PaperlayerView exact revision on arXiv

Source-linked AI summary

MultiphaseHUB for Evidence-Grounded Metadata Curation and Task-Aware Compression of Multiphase Data

Yingyan Zeng, J. P. Panda, Avranshu Mahanty, Han Hu, Depeng Xu, Ying Sun

arXiv:2610.28644v1cs.DL

TL;DR

Multiphase transport data are difficult to reuse when experimental context is scattered and measurements are costly or incomplete. MultiphaseHUB connects distributed datasets with literature-grounded metadata and modality-specific compression. Preliminary results include literature extraction and compact representations, while the authors identify evaluation gaps and a single-dataset compression scope.

  • Problem

    Scattered experimental context and high-volume or incomplete measurements complicate multiphase-data reuse and increase the effort or cost of access and analysis.

    Source: §1 · p. 1

  • Method

    MultiphaseHUB connects externally hosted resources with metadata and provenance, combines literature analysis and curation with modality-specific compression, and outlines extensions including cross-facility harmonization and conditional reconstruction.

    Source: §1 · pp. 1–2, §3.1 · p. 4

  • Results

    The hub hosts 13 datasets; across 67 papers, two-figure digitization deviations averaged 0.31% and 0.35% of axis span; packed-index ratios were 10.7:1 for temperature/audio and 279.3:1 for video.

    Source: §7 · p. 15

  • Takeaways & Limitations

    The infrastructure connects distributed resources to physical metadata and compact representations for assessing dataset suitability and supporting physically meaningful comparison and cross-study discovery.

    Source: §1 · pp. 1–2

    Compression was studied on one multimodal boiling dataset; QoI preservation and independent scoring of metadata extraction remain unevaluated, while digitization accuracy was assessed on only two figures.

    Source: §6 · p. 14, §7 · p. 15

Abstract

from arXiv · show

Multiphase transport experiments and simulations generate high-dimensional data whose scientific value depends on the physical conditions, sensing configuration, calibration, synchronization, uncertainty, and provenance accompanying the raw signals. These data are difficult to reuse when experimental context is scattered across publications and repositories, while large measurement volumes increase the cost of transfer and analysis. In this work, we introduce MultiphaseHUB as a community infrastructure and an integrated framework for organizing, discovering, and reusing multiphase transport data. The infrastructure connects externally hosted datasets with modality-specific metadata, publications, analysis tools, derived features, source repository links, and boundaries for reuse. Within this framework, we develop two complementary components: literature analysis and data curation, comprising Large Language Model (LLM)-assisted metadata extraction and an agentic literature-analyzer for quantitative values from figures; and modality-specific scientific data compression guided by the quantities of interest needed for downstream analysis. These components address the effort required to interpret heterogeneous sources and the cost of accessing useful information from high-dimensional measurements. Public multiphase datasets provide seed resources for illustrating the framework and its workflows. Building on these components, we outline a research roadmap encompassing cross-facility harmonization, multimodal alignment, task-aware retrieval and dataset valuation, and conditional reconstruction of missing modalities. MultiphaseHUB provides a foundation for connecting distributed data resources to physically meaningful comparison and cross-study scientific discovery.

1 Introduction

Multiphase data reuse is hindered by scattered experimental context and high-volume or incomplete measurements. MultiphaseHUB connects distributed resources to literature-grounded metadata and compression, with a roadmap toward broader comparison and discovery.

1 Introduction

  • Scattered data, annotations, procedures, and experimental descriptions make it difficult to determine whether datasets suit a new question or support meaningful cross-study comparison.Differences in variables, units, timestamps, sensing configurations, and operating regimes add to the difficulty.Source: §1 · p. 1
  • High-volume measurements raise transfer and analysis costs, while sensor limitations or acquisition failures can leave segments or entire modalities missing.Users may need only a subset of the information in high-speed video, temperature measurements, or acoustic signals.Source: §1 · p. 1
  • MultiphaseHUB links distributed datasets with metadata, publications, tools, derived features, source repositories, provenance, and reuse boundaries.The framework emphasizes physical context and measurement information in line with FAIR principles.Source: §1 · pp. 1–2
  • The technical components combine literature analysis and curation—including metadata extraction and figure-value recovery—with modality-specific compression and proposed cross-modal imputation.The framework connects operating conditions and measurement availability to quantities of interest, reconstruction fidelity, imputation uncertainty, and downstream scientific utility.Source: §1 · p. 2
  • The roadmap proposes cross-facility harmonization, multimodal alignment, task-aware retrieval and dataset valuation, and conditional reconstruction to support physically meaningful comparison and cross-study discovery.These are extensions building on the paper's infrastructure and preliminary methods.Source: §1 · p. 2

2 Related Data and AI Efforts

Existing multiphase data resources and AI methods support focused analyses, but differences in data scope and experimental context complicate reuse. Related work motivates linking datasets to literature evidence and evaluating compression by the physical quantities needed downstream.

2 Related Data and AI Efforts

  • BubbleID, annotated bubble images, and BubbleML provide task-focused resources, but differences in scope, variables, formats, labels, sensor conditions, and quality controls limit interoperability.These resources span imagery, annotations, synthetic data, simulations, benchmarks, and reproducibility materials.Source: §2 · p. 2
  • BoilingBench illustrates aligned multimodal measurements, including temperature, pressure, hydrophone and acoustic-emission signals, heat flux, spectra, and video.The heater power is held constant from about 5 s until shutoff at about 670 s; heat flux is derived from a linear fit to thermocouple readings.Source: §2 · p. 2
  • Analysis methods such as segmentation, tracking, optical flow, surrogate modeling, and nonintrusive sensing require physical and procedural context to assess suitability on a new dataset.Differences in conditions, sensing, preprocessing, and evaluation can affect method suitability and interpretation of reported performance.Source: §2 · pp. 2–3
  • Prior data hubs and repositories combine data access with scientific context and tools, motivating a multiphase hub that builds on existing repositories and FAIR principles.Examples across turbulence, materials science, structural biology, genomics, and general research data illustrate approaches to access, curation, harmonization, and organization.Source: §2 · p. 3
  • Literature retrieval and agentic workflows can extract candidate metadata, check units and experimental scope, and revisit sources when evidence is missing or conflicting.Evaluation should consider field-level correctness, evidence support, and curation effort; retrieved passages alone do not establish that values belong to the right run or condition.Source: §2 · p. 4
  • Scientific compression should assess reconstruction error alongside preservation of physical quantities of interest, whose sensitivities vary across modalities and operating conditions.The proposed framework couples modality-specific compression with cross-modal imputation using available measurements and physical context.Source: §2 · p. 4
  • MultiphaseHUB links distributed resources with physical metadata and compact representations to investigate dataset suitability, preservation of analysis-relevant information, and effects of missing-signal uncertainty.These questions connect current curation and compression capabilities to reliable reuse and cross-study discovery.Source: §2 · p. 4

3 MultiphaseHUB Architecture and Core Components

MultiphaseHUB connects distributed datasets with physical context, literature-derived evidence, and compact representations through three functional components. Its prototype links metadata and provenance to datasets and tools, while auditing the information exposed by hosted dataset pages.

3.1 Functional components and information flow

  • The hub separates dataset discovery and organization, literature analysis and data curation, and scientific data compression; metadata extraction and literature analysis are modules within the literature component.A component is a major functional capability; a module is a tool or workflow within one.Source: §3.1 · p. 4, Table I · p. 5

3.2 Physical context, metadata, and provenance

  • Metadata values should link to the study, case, measurement, or interval they describe, while derived features and compression results should identify source data and processing settings.These links support selection, reproducible analysis, and comparison across studies.Source: §3.2 · p. 4

3.3 Prototype and a documented seed resource

  • The prototype keeps original files in external repositories and connects them to publications, physical metadata, quantitative data, tools, and derived artifacts.Its 13 hosted datasets span experimental, digitized-literature, and simulation resources and diverse sensing modalities.Source: §3.3 · pp. 4–5
  • All 13 dataset pages provide a description, added date, and README; 9 link a notebook, and 7 expose physical-parameter facets.The four newest boiling datasets lack notebooks; BoilingBench-Multimodal's only physical facet selects its sub-collection.Source: §3.3 · p. 5

3.4 Connecting metadata and compact representations

  • Metadata extraction supplies experimental context and figure analysis supplies quantitative values, while compression adds compact representations with reconstruction and storage-rate information.Source links and provenance connect these products without joint model training; only verified metadata should condition codecs or evaluation splits.Source: §3.4 · pp. 5, 7

4 Literature Analysis and Data Curation

MultiphaseHUB links distributed data resources to physical metadata, publications, tools, and derived features. Its literature-curation workflow extracts structured metadata and figure-derived values, with preliminary boiling-curve comparisons illustrating the compiler.

4.1 Metadata extraction from text and tables

  • Four fixed, schema-constrained language-model passes identify paper type, numerical operating envelopes, taxonomy facets, and stated or inferred applications from extracted text.The passes return structured metadata records using stage-specific instructions and output schemas.Source: §4.1 · p. 7
  • The catalog covers flow and pool boiling and evaporation, with prominent fluids including dielectric liquids and water and applications led by electronics thermal.Reported measurements include heat-transfer coefficients, wall superheat, boiling curves, and critical heat flux; an example records contact angles of 158° and 165° for distinct surfaces.Source: §4.1 · p. 7

4.2 Agentic literature-analyzer: quantitative extraction from figures

  • The figure analyzer extracts paired wall-superheat and heat-flux values, retaining source, series, units, extraction method, and reported physical conditions for comparison.Common units facilitate comparison, but geometry, surface state, pressure, and measurement definitions still require consideration.Source: §4.2 · p. 8

4.3 Validation and human review

  • Validation combines schema checks, SI conversion, plausibility screening, and optional language-model audits; human review addresses ambiguities numerical checks cannot resolve.Automated confirmation flags indicate triage outcomes, not completed expert review, and model-reported confidence is not a calibrated probability of correctness.Source: §4.3 · pp. 8–9

4.4 Output and evaluation

  • Structured records preserve source identifiers, evidence locations, original and normalized values, taxonomy assignments, derived tags, and extractor and taxonomy versions.Proposed evaluation includes hierarchical tagging precision and recall, numerical and unit accuracy, evidence support, unresolved-field rates, reviewer time, and residual error.Source: §4.4 · pp. 9–10

4.5 Preliminary literature-compiler application: boiling-curve comparison

  • 0.29% mean deviation (0.65% maximum) for Allred and 0.4% (2% maximum) for Berce; the largest Berce deviations occur on run 1’s steep backbend.Agreement is the nearest-point distance between digitized curves, normalized by the source plot’s axis span.Source: Figure 4 · p. 10, §4.5 · p. 11
  • The preliminary compiler jointly visualizes boiling curves from 67 papers with source labels, supporting inspection of curve shapes and superheat ranges without establishing cross-study harmonization or extraction accuracy.Differences in fluid, pressure, surface state, and measurement definitions must be checked before attributing curve differences to physical mechanisms.Source: §4.5 · p. 10

5 Scientific Compression of Multiphase Transport Data

MultiphaseHUB combines modality-specific residual-quantized codecs with dataset context and provenance to support reuse of multiphase measurements. Preliminary experiments assess reconstruction and index-payload compression on synchronized measurements from one boiling dataset.

5.1 Modality-specific encoding and residual quantization

  • Modality-specific RVQ-VAEs encode temperature, hydrophone audio, and boiling video as discrete indices, with later residual stages adding corrections and increasing payload.The preliminary implementation uses two stages; variable active stage counts and QoI-guided training are proposed extensions.Source: §5 · p. 11, §5.1 · p. 11
  • A proposed QoI penalty prioritizes selected scientific quantities alongside signal reconstruction, with validation needed to assess achievable errors at each active stage count.Candidate quantities include projected bubble area, acoustic band energy, and temperature statistics; some may require post-reconstruction evaluation or a stated differentiable surrogate.Source: §5.1 · p. 11

5.2 Boiling dataset and experimental setup

  • The compression evaluation uses 10,311 overlapping windows from one transient polished-copper boiling run, with synchronized temperature, hydrophone, and video inputs.The 80/20 within-dataset results are preliminary; source-disjoint generalization requires preventing shared raw samples across subsets.Source: §5.2 · p. 12

5.3 Preliminary reconstruction results

  • Reconstructed examples retain local temperature trends, main audio excursions, and major bubble structures, while smoothing rapid fluctuations, waveform details, and fine video interfaces.The illustrated video example has an MSE of 0.0765 at epoch 22; this is a sample-level metric, not a dataset-wide average.Source: §5.3 · p. 13
  • Video uses 1,739 of 2,048 stage-1 codes and all 2,048 stage-2 codes, while temperature and audio show different usage patterns.Code usage describes representation utilization, not the accuracy gain from adding a second stage.Source: §5.3 · p. 13

5.4 Compression ratios and storage accounting

  • The combined index payload falls from 6,312,192 to 24,472 bits, an approximately 257.9:1 ratio; video accounts for most of the storage saving.The reported payload ratios are 10.7:1 and 279.3:1; these are index-payload figures and exclude deployment side information and model files.Source: §5.4 · p. 13

5.5 Integration with MultiphaseHUB

  • Each compressed representation is linked to its source dataset, preprocessing, experimental conditions, decoder, and codebook versions for users to trace reconstruction context.The preliminary results provide a starting point for evaluating effects on scientific quantities and selecting representations for specified analyses.Source: §5.5 · p. 13

6 Research Roadmap

The roadmap extends current curation and compression toward comparison across facilities, multimodal alignment, task-aware resource selection, and missing-data reconstruction. These directions remain future work, with broader evaluation needed beyond the single boiling dataset used in the current compression study.

Research roadmap scope

  • Compression evaluation currently uses one multimodal boiling dataset; extension across operating regimes, sensors, and facilities is a central next step.Broader seed resources, reproducible evaluation workflows, and community contributions are also identified as needed for progress.Source: §6 · p. 14

Cross-facility harmonization and multimodal alignment

  • Cross-facility harmonization and multimodal alignment would connect physical conditions, variables, units, calibration, sampling, and differing time bases, spatial resolutions, and sensing mechanisms.Evaluation across independent studies and facilities is proposed to assess reliable comparison and transfer.Source: §6 · p. 14

Task-aware retrieval and dataset valuation; Scientific utility and missing-data reconstruction

  • Task-aware retrieval and dataset valuation would assess resource relevance, added information, operating-regime coverage, measurement quality, missing information, and reconstruction uncertainty.The roadmap also proposes evaluating rate-utility tradeoffs and estimating missing segments or modalities from available measurements and physical context, with uncertainty and supporting observations attached to inferred signals.Source: §6 · p. 14

7 Conclusions

MultiphaseHUB connects distributed multiphase datasets with context and reuse tools, while its literature workflows and compression methods show preliminary results. The paper also identifies evaluation gaps that constrain those results.

7 Conclusions

  • The hub hosts 13 datasets across six application areas; all have descriptions and READMEs, 7 expose physical-parameter facets, and none has a persistent identifier.The areas include boiling, condensation, drop impact, immersion cooling, particle deposition, and liquid cooling.Source: §7 · p. 15
  • Literature workflows extracted structured metadata from 67 heat-transfer papers; digitization of two boiling-curve figures differed from manual extraction by mean values of 0.31% and 0.35% of axis span.Source: §7 · p. 15
  • For one multimodal boiling dataset, two-stage RVQ achieved packed-index compression ratios of approximately 10.7:1 for temperature and audio and 279.3:1 for video.These ratios exclude model storage and side information; reconstruction examples and codebook utilization provide initial evidence for compact representations.Source: §7 · p. 15
  • Evaluation remains preliminary: compression used one transient run on one surface, QoI preservation is untested, metadata extraction lacks independent scoring, and digitization accuracy covers only two figures.Reconstruction is characterized by examples rather than dataset-wide metrics.Source: §7 · p. 15

Data and Code Availability

The platform, BubbleID source data, and code and workflows are publicly available through their respective sources.

Data and Code Availability

  • The MultiphaseHUB platform is publicly available, and BubbleID source data remain available through the original Dryad record.Source: Data and Code Availability · p. 15
  • The source code, literature-analyzer workflows, and AgenticTool execution framework are publicly available with implementations and materials to reproduce the reported experiments.Source: Data and Code Availability · p. 15

Related papers

Loading 2610.28644v1…