MultiphaseHUB：面向多相数据证据化元数据整理与任务感知压缩的社区基础设施

Yingyan Zeng Affiliation: Department of Mechanical, Materials, and Industrial Engineering, University of Cincinnati, Cincinnati, OH    J.P. Panda Affiliation: Department of Mechanical Engineering and Engineering ScienceUniversity of North Carolina at Charlotte, Charlotte, NC    Avranshu Mahanty Affiliation: Department of Mechanical, Materials, and Industrial Engineering, University of Cincinnati, Cincinnati, OH    Han Hu Affiliation: Department of Mechanical Engineering, University of Arkansas, Fayetteville, AR Affiliation: Arkansas Research Institute for Electronics Systems, University of Arkansas, Fayetteville, AR    Depeng Xu    Ying Sun Affiliation: Department of Mechanical Engineering and Engineering ScienceUniversity of North Carolina at Charlotte, Charlotte, NC

Abstract

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-analysis component 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 transport is central to boiling and condensation, immersion cooling, drop impact, particulate deposition, and many engineering processes that couple momentum, heat, mass, and interfacial dynamics. Investigating these processes requires observations across multiple spatial and temporal scales. High-speed imaging captures interface motion, infrared thermography measures spatial thermal response, acoustic sensing detects transient events, and pressure or electrical measurements characterize operating conditions and system behavior. Experiments and simulations therefore generate diverse, multimodal, high-dimensional datasets that can support physical understanding, model development, and engineering design. Their scientific value depends on the context needed to interpret the observations, including the geometry, working fluid, surface state, boundary conditions, sensor configuration, calibration, and uncertainty.

Despite the growing availability of multiphase transport data, reuse across studies remains difficult. Raw measurements, annotations, preprocessing scripts, and experimental descriptions are often distributed across laboratory websites, publication supplements, and general-purpose repositories. Important details may appear only in a paper’s methods section or in a separate calibration document. As a result, users must reconstruct the connections among data, physical conditions, and analysis procedures before determining whether a dataset is appropriate for a new question or can be used for one downstream analytic tasks. Differences in variable definitions, units, time stamps, sensing configurations, and operating regimes further complicate comparison. Persistent access to the measurement data is an essential starting point, but physically meaningful reuse also requires sufficient information about what was measured, how it was measured, and under which conditions.

The volume and incompleteness of multiphase transport measurements create coupled barriers to reuse. On one side, high-speed videos, temperature measurements, and acoustic signals can be costly to transfer and analyze, particularly when a user needs only a subset of their information. On the other side, these measurements may also contain missing segments or lack entire modalities because of sensor limitations, acquisition failures etc.

We introduce MultiphaseHUB [1] as a community infrastructure and an integrated framework for organizing, discovering, and reusing multiphase transport data. The infrastructure supports access to distributed data resources and their associated metadata, publications, analysis tools, and derived features, while retaining source repository links, provenance, and reuse boundaries. The framework integrates literature-grounded metadata curation and scientific data compression to support the interpretation and efficient reuse of these resources across physical systems and sensing modalities. Its development follows the goals of findability, accessibility, interoperability, and reusability articulated by the FAIR principles [2], with particular attention to the physical context and measurement information required for multiphase analysis.

The framework brings together three functional components: dataset discovery and organization, literature analysis and data curation, and scientific data compression. The latter two form the technical focus of this paper. Dataset discovery and organization connects source resources with their physical context and associated tools. The literature-analysis component includes asgentic metadata extraction from text and tables and a agentic literature-analyzer that recovers quantitative values from figures and expresses them in canonical units. Both modules connect extracted information with source evidence for verification and curation. Meanwhile, a modality-specific scientific compression method is proposed as one component in MultiphaseHUB with cross-modal imputation, using available measurements and physical context to estimate missing signals. The resulting proxies (i.e., compressed representations) would provide compact representations of observed measurements together with estimates of missing information. Their scientific usefulness depends on both reconstruction fidelity and imputation reliability. For example, smoothing a bubble interface can affect an estimated projected area, while attenuating or incorrectly imputing a weak acoustic transient can alter event detection. The relevant tradeoff therefore involves storage rate, reconstruction fidelity, imputation uncertainty, and downstream scientific utility. The extracted metadata from the agentic literature analyzer connects these objectives by describing operating conditions and measurement availability, helping determine which quantities matter and when relationships across modalities can support imputation. Together, these components support a workflow from identifying relevant data and interpreting their experimental conditions to constructing compact representations for subsequent analysis. They also provide a basis for investigating how metadata reliability and representation error affect downstream scientific utility.

This paper presents the infrastructure, the methodologies for the two technical components, and their application and preliminary results on seed multiphase transport datasets. Building on this work, we outline a research roadmap that extends the current curation and compression capabilities toward reuse across datasets and knowledge discovery in diverse application domains for multiphase transport. Cross-facility harmonization and multimodal alignment address differences in operating conditions, sensing configurations, and representations. Task-aware retrieval and dataset valuation investigate which resources can best inform a scientific question or resolve an evidence gap. Conditional reconstruction extends the compression workflow toward the proposed proxies by estimating missing segments or modalities from available measurements and physical context. Together, they connect the present infrastructure and preliminary methods to the longer-term goals of physically meaningful comparison, hypothesis development, and cross-study scientific discovery.

2 Related Data and AI Efforts

Open multiphase data resources already span experimental images, annotations, simulations, and machine-learning benchmarks. BubbleID couples boiling imagery, annotations, trained models, and tutorials for interface-dynamics analysis [3, 4]. Hessenkemper et al. released annotated and synthetic bubble images with a persistent dataset identifier to support image-based bubble identification [5]. BubbleML provides simulated boiling fields, baseline scientific-machine-learning tasks, and reproducibility materials for multiphysics learning [6]. While these efforts demonstrate that useful multiphase resources can be built around a well-defined physical system and task, they also expose a persistent interoperability problem: data scope, variables, formats, labels, sensor conditions, and quality controls differ substantially across resources.

Figure 1 summarizes the multimodal data of BoilingBench-1 [7], a subcooled pool-boiling run on a flat copper surface. All modalities are aligned to the thermocouple clock, and the heater power is held constant from about 5 s until it is switched off at about 670 s. Panel (a) shows the raw measurements: the liquid temperatures, the recorded pressure, the hydrophone voltage, and the amplitude of the acoustic-emission (AE) hits. Panel (b) shows the heat flux and DC input power, the hydrophone and AE spectrograms, and the band-integrated hydrophone and AE power. The heat flux is derived from a linear fit to the thermocouple readings. Panel (c) shows the images collected from recorded video.

In recent years, AI methods are increasingly applied to segmentation, tracking, optical-flow estimation, surrogate modeling, and nonintrusive sensing in multiphase systems. For example, BubbleID supports image-based analysis of bubble interface dynamics [3], BubbleML provides simulation datasets and benchmarks for scientific machine learning [6], and Wallen et al. combine acoustic-emission sensing with unsupervised learning to investigate flow condensation [8]. These efforts illustrate how different data modalities and analysis methods can support the study of multiphase phenomena. However, applying an existing method to a new dataset requires understanding the physical conditions, sensing configuration, preprocessing, and evaluation procedures used in its development. Differences in these factors can affect whether the method remains suitable and how its reported performance should be interpreted. This motivates connecting datasets with associated AI tools, publications, and experimental context within MultiphaseHUB, helping users identify relevant analysis methods and determine what adaptation or additional validation may be needed for their intended use.

The development of datahubs and data banks has been demonstrated in turbulence, materials science, structural biology, genomics, and other research fields. The Johns Hopkins Turbulence Databases provide programmatic access to large simulation fields through virtual-sensor services rather than requiring users to handle the entire archive [9]. The Materials Project combines structured material-property records with programmatic access [10]. The Protein Data Bank demonstrates community deposition, curation, standardized representation, and rich search for experimentally determined structures [11]. The Genomic Data Commons harmonizes cancer genomic and associated clinical data and supports cross-study analysis under explicit access controls [12]. Dataverse provides persistent identification, citation, versioning, and collection-level organization for research data [13]. Together, these dataset and datahub efforts demonstrate the value of combining reliable data access with scientific context and analysis tools. They motivate MultiphaseHUB as a community infrastructure and integrated framework that builds on existing repositories to support the discovery, interpretation, and reuse of multiphase data across physical systems and sensing modalities. The FAIR principles provide a foundation for this effort by emphasizing persistent identifiers, descriptive metadata, provenance, and clear access and reuse information [2].

Literature-grounded AI offers a way to recover relevant context when experimental conditions and measurement definitions are distributed across papers, supplements, and repository descriptions. Retrieval-augmented generation (RAG) combines document retrieval with language generation, allowing an output to be conditioned on external evidence [14]. Agentic approaches extend this capability by selecting subsequent actions in response to intermediate observations. ReAct, for example, interleaves reasoning with actions that access external information [15]. For multiphase metadata curation, these capabilities motivate workflows that retrieve evidence for a requested field, extract a candidate value, check units and experimental scope, and revisit the sources when information is missing or conflicting. However, retrieving a relevant passage does not establish that a value belongs to the correct run, surface, or operating condition. The domain-specific challenge is to produce structured assertions with traceable source locations, explicit unresolved states, and reviewable validation outcomes. Accordingly, a metadata extraction tool should be evaluated through field-level correctness, evidence support, and curation effort.

Scientific data compression techniques have been developed to reduce the storage and transfer costs of high-dimensional experimental measurements and simulation outputs, facilitating data access and reuse. Compression methods such as the fixed-rate floating-point representation introduced by Lindstrom reduce the size of numerical arrays while supporting efficient access [16]. Learned approaches jointly optimize an encoder, quantized representation, and decoder through rate-distortion objectives, as demonstrated by Ballé et al. for image compression [17]. Scientific utility nevertheless depends on the analysis performed on the reconstructed data. Laney et al. explicitly examined the effects of lossy compression using physically motivated simulation metrics [18]. These studies highlight the importance of evaluating scientific compression through both reconstruction error and preservation of physical quantities of interest (QoIs), i.e., scientific quantities computed from the data for a downstream analysis, such as the projected bubble area extracted from high-speed boiling video. However, preserving these quantities in multiphase transport remains challenging because bubble-interface geometry, event timing, acoustic spectral features, and thermal statistics can exhibit different sensitivities to compression across modalities and operating conditions. Missing signal segments or unavailable modalities further limit the information needed to estimate these quantities reliably. Motivated by these challenges, we propose to couple modality-specific compression with cross-modal imputation in MultiphaseHUB, using available measurements and physical context to construct compact proxies containing reconstructed observations and estimates of missing signals.

Together, these efforts establish MultiphaseHUB [1] as a community infrastructure and integrated framework for connecting distributed multiphase resources with literature-grounded physical metadata and compact representations whose scientific utility can be assessed. This integration provides a basis for investigating whether a dataset is suitable for an intended analysis, whether compression preserves the physical information that analysis requires, and how uncertainty in estimates of missing signals affects scientific conclusions. These questions connect the present curation and compression components to a broader research roadmap for reliable reuse and cross-study discovery.

3 MultiphaseHUB Architecture and Core Components

3.1 Functional components and information flow

MultiphaseHUB organizes three functional components, summarized in Table I. Dataset discovery and organization connect externally hosted resources with the information needed to locate and interpret them. Literature analysis and data curation provide physical metadata and quantitative evidence from publications. Scientific data compression provides compact representations of high-dimensional measurements. Here, a component denotes a major functional capability, while a module denotes a tool or workflow within that capability. External repositories and seed datasets are sources used by these components; cross-facility alignment, dataset valuation, and cross-modal imputation are research extensions.

3.2 Physical context, metadata, and provenance

Table II summarizes the information supporting these components, including source identity, physical conditions, measurement semantics, derived artifacts, and provenance. These categories describe the information needed for interpretation and reuse.

Each metadata value should be linked to the study, experimental case, measurement, or time interval it describes. Derived features and compression results should also identify their source data and processing settings. These links support data selection, reproducible analysis, and comparison across studies.

3.3 Prototype and a documented seed resource

The public MultiphaseHUB site exposes dataset and agentic-tool entry points [1]. In the architecture considered here, original files remain with their external repositories. The infrastructure supplies the connections among publications, physical metadata, compiled quantitative data, tools, and derived artifacts.

BubbleID offers a concrete seed example. The original paper describes bubble-interface analysis, and its Dryad record supplies a persistent dataset identifier and associated documentation [3, 4]. A reusable hub entry links those resources and associates the images and annotations with their documented experimental context. Paper-derived metadata remain distinguishable from labels distributed with the dataset. Any compressed derivative carries its own codec and evaluation information rather than inheriting the original asset’s status as an observation.

The MultiphaseHUB datasets page hosts 13 datasets: a seven-dataset “Boiling” collection and six stand-alone datasets covering condensation, drop impact, immersion cooling, particle deposition, cold-plate simulation and an additively manufactured liquid cooler. The datasets are predominantly experimental, with two literature-digitized compilations (Boiling Curve and Microbubble Emission Boiling) and one simulation set (Cold Plate). They span high-speed and infrared video, fluorescence and total-internal-reflection microscopy, hydrophone and acoustic-emission records, thermocouple, pressure and power time series, annotated images with trained machine-learning models, CFD fields and CAD geometry. Each dataset has its own page with a description, an “added” date, a README, and a filterable file table. Table III identifies each dataset by its name with the URL; listed entry counts are the file and folder rows shown on each dataset page.

Table IV audits what each dataset page exposes. The checklist covers (i) a text description, (ii) an “added” date, (iii) a README, (iv) a companion notebook, (v) faceted file metadata that describes physical experimental parameters (e.g. power, pressure, surface, Weber number) rather than only generic file kind or format, and (vi) a populated file-level “created” timestamp. ✓ = present, ∼ = partial, × = absent. “Facets” is the number of filterable metadata attributes on the page; every attribute carries a human-readable definition.

All 13 datasets have a description, an added date and a README (100%); 9 of 13 (69%) link a notebook, the exceptions being the four newest boiling datasets (BoilingBench-Multimodal, Sub-cooled Pool Boiling, Transient Boiling and Subatmospheric Boiling), which are marked “Notebook unavailable”; 7 of 13 (54%) expose physical-parameter facets, with BoilingBench-Multimodal partial because its only such facet selects the BoilingBench sub-collection.

3.4 Connecting metadata and compact representations

Figure 2 shows the relationship between the literature-analysis and compression workflows. Metadata extraction supplies experimental context, while the agentic literature-analyzer supplies quantitative values for analysis across publications. Compression produces compact representations accompanied by reconstruction and storage-rate information. Source links and provenance connect these products without requiring joint model training. Their integration can support subsequent evaluation of scientific quantities. Only verified metadata should be passed as conditions to a codec or used to define an evaluation split; otherwise, extraction errors can become hidden experimental confounders.

注意转载授权 · 阅读原文