. The higher recall but lower precision of the ablated variant reflects a precision–recall tradeoff: less structured extraction recovers more records at the cost of data integrity. Enabling automated validation rollback—which directly exercises the Validation Criteria (\(VC\)) dimension—further improves end-to-end F1 to 0.619, demonstrating \(VC\)’s contribution to overall reliability. Rollback results are reported here rather than in Table [tab:main95result], as baselines are evaluated under single-pass extraction for fair comparison.

5.2 Deployment Results↩︎

We deploy Compass with the Qwen2.5-32B backbone for full-scale marine Pb data integration, selected for its open-source nature and favorable performance-cost balance.

5.2.1 Integration Overview↩︎

We retrieve over 230,000 open-access papers using Pb-related keywords. Compass hierarchically processes these papers and identifies 110 containing target marine Pb data, extracting 3,751 new records. Combined with 12,704 records from 46 public datasets and 19,108 records from GEOTRACES [6], the integration yields a total of 35,563 marine Pb records.

Compared to the 19,108 records in GEOTRACES—the most authoritative existing data product—our effort introduces 16,455 new records, representing an 86% increase in available data. We categorize all records into eight types: Pb concentration, \(^{210}\)Pb concentration, and six isotope ratios (\(^{206}\)Pb/\(^{204}\)Pb, \(^{207}\)Pb/\(^{204}\)Pb, \(^{208}\)Pb/\(^{204}\)Pb, \(^{206}\)Pb/\(^{207}\)Pb, \(^{208}\)Pb/\(^{206}\)Pb, \(^{208}\)Pb/\(^{207}\)Pb). As shown in Figure 2, this significantly enhances global coverage, particularly in previously undersampled regions such as the East China Sea, Arabian Sea, and Southern Ocean. The complete source paper list is in Appendix 9. On benchmark papers containing target data, Compass achieves an average per-paper recall of 0.745, offering a practical proxy for within-paper extraction completeness.

The full pipeline required approximately 52 GPU hours on dual RTX 3090 GPUs, an efficiency improvement of over four orders of magnitude compared to manual curation [46]. The one-time Knowledge Tree construction cost is detailed in Section 4.3.

5.2.2 Data Quality Assessment↩︎

Automated Constraint Checks. All 3,751 extracted records are validated against physical constraints encoded in the Knowledge Tree’s Validation Criteria. Value range verification and geographical boundary checks achieved 100% pass rates, while unit conversion validation reached 95%. Records that failed automated checks were flagged, rolled back, and re-processed by Compass. Rollback was triggered in approximately 0.06% of processed cases across the full 230,000-paper corpus, with each instance limited to re-processing a single step, resulting in negligible additional computational overhead.

Expert Manual Validation. We conducted stratified sampling across the 110 identified papers, reviewing 22 papers (20%): 12 Pb conc., 4 \(^{210}\)Pb conc., 5 Pb isotopes ratios, and 1 mixed. These 22 papers yielded 946 data points, of which 869 were confirmed correct, giving 92% accuracy (95% CI: \(\pm\)​1.7%). The review was led by one marine scientist with a second providing overall consistency checks. Of the 77 erroneous data points (\(\sim\)​8%), errors stem primarily from semantic confusion (43, \(\sim\)​56%), e.g., rainwater samples collected during marine cruises misclassified as seawater, or sediment data confused with dissolved-phase measurements; followed by data association errors (26, \(\sim\)​34%), e.g., mismatched latitude/longitude or depth across tables; and PDF parsing failures (8, \(\sim\)​10%), e.g., row misalignment in upstream parsing. Note that Table [tab:main95result] reports single-pass extraction for fair baseline comparison; with rollback enabled, Compass achieves F1: 0.619 (see Table [tab:ablation]). The remaining gap with the 92% deployment accuracy stems from two factors: the benchmark penalizes unextracted data regardless of paper relevance, and expert validation accepts minor coordinate variants (e.g., degree-minute conversions) that automated metrics penalize.

5.2.3 Case Study↩︎

Figure 6 illustrates the Compass workflow on Langford (1971) [45]. Compass identifies two relevant data tables and performs intra-table data association for each: in the first table, Pb measurements are linked with geographic coordinates recorded in its footnotes; in the second, data values are associated with the depth information specified in its caption. The independently resolved records from both tables are then consolidated into a unified, standardized dataset.

5.3 Online Data Platform↩︎

We launch an interactive online platform³ integrated into the open-source JingWei marine data system for visualization and exploration of the Pb database. Users can query records with full metadata and source provenance (Figure 7). Since launch, the platform has received over 1,590 visits, demonstrating its practical utility for marine scientific research.

5.4 Scientific Insights↩︎

The integrated marine Pb database provides a substantially enriched data landscape compared to existing resources, enabling scientific analyses previously infeasible due to data limitations.

Enhanced Global Coverage. As shown in Figure 2, the integrated database extends spatial coverage across all eight Pb measurement types. Among the 3,751 newly extracted records, concentration measurements constitute the largest share, with 1,228 Pb concentration and 472 \(^{210}\)Pb activity concentration records that directly support global Pb inventory estimation and vertical flux studies. The three \(^{204}\)Pb-normalized isotope ratios (\(^{206}\)Pb/\(^{204}\)Pb: 562; \(^{207}\)Pb/\(^{204}\)Pb: 532; \(^{208}\)Pb/\(^{204}\)Pb: 476) provide the most diagnostic power for source apportionment, as \(^{204}\)Pb is the only non-radiogenic stable Pb isotope. Additional paired ratios (\(^{206}\)Pb/\(^{207}\)Pb: 246; \(^{208}\)Pb/\(^{206}\)Pb: 139; \(^{208}\)Pb/\(^{207}\)Pb: 96), though fewer in number, complement source identification in regions where \(^{204}\)Pb measurements are unavailable. This broader coverage reduces reliance on interpolation for global Pb budget estimation, particularly in the Southern Hemisphere and marginal seas.

Regional Data Enrichment. The integration reveals previously under-sampled patterns in several key regions. In the East China Sea, newly incorporated Pb concentration and isotope data capture the influence of Yangtze River discharge and East Asian industrial emissions, filling a critical gap in global databases for this highly impacted marginal sea. In the Southern Ocean, recovered records from historical expedition reports provide valuable constraints on pre-industrial Pb baselines and the penetration of anthropogenic Pb into deep waters. In the Arabian Sea, new records complement GEOTRACES transects by adding temporal depth, enabling investigation of how regional Pb sources have evolved over decades.

Implications for Anthropogenic Impact Assessment. The temporal span of the dataset—from the 1960s to the present—enables multi-decadal trend analysis. The global phase-out of leaded gasoline provides a natural experiment for studying oceanic recovery, and our data capture both the peak contamination and recovery phases across diverse regions. The comprehensive isotope coverage supports source apportionment, linking observed Pb patterns to anthropogenic emissions. These insights demonstrate that systematically recovering scattered historical records enables a more complete understanding of human impacts on ocean chemistry.

6 Conclusion↩︎

In this paper, we propose Compass, a Knowledge Tree-enhanced LLM Agent framework that operationalizes an expert-guided adaptation approach for fine-grained scientific data integration from academic papers. By co-designing a domain-specific Knowledge Tree with marine scientists, Compass bridges the gap between general-purpose LLMs and high-stakes scientific domains without fine-tuning. Deploying Compass across over 230,000 open-access papers, we recover 3,751 previously unincorporated Pb records, substantially expanding the integrated marine Pb database, with 92% accuracy confirmed through expert validation. The newly integrated data reveal previously obscured patterns in under-sampled regions such as the East China Sea and the Southern Ocean. In the future, we plan to extend Compass to additional scientific domains and integrate multi-modal processing capabilities.

7 Limitations and Ethical Considerations↩︎

Despite Compass’s effectiveness, we acknowledge certain limitations delineating our current scope. Our implementation focuses on extracting data from tables and text, excluding figures. This is a strategic choice: the vast majority of marine Pb records are reported in tabular or textual format, whereas graphical representations are less commonly used for data reporting in this domain. Furthermore, extraction quality is partially bounded by upstream PDF parsing tools. While layout recognition errors can propagate through the pipeline, our modular design ensures that Compass will naturally benefit from future advancements in general document analysis. Finally, the Knowledge Tree requires expert-guided initialization; while this incurs a setup cost, the construction process is iterative and accumulates reusable experience—including core module design and structural patterns—such that subsequent updates or generalization to other scientific domains primarily require domain-specific knowledge adaptation rather than rebuilding the entire framework from scratch.

Ethical Considerations. All data are sourced from publicly available open-access papers with no personally identifiable information. Our platform ensures transparency by linking every integrated record to its original source to support reproducibility.

8 GenAI Disclosure↩︎

During the preparation of this manuscript, generative AI tools were used to assist with language refinement and grammatical correction of certain draft passages. All scientific contributions, including methodology, experiments, analysis and conclusions, are solely the work of the authors.

9 Complete Paper List↩︎

Table 2 presents the complete list of 110 papers that were identified by Compass during its deployment in marine Pb data integration.

References↩︎

---

1. Please visit our online platform at https://jingwei.acemap.cn/lead.↩︎

2. Our code is available at https://github.com/liuyiming01/COMPASS.↩︎

3. https://jingwei.acemap.cn/lead↩︎