SciExplorer: Interactive Scientific Exploration

• SciExplorer is a framework for exploratory scientific interaction that exposes latent connections among research entities, papers, and experimental setups.
• It leverages diverse methodologies—from faceted biomedical navigation to agentic physical model discovery—to transform traditional search into interactive sensemaking.
• By integrating semantic navigation, graph analytics, and iterative hypothesis testing, SciExplorer turns static corpora into dynamic maps for discovery.

Searching arXiv for the named systems and papers to ground the article in the relevant literature. SciExplorer is a recurrent label for computational systems that support exploratory interaction with scientific knowledge rather than conventional query–response retrieval. In the literature, the name appears in settings that include faceted navigation over COVID-19 research, sentence-level extractive querying of biomedical corpora, visualization-oriented scientometric mapping, and LLM-based agentic discovery of physical laws. Across these settings, the shared objective is to expose latent structure—associations among biomedical entities, research groups, themes, citations, collections, or dynamical hypotheses—so that scientific corpora and scientific phenomena become navigable objects of inquiry rather than static result lists (Hope et al., 2020, Taub-Tabib et al., 2020, Xu et al., 11 Oct 2025, Nägele et al., 29 Sep 2025).

1. Exploratory search as a research paradigm

SciExplorer-oriented systems are motivated by a common diagnosis: targeted search is effective when the query is already well formed, but it is less effective when the task is discovery. In fast-moving domains such as COVID-19 research, relevant knowledge is scattered across papers and communities, users may not know the right terms or authors in advance, and scientific discovery often depends on finding surprising relationships between concepts, papers, and research groups. The exploratory-search framing therefore emphasizes open-ended sensemaking, associative browsing, and the ability to discover who is working on what, and where, rather than merely retrieving documents that lexically match a query (Hope et al., 2020).

The same distinction reappears in broader literature-exploration systems. Query-conditioned frameworks are designed to answer not only which papers match, but also what topics are present, which authors and institutions are central, and how documents relate through citations, geography, and time. This shifts the unit of interaction from isolated papers to a structured scholarly landscape (Jani et al., 14 Dec 2025).

A further extension appears in the agentic-discovery formulation for physics. There, the central claim is that scientific discovery is not a single prediction problem but an iterative loop: choose an experiment, observe the outcome, analyze it, generate or revise a hypothesis, and repeat until the model stabilizes. SciExplorer is therefore defined not by one interface style alone, but by a more general commitment to iterative exploration under uncertainty (Nägele et al., 29 Sep 2025).

2. Literature-centric implementations in biomedicine and COVID-19

A prominent literature-centric implementation is "SciSight: Combining faceted navigation and research group detection for COVID-19 exploratory scientific search" (Hope et al., 2020). SciSight is a prototype system for exploratory scientific search over COVID-19 literature that combines two capabilities. The first is faceted navigation over automatically extracted biomedical entities and PICO-inspired facets; the second is research-group detection and visualization using textual and network signals. Its workflow is explicitly staged: extract entities and topics from papers, build association networks among entities, cluster authors into research groups, represent each group using textual and network signals, link groups using topical and social meta-edges, support interactive exploration through faceted search and visualization, and allow drilling down from groups and facets to papers.

The biomedical facet layer converged on three widely used topical facets in biomedicine—patient/problem characteristics, interventions, and outcomes—while also extracting proteins, genes, cells, drugs, and diseases from titles and abstracts. Entity extraction uses S2ORC-BERT fine-tuned on BC5CDR and JNLPBA, and the system shows only entities that are collocated at least twice with other entities. PICO-style facets are extracted automatically from biomedical abstracts using the distant supervision model from Wallace et al. (2016). Metadata facets include journal, affiliation, and author. In the collocation explorer, term associations are built from co-appearance in the same sentence, edge weights correspond to collocation frequency, and the visualization is a customized chord diagram in which all pairwise interrelations are shown. Group exploration is driven by clustering the co-authorship graph through an overlapping community detection approach; group cards display the top three authors, topics, and affiliations, while green edges denote social affinity and purple edges denote topical affinity.

SciSight also formalizes group retrieval. One ranking signal is normalized facet overlap,

$\frac{|\{q: q \in f \}|}{|f|},$

where query facets $q$ are compared against a group’s top-$K$ salient facets $f$. A second signal uses weighted PageRank over the graph of groups and meta-edges, computed separately for topical affinity and social proximity, after which the system averages topical PageRank, social PageRank, and facet-overlap scores. Query semantics are conjunctive across facets and disjunctive within a facet, and the interface includes a time graph of publication counts so that users can inspect trends and peaks. The abstract reports that SciSight had served over $15K$ users with over $42K$ page views and $13\%$ returns (Hope et al., 2020).

A distinct but related implementation appears in "Interactive Extractive Search over Biomedical Corpora" (Taub-Tabib et al., 2020). Here the interaction unit is not primarily the paper but the sentence expressing a fact of interest. The system supports three query modes: boolean keyword queries, sequential or token queries, and example-based syntactic queries over dependency structures. Its signature design choice is a lightweight query language that does not require users to understand dependency graphs directly. Instead, users provide an example sentence and mark tokens as anchors, captures, and optionally expanded captures. A syntactic query such as BMP-6 induces the phosphorylation of Smad1 can therefore be transformed into a query graph without forcing the user to write dependency labels manually.

The extractive system is optimized for interactive speed through precomputed linguistic annotations and graph indexing. The corpus is processed with sentence splitting, tokenization, POS tagging, dependency parsing, and named entity recognition; the implementation uses SciSpacy with all four available NER models, converts syntactic trees to BART format using pyBART, and indexes the result through the Odinson framework. Contextual restrictions can be applied over the paragraph containing the sentence, title, abstract, authors, publication data or year, venue, and MeSH terms. The paper demonstrates the system on 14,446,243 PubMed abstracts and on the CORD-19 dataset of over 45,000 papers, including over 33,000 full-text papers. In a workflow for extracting stroke risk factors, the resulting file contained 1,212 rows, 640 unique risk factors, and 114 occurring more than once (Taub-Tabib et al., 2020).

Taken together, these systems instantiate two complementary modes of literature exploration. One treats the corpus as a network of concepts, facets, and research communities; the other treats it as a linguistically annotated reservoir of extractable relational statements.

3. SciExplorer as a scientometric mapping environment

In "Mapping the Urban Mobility Intelligence Frontier: A Scientometric Analysis of Data-Driven Pedestrian Trajectory Prediction and Simulation" (Xu et al., 11 Oct 2025), SciExplorer functions as a visualization-oriented bibliometric platform rather than a literature-search interface. It is jointly applied with Bibliometrix to analyze a Web of Science Core Collection dataset retrieved on 2025-08-24. The search returned 572 publications; 12 review articles were excluded, leaving 560 original research papers for analysis.

Within this workflow, SciExplorer is the science-mapping layer. It supports keyword co-occurrence networks, disciplinary or global science maps, thematic clusters, evolution of topics over time, and intellectual lineage from landmark papers. The study uses this apparatus to address five questions concerning whether the field is emerging, which disciplines influence it, which venues, countries, and authors are most influential, what research hotspots dominate, and which paradigms are likely to become the next milestones.

The reported scientometric indicators characterize the field as young and expanding: annual growth rate 18.44%, sharp increase in publications since 2017, average document age 4.19 years, average citations per document 19.58, average co-authors per document 4.07, and international co-authorship rate 22.9%. A Global Map of Science clusters the literature into 12 academic disciplines, with prominent clusters in Computer Science and Engineering, Civil and Construction Engineering, Mathematics and Physics / Physical Sciences, and Environmental Science and Technology. Peripheral but meaningful contributions come from Biology and Medical Sciences, Social Studies, Neurosciences, and Food and Drug Sciences. The keyword network identifies three main thematic clusters: a blue cluster centered on trajectory, pedestrians, autonomous vehicles, generative adversarial networks, and human trajectory prediction; a green cluster centered on deep learning, transformers, graph neural networks, and attention; and a red cluster centered on crowd simulation, reinforcement learning, path planning, agent-based modeling, and collision avoidance (Xu et al., 11 Oct 2025).

SciExplorer also supports contributor profiling. The country statistics reported in the study include China with 259 articles (46.3%), the USA with 67 articles (12%), and Japan with 30 articles (5.4%). In citation impact, the USA leads with 4083 total citations and 60.9 average citations per article, while Australia shows 49.6 average citations per article. Author influence is measured with productivity $N$, $H$-index, and total citations $TC$; the examples reported are Kapadia, Mubbashir as most prolific ($q$0), Liu Hong, Manocha Dinesh, and Zheng Nanning with highest $q$1-index ($q$2), and Alahi Alexandre as most cited ($q$3, with 4 papers). In this usage, SciExplorer makes the field’s intellectual structure legible as a network of themes, disciplines, and influential works rather than as a ranked list of publications (Xu et al., 11 Oct 2025).

4. Agentic exploration of physical models

"Agentic Exploration of Physics Models" (Nägele et al., 29 Sep 2025) gives the most expansive meaning of SciExplorer: a general-purpose agent for open-ended scientific discovery. The system is designed to automate the iterative scientific loop without domain-specific blueprints, fine-tuning, or task-specific instructions. In the main experiments, the backbone is GPT-5 Pro, prompted to act like a careful computational physicist. The prompt is generic across domains and encourages the agent to plan broadly, test many hypotheses, probe wide ranges of initial conditions, and delay final commitment until confidence is high.

The architecture is intentionally minimal. In each cycle, the LLM reasons about the current evidence, chooses an experiment or analysis step, invokes a tool, and stores the result in an external memory. The core tools are execute_code, which runs arbitrary Python/JAX/SciPy code; plot_from_code, which generates plots from saved arrays; and approx_equal, which provides a coarse similarity assessment between arrays. These generic tools are supplemented only by a domain simulator: an ODE solver for classical mechanical systems, a split-step field simulator for waves, or a quantum simulator for spin systems. The final answer is required to be executable code: a Python rhs function for mechanical systems, kinetic and potential propagators for wave systems, or a Python Hamiltonian construction for quantum systems (Nägele et al., 29 Sep 2025).

The benchmark spans three domains. The mechanical suite includes one- to three-dimensional ODE systems, planar particle systems, hidden-degree-of-freedom systems, and nonlinear oscillators such as damped Duffing, pendulum, asymmetric double-well, nonlinear damping, driven oscillator, parametric oscillator, double pendulum, coupled oscillators, Mexican-hat potential, off-center gravity, arbitrary 2D potential, multi-particle gravity systems, a 10-particle exponential interaction, and partially observed systems. The wave suite covers complex-field evolution on a 1D tight-binding lattice with periodic boundary conditions, including linear and nonlinear Schrödinger-type equations, Ginzburg–Landau equations, next-nearest-neighbor couplings, confining and periodic potentials, and a sinusoidal relaxation model. The quantum suite covers spin-$q$4 Hamiltonians inferred either from dynamics of spin expectation values or from ground-state expectation values of chosen operators, including transverse-field Ising chains, Heisenberg models, topological Ising chains, arbitrary spin Hamiltonians, and models with tunable couplings or system size (Nägele et al., 29 Sep 2025).

Evaluation is domain-specific. For mechanical systems, the target is a second-order ODE written in state-space form,

$q$5

and performance is scored with a coefficient of determination between true and predicted right-hand sides,

$q$6

For wave systems, the agent specifies split-step propagators such as $f$1 and the score again compares $q$7 under the true and predicted models using $q$8. For quantum ground-state tasks, one representative Hamiltonian is the transverse-field Ising chain,

$q$9

and the main score is fidelity per spin,

$K$0

For dynamical quantum tasks, the paper compares Hamiltonians using a normalized Frobenius inner product after trace subtraction (Nägele et al., 29 Sep 2025).

The reported results are broad. In mechanics, SciExplorer often recovers correct equations with near-perfect $K$1. One highlighted case is a tilted damped Duffing oscillator, for which the agent identifies

$K$2

with numerical parameters close to

$K$3

In a partially observed two-degree-of-freedom linear oscillator, it infers

$K$4

with

$K$5

and hidden initial conditions

$K$6

In the wave benchmark, one notable reconstruction is a complex Ginzburg–Landau model with next-nearest-neighbor interactions,

$K$7

with inferred complex coefficients

$K$8

In the quantum ground-state setting, the agent correctly identifies an open-boundary transverse-field Ising chain with tunable coupling and system size,

$K$9

and in the dynamical setting it reconstructs a 10-spin Heisenberg chain with transverse field,

$f$0

The paper is explicit that the strongest enabling factor is arbitrary code execution rather than a large catalog of specialized scientific tools. The agent writes its own routines for acceleration estimation, Fourier analysis, parameter fitting, residual checks, and numerical validation. Ablations show that removing arbitrary code execution severely degrades performance, and removing plotting is also harmful. At the same time, the method is limited by the LLM backbone, can take from a few minutes to more than 1.5 hours per exploration in the mechanical tasks, and can fail when it commits prematurely to an incorrect model or overlooks qualitative cues in plots (Nägele et al., 29 Sep 2025).

5. Adjacent exploration systems and infrastructural analogues

Several related systems, although not always named SciExplorer, define the broader technical landscape of scientific exploration tools.

System	Domain	Primary exploration mechanism
ChemCloud	Chemistry, pharmacy, biochemistry	RDF, ontologies, Linked Open Data, semantic autocomplete, semantic faceted search
scX	scRNA-seq	Shiny-based interactive analysis and visualization over Seurat, SingleCellExperiment, or count matrices
CollEX	Scientific collections	Multimodal agentic RAG with chat, BM25, HNSW, SigLIP, and LVLM tools
ISLE	Scientific literature	BM25 + semantic retrieval + RRF, BERTopic or NMF, query-conditioned knowledge graph

ChemCloud is a Semantic Web–based e-science information cloud that integrates resources such as ChemInform, InfoTherm, Chemisches Zentralblatt, ChemgaPedia, the W3C HCLS knowledge base, DBpedia Germany, and Bio2RDF sources including ChEBI and PubChem. It uses RDF as a common representation layer, derives ontologies from XML schemas, aligns them with expert-made ontologies and external knowledge bases, and supports semantic autocomplete and ontology-based faceted search (Todor et al., 2010).

scX is an R package implemented as a Shiny web application for interactive scRNA-seq exploration and visualization. Its preprocessing function createSCEobject can perform normalization, feature selection, clustering, dimensionality reduction, marker detection, and differential expression, largely through scran and related Bioconductor functionality. The interactive interface exposes modules such as Summary, Exploratory Data Analysis (EDA), Markers, Gene Expression, Differential Expression, and Visual tools; the largest tested dataset required up to 24 GB RAM during preprocessing, and a subsampling strategy with a default setting of 50,000 cells is used for smooth interactivity (Waichman et al., 2023).

CollEX generalizes exploration to multimodal scientific collections. It is a web application with a client–server architecture built from Vite, React, TypeScript, Material UI, FastAPI, Weaviate, LitServe, and LVLMs such as Gemma 3, Gemini 1.5 and 2.0, GPT-4o, and o1. Retrieval combines BM25 lexical search with semantic search over HNSW indexes; embeddings are computed with SigLIP; and specialized tools support database lookup, similarity search, image captioning, visual question answering, OCR, and object detection. The proof-of-concept deployment contains 64,469 unique records across 32 collections (Schneider et al., 10 Apr 2025).

ISLE, the "Intelligent Scientific Literature Explorer," provides a query-conditioned literature-exploration pipeline that merges arXiv full text or abstract-level content with OpenAlex metadata, stores the merged corpus in PostgreSQL, indexes it in Elasticsearch, fuses BM25 lexical search with dense semantic retrieval using Reciprocal Rank Fusion, performs BERTopic or NMF topic modeling, and constructs a heterogeneous knowledge graph over papers, authors, institutions, countries, topics, and years. The reported corpus scale is around 1.73 million papers, over 1.6 million authors, 36,531 institutions, 218 countries, and more than 3.4 million citation links (Jani et al., 14 Dec 2025).

A domain-specific design principle appears in the exoplanet paper "Indexing Exoplanets with Physical Conditions Potentially Suitable for Rock-Dependent Extremophiles" (Jagadeesh et al., 2020). There the Rock Similarity Index (RSI) is introduced as an organism-specific ranking metric for rock-dependent extremophiles rather than broad Earth similarity. A plausible implication is that SciExplorer-style systems need not rely on generic relevance alone; they can embed domain-informed ranking features tailored to particular biological, physical, or methodological targets.

6. Evaluation patterns, limitations, and significance

The literature presents a consistent pattern: exploratory systems are technically ambitious, but evaluation is often partial, qualitative, or deployment-specific. SciSight reports informal user studies with five experts or users from virology, viral infections, oncology, medical practice, and viral diseases or proteins, and the findings emphasize discovery of unknown associations and identification of relevant labs or collaborators (Hope et al., 2020). The extractive biomedical system demonstrates realistic workflows and exported outputs, but it does not claim to solve relation disambiguation automatically or guarantee high precision without iterative user refinement (Taub-Tabib et al., 2020). The scientometric use of SciExplorer is limited by the scope of Web of Science and by the fact that citation-based analysis underrepresents very recent work (Xu et al., 11 Oct 2025).

The same caution appears in more recent systems. CollEX explicitly notes that no formal system-level or component-level evaluation has yet been performed; its claims are grounded in qualitative demonstrations, and performance depends heavily on LVLM quality, with proprietary LVLMs working better but raising privacy and cost concerns (Schneider et al., 10 Apr 2025). ISLE lacks comprehensive relevance judgments at scale, notes topic-modeling sensitivity to the retrieved subset, and identifies visualization scalability as a challenge for very large query-conditioned graphs (Jani et al., 14 Dec 2025). In the physics setting, SciExplorer depends strongly on the reasoning quality of the backbone model, exhibits substantial runtime variability, and can fail when hidden degrees of freedom or structured potentials are not correctly inferred from the evidence (Nägele et al., 29 Sep 2025).

These limitations do not reduce the conceptual coherence of the field. Across implementations, SciExplorer denotes a shift from static retrieval toward interaction with structured scientific objects: entities and PICO facets, syntactic relations, thematic clusters, citation and affiliation graphs, multimodal records, or executable hypotheses. The systems differ sharply in substrate—literature, collections, single-cell data, chemistry databases, exoplanet indices, or physical simulators—but they converge on the same technical premise: discovery benefits when the underlying corpus or phenomenon is rendered explorable as a network, graph, map, or iterative experimental workspace rather than as a flat list of documents. In the available literature, the label therefore spans multiple implementations rather than a single canonical codebase, but the unifying idea is stable: scientific knowledge should be computationally navigable at the level where new relations, new hypotheses, and new research directions can emerge.