InteractScience: Interactive Scientific Systems

• InteractScience is an umbrella concept for dynamic scientific environments that merge models, data, and interactive visualization.
• It employs unified design principles such as integrated theory, immediate feedback, and progressive complexity to reveal latent structures.
• It supports reproducible research and education through executable publications, virtual labs, and benchmarks for scientific demonstration quality.

Across the cited literature, “InteractScience” can be understood as an umbrella concept for scientific environments in which models, data, instruments, and explanatory media are made directly manipulable, executable, or visually inspectable. In this sense, it spans interactive textbook simulations for biochemistry, physical visualizations of otherwise invisible phenomena, modular laboratories for mixed virtual and augmented contexts, overlays and maps of scientific literature, visual analytics systems for science–technology linkages, executable publication environments, formal studies of interactional properties, and benchmarks for interactive scientific demonstration code generation (Jakubowski et al., 2023, Montalbano, 2014, Kets et al., 2021, Leydesdorff et al., 2013, Wang et al., 2023, Klein et al., 2024, Rivière, 18 Jul 2025, Chen et al., 10 Oct 2025). This suggests that InteractScience is less a single platform than a programmatic orientation: scientific understanding is advanced by coupling formal representations to interaction, immediate feedback, and reproducible computational or experimental substrates.

1. Conceptual scope and recurrent design principles

A recurring premise is that scientific understanding often fails when structural, chemical, mathematical, textual, spatial, and temporal representations remain disconnected. In biochemistry education, this appears as difficulty connecting reaction schemes, ordinary differential equations, and progress curves; in physics education, as difficulty understanding magnetic fields, energy transport, or induction when these remain inaccessible to the senses; in science mapping, as difficulty seeing where a document set sits within the global journal system; and in innovation analytics, as difficulty tracing how patents rely on scientific papers (Jakubowski et al., 2023, Montalbano, 2014, Leydesdorff et al., 2013, Wang et al., 2023).

Across these domains, several design principles recur. One is tight integration with theory and narrative: equations, reaction schemes, and graphs are presented together, or visual analytics views are linked to metrics and entity profiles. A second is immediate feedback and manipulability: slider changes re-run biochemical simulations in real time; journal overlays can be explored interactively in VOSviewer; wavefield-synthesis and VR/AR experiments in ASIL depend on synchronized sensing and rendering; and benchmarked scientific demonstrations are defined precisely by their dynamic response to user actions (Jakubowski et al., 2023, Leydesdorff et al., 2013, Kets et al., 2021, Chen et al., 10 Oct 2025). A third is progressive complexity: simple binding or first-order kinetics lead into pathway models; local maps complement global maps; low-level HCI tasks are paired with higher-level or representative tasks; and educational notebooks can evolve from reproduction-oriented artifacts into broader pedagogical resources (Jakubowski et al., 2023, Leydesdorff et al., 2013, Rivière, 18 Jul 2025, Klein et al., 2024).

A plausible synthesis is that InteractScience is organized around the controlled exposure of latent structure. What is latent may be reaction dynamics, invisible physical fields, journal relations, science–technology dependence, interaction-loop effects, or executable dependencies within a publication. The common operation is to make these structures directly inspectable and experimentally variable.

2. Interactive representation, visualization, and learning

In educational settings, InteractScience is exemplified by systems that allow learners to manipulate parameters and observe dynamic consequences immediately. In Fundamentals of Biochemistry, two main classes of tools are embedded directly into the LibreTexts resource: CalcPlot3D for simple binding and kinetic reactions, and miniSidewinder for concentration-versus-time progress curves in systems ranging from simple reactions to metabolic and signal-transduction pathways represented in SBML. Users can move sliders for dissociation constants, kinetic constants, inhibition constants, Michaelis–Menten parameters, and initial concentrations, and can export .csv data for spreadsheet-based construction of derivative Lineweaver–Burk and traditional Michaelis–Menten plots (Jakubowski et al., 2023).

The biochemical cases make formal relations visible rather than merely symbolic. For an irreversible first-order process,

$A \xrightarrow{k} P,\qquad \frac{dP}{dt}=k[A],$

students can inspect how changing $k$ alters the curvature of $[A](t)$ and the approach of $[P](t)$ to its plateau. For binding equilibria and enzyme kinetics, the same environment supports inspection of

$$\theta=\frac{[L]}{K_d+[L]},\qquad v_0=\frac{V_{\max}[S]}{K_M+[S]},$$

and comparison of mechanistic and effective models whose macroscopic progress curves may largely overlap (Jakubowski et al., 2023). The explicit pedagogical point is that progress curves are closer to what experiments directly measure, and that $v_0$ is better understood as a derivative of these curves than as an isolated abstraction.

A parallel logic appears in physics education, where direct interaction with phenomena that are normally invisible is presented as a powerful learning tool. The paper on “seeing and interacting with the invisible” argues for physical systems that are directly coupled to the target phenomenon and respond in real time. Its examples include microphone-to-oscilloscope sound-wave visualization, magnetic-flux measurements for a falling magnet, ferrofluid displays of dipoles and multipoles, solenoid-current control with ferrofluid, Shive wave machines for energy transport, infrared imaging of electromagnetic braking, and oscilloscope-based visualization of electromagnetic damping (Montalbano, 2014). The distinction drawn there between schematic images, simulations, and physically coupled visualization is important: pure computer simulations may be experienced as artificial contexts, whereas direct physical visualization can anchor formal laws such as

$$\Phi_B=\int \vec{B}\cdot d\vec{S},\qquad \mathcal{E}=-\frac{d\Phi_B}{dt}.$$

These strands jointly support a broader interpretation: InteractScience in learning is not simply the addition of interactivity to instruction. It is the systematic alignment of formal models, parameter variation, visual response, and analytic workflow. The statement that it is “not just about flashy graphics” but about environments in which students connect multiple representations, experiment with models, and practice research-like reasoning is explicit in the biochemistry case (Jakubowski et al., 2023).

3. Infrastructures, laboratories, and executable environments

InteractScience also denotes a class of technical infrastructures in which interaction is not only represented but instrumented. The Art and Science Interaction Lab is a purpose-built interaction-science facility embedded in Ghent’s De Krook building: a $10\,\text{m}\times10\,\text{m}$, approximately $6\,\text{m}$-high, acoustically treated, reconfigurable lab with five motorized trusses, patch-box connectivity, 80 calibrated speakers, Qualisys motion capture, untethered EEG, eye tracking, physiological sensors, and synchronized 48 kHz audio and 120 Hz lab clocks (Kets et al., 2021). Its stated function is to support tightly controlled yet richly reconfigurable studies of interaction in mixed virtual and augmented contexts, including VR/AR/MR, wavefield synthesis, biofeedback art, telepresence, and multi-user embodied experiments.

The technical significance of such facilities lies in synchronized multi-sensor measurement and analysis. ASIL’s central backend supports acquisition, time-stamping, real-time processing, storage, and remote connectivity via dark fiber; its architecture allows 6DoF object tracking, avatar control, physiological footprint analysis, and experimental configurations that balance control and ecological validity (Kets et al., 2021). This suggests a strong infrastructural reading of InteractScience: interaction becomes a measurable scientific object when visual, auditory, kinematic, physiological, and computational layers share a common timing and data backbone.

A related, but publication-oriented, infrastructure appears in LiveDocs. There, research findings are turned into remotely executable development environments built from repositories containing notebooks, code, data, and deployment machinery. The initiative distinguishes Reproduction LiveDocs, which focus on re-running the results of a specific publication, from Educational LiveDocs, which add narrative, examples, and didactic framing. Deployment “flavours” include BinderHub or MyBinder, Docker, JupyterLite, and static HTML via nbconvert; the main interaction medium is the Jupyter ecosystem, optionally augmented by ipywidgets and Voilà dashboards (Klein et al., 2024).

The following examples illustrate the infrastructural breadth of InteractScience:

Infrastructure	Key components	Primary function
LibreTexts biochemical simulations	CalcPlot3D, miniSidewinder, SBML	Interactive reaction and pathway modeling
ASIL	Motorized trusses, synchronized backend, VR/AR, Qualisys, 80 speakers	Controlled multimodal interaction experiments
LiveDocs	Git repositories, notebooks, Docker, BinderHub, JupyterLite	Reproducible and reusable executable publications

Taken together, these systems show that InteractScience can be instantiated as an embedded textbook widget, a city-scale immersive laboratory, or a repository-backed executable paper. The unifying condition is not form factor but the provision of an environment in which scientific objects can be run, perturbed, and inspected under controlled conditions.

4. Mapping, measuring, and discovering interactions in science

A further branch of InteractScience concerns scientific interaction at the level of literatures, journals, papers, patents, and entities extracted from text. One foundational approach is the journal overlay methodology based on aggregated journal–journal citations. Using the 2011 Journal Citation Reports, the cited and citing dimensions of a $10{,}675 \times 10{,}675$ matrix are cosine-normalized, projected into two dimensions with VOSviewer, and colored by Louvain communities. Interdisciplinarity is then measured with Rao–Stirling diversity,

$k$0

where $k$1 is the share of the document set in journal $k$2 and $k$3 is the normalized Euclidean distance between journals on the map (Leydesdorff et al., 2013). The same work argues that the citing dimension provides a more comprehensive description than the cited archive for many mapping purposes, while also stressing that Rao–Stirling captures spread across categories and distances rather than intermediation or coherence.

InnovationInsights generalizes the idea of interactive overlays from journals to the dual frontier of science and technology. It constructs a multilayer graph of scientific papers, patents, authors, inventors, organizations, and fields, and visualizes paper–patent reliance through coordinated views centered on the Interplay Graph. Its quantitative layer includes patent reliance counts for papers, science reliance for patents, field-normalized impact, community structure, and local outlier detection; its system layer uses MongoDB for storage and a React.js plus D3.js front end for coordinated multiple views (Wang et al., 2023). The result is an environment in which users can trace the scientific roots of patented technologies, identify technologically influential papers, and inspect institution-level science–technology flows.

At the level of unstructured text, INtERAcT shows how interaction networks can be inferred without annotation or manual curation. It trains domain-specific Word2Vec embeddings on corpora of biomedical abstracts, clusters the embedding space with K-means, builds cluster-distribution fingerprints from the $k$4 nearest neighbors of each term, and defines an interaction score through Jensen–Shannon divergence:

$k$5

Across 10 cancer types, the resulting metric outperforms cosine and Euclidean similarity against STRING in reconstructing known molecular interactions (Manica et al., 2018).

These approaches extend InteractScience beyond direct user manipulation of a local simulation. They make the structure of science itself interactive: journals become coordinates in a global map; paper–patent reliance becomes a navigable frontier; biomedical interaction claims become weighted networks inferred from literature. A plausible implication is that the interactive object here is not a physical or biochemical process but the organization of scientific knowledge.

5. Formal models of interaction and interdisciplinary organization

InteractScience also has a formal, theory-building dimension in which interaction is modeled explicitly rather than merely visualized. In the agent-based model of interdisciplinary interactions, scientists are agents $k$6 endowed with disciplinary profiles represented by Gaussian distributions

$k$7

on a one-dimensional knowledge axis. The overlap

$k$8

defines project depth, while a broader covered area corresponds to interdisciplinarity. Collaboration choice is driven by a utility

$k$9

with collaborator selection governed by a multinomial logit parameterized by $[A](t)$0 (Raimbault, 2020). Across 158,400 NetLogo/OpenMole simulations on random, small-world, and scale-free networks, the model yields a non-linear Pareto front between depth and interdisciplinarity and indicates that intermediate values of $[A](t)$1 can provide better compromises than either random or strongly deterministic behavior.

A complementary formalization appears in the proposal for a science of interactional properties. There, the focus shifts from interfaces to interaction loops between person and system, and from isolated UI findings to replicable properties about the effect of one parameter on one loop. The revealing method is interaction loop diffraction, which introduces an “interaction prism” on an input or output loop and varies a single parameter while holding the rest of the interaction context fixed. Property records then specify prism type, loop, prototype scope, parameter values, tasks, human faculty, and the property itself, often as an ordering such as

$[A](t)$2

for AR depth cues, or

$[A](t)$3

for certain tactile-haptic speed profiles (Rivière, 18 Jul 2025). The method is explicitly oriented toward replication across prototypes, technologies, tasks, and user profiles.

What these two lines share is a micro–macro logic. In the ABM, local collaborator choices under resource constraints generate system-level disciplinary structure. In interaction loop diffraction, local parameter variations on a single loop generate claims that may later populate a catalog of interactional properties. This suggests a formal meaning of InteractScience in which interaction is treated as a primary explanatory variable, whether the system under study is a research community or a user–system loop.

6. Evaluation, limitations, and prospective synthesis

The most explicit evaluative formalization of InteractScience appears in the 2025 benchmark for interactive scientific demonstration code generation. That benchmark defines a task in which a model receives a detailed implementation plan and must produce a single self-contained HTML artifact implementing an interactive scientific demonstration. Evaluation is hybrid: Programmatic Functional Testing executes Playwright-based action–assertion scripts over the generated DOM, while Visually-Grounded Qualitative Testing executes action sequences to reach reference states, compares screenshots with CLIP, and uses a vision-language judge with reference snapshots and checklists (Chen et al., 10 Oct 2025). The benchmark contains 150 problems across mathematics, physics, chemistry, earth science, and computer science, and evaluates 30 open- and closed-source models.

The reported results clarify an important misconception. Many models achieve high VQT Action Success Rate, generally above 85%, meaning that sliders, buttons, and other controls can often be operated. Yet the best PFT Overall Pass Rates remain around 41.47% for Claude-Sonnet-4 and 39.47% for GPT-5, while VLM-judge scores remain well below perfect, with GPT-5 at 57.02 and Claude-Sonnet-4 at 55.42 (Chen et al., 10 Oct 2025). The contrast between the Huffman-tree case and the spring–mass–damper case shows why: visual layout can be plausible while the scientific rendering is wrong. In other words, surface interactivity is not equivalent to scientifically correct interactivity.

The broader literature repeatedly adds caution. Journal-based maps are more precise than Web of Science subject categories but remain coarse relative to article-level classification, and Rao–Stirling diversity should be treated as a relative indicator rather than an exact metric (Leydesdorff et al., 2013). LiveDocs improve accessibility and reuse, but JupyterLite is constrained by WebAssembly and some workflows require specialized hardware such as CUDA-capable GPUs (Klein et al., 2024). Highly modular laboratories such as ASIL enable unusual combinations of control and ecological validity, but this flexibility brings setup complexity and integration overhead (Kets et al., 2021). In the LibreTexts biochemistry work, formal assessment data are not yet available, even though the authors situate the approach in a broader literature on computational models and simulations in teaching (Jakubowski et al., 2023).

A plausible synthesis is that InteractScience is converging toward a layered program. At one layer are interactive representations that link equations, data, and visualization. At another are interoperable substrates—SBML, Jupyter, Docker, Playwright, or synchronized laboratory clocks—that make these representations executable and measurable. At a third are analytical formalisms—Rao–Stirling diversity, agent-based utilities, interactional properties, or JSD-based network scores—that convert interaction into a quantitative research object. The literature does not present these layers as a single unified framework, but taken together it outlines a coherent research agenda: science is increasingly understood, taught, evaluated, and organized through systems in which explanation and interaction are inseparable.