AI Text-to-Draw
Synthesizes normalized contour coordinates, ARGB fills, and initial behavior blocks directly from natural language.
Prompt“A red race car and a green tree”
AiObjectFactory · JSON2 ModelUnits
Educational Modeling Lineage (2013–2019) · 2026 Revitalization
SimSketch: Sketch-Based Scientific Modeling, Revitalized for 2026
Originally pioneered in science education research to let students learn by modeling through freehand drawings—now rebuilt from the ground up in Flutter/Dart with four specialized domain studios, 2D rigid-body physics, and Socratic multimodal AI scaffolding.
Investigated in IEEE TLT (2013) · JSET (2015, 2018) · Springer (2019) · Interactive 2026 Studio at /simsketch
SimSketch — Sandbox Studio (2026)
Socratic AI
Behaviors
Go
Follow
Eat
Circle
Toroidal WrapActive · 60Hz
perfect_freehand · Ink Canvas3 Units
Running · t = 14.2s
InspectorFox_01
FOV Radar120px
Speed3.2 px/f
Targettype: Rabbit
Energy84%
Snapshot60s ring
Peer-Reviewed Lineage (2013 → 2019 → 2026)
Why SimSketch Matters in Science Education—and Why We Revived It
Conventional modeling environments require learners to master programming code or differential equations before testing a hypothesis—confining classroom practice to passive “learning with models” rather than active “learning by modeling.” Conceived at the University of Twente and extensively researched at Utrecht University's Freudenthal Institute, SimSketch demonstrated across multiple classroom and museum studies that annotated freehand sketches enable young learners to construct, execute, and revise genuine computational models.
In 2026, we revitalized this foundational research software in Flutter/Dart—preserving its constructivist sketch-to-agent pedagogy while introducing domain-specific physics, ecology, kinematics, mathematics, and Socratic AI scaffolding.
- 2013Foundational Paradigm
Externalized Mental Models & Multi-Agent Sketching
Bollen & van Joolingen formalized sketch-based modeling: decomposing learner drawings into autonomous
ModelUnitglyphs governed by composableAbstractBehaviorprimitives across astronomy, predator-prey, and traffic systems.Primary PublicationIEEE TLT · doi:10.1109/TLT.2013.9 - 2015 – 2019Empirical Validation
Epistemological Competence, Cognitive Load & Scaffolding
Evaluated across science-center astronomy cohorts and lower-secondary biology classrooms—establishing measurable gains in students' epistemological grasp of scientific models while revealing how interface parsimony and teacher scaffolding shape causal reasoning.
Empirical Studies (DOIs) - 2026Contemporary Revitalization
Multi-Domain Formalisms & Socratic VLM/LLM Scaffolding
Translating a decade of design-based research into a modern cross-platform architecture: four domain-tailored laboratories (Sandbox, progressive Ecology, involute GearSketch, and MathAnim 2D) augmented by multimodal vision and Socratic conversational inquiry.
Interactive Research EnvironmentOpen Live 2026 Studio (/simsketch)
2013–2019 Research InstrumentJava & Early Web Prototypes
Original SimSketch (Bollen, van Joolingen & Colleagues)
Pioneered drawing-based inquiry in primary classrooms, science museums, and biology field studies.
- Pen-tablet stroke canvas with stroke-level erasing, lasso grouping, and density-based stroke segmentation.
- Color-classified behavior taxonomy: meta-behaviors (red), independent actions (blue), type-targeting (orange), and object-bound rules (green).
- Domain adaptations crafted per classroom study (planetary orbits, evolutionary snail camouflage, predator-prey).
- Demonstrated that learners require structured complexity staging and interactive Socratic prompts from teachers.
What's New in 2026Flutter / Dart · Web, Desktop & Mobile
SimSketch (2026 Revitalization)
Uniting the validated pedagogical paradigm with modern physics, analytical engines, and multimodal AI.
- Four specialized modeling studios in one workspace: Sandbox (19 behaviors), 3-Tier Ecology, GearSketch kinematics, and 60 FPS MathAnim 2D.
- Pure-Dart 2D rigid-body physics (restitution, friction, 3x sub-stepping), live FOV sensory radar rings, and 2D raycasted optical shadows.
- Multimodal AI co-creation across Gemini, ChatGPT, and Claude: Socratic Modeling Assistant, notebook Photo Import, and Text-to-Draw.
- 60-second circular snapshot ring buffer for step-backward time-travel ('['), cross-mode Undo/Redo, and local JSON scene serialization.
Complete Peer-Reviewed Bibliography
- Bollen, L., & van Joolingen, W. R. (2013). SimSketch: Multiagent Simulations Based on Learner-Created Sketches for Early Science Education. IEEE Transactions on Learning Technologies, 6(3), 208–216. https://doi.org/10.1109/TLT.2013.9
- van Joolingen, W. R., Aukes, A. V. A., Gijlers, H., & Bollen, L. (2015). Understanding Elementary Astronomy by Making Drawing-Based Models. Journal of Science Education and Technology, 24(2–3), 256–264. https://doi.org/10.1007/s10956-014-9540-6
- Heijnes, D., van Joolingen, W. R., & Leenaars, F. (2018). Stimulating Scientific Reasoning with Drawing-Based Modeling. Journal of Science Education and Technology, 27(1), 45–56. https://doi.org/10.1007/s10956-017-9707-z
- van Joolingen, W. R., Schouten, J., & Leenaars, F. (2019). Drawing-Based Modeling in Teaching Elementary Biology as a Diagnostic Tool. Chapter 8 in Towards a Competence-Based View on Models and Modeling in Science Education (Springer), Vol. 12, pp. 131–145. https://doi.org/10.1007/978-3-030-30255-9_8
Empirical Evidence & Design Translation
Published Findings & How They Shaped 2026
SimSketch is one of the few educational modeling tools tested repeatedly across science museums and secondary classrooms. Each empirical finding directly guided a concrete capability in our 2026 re-architecture.
Wilcoxon Z = 5.432019
Statistically significant gain in modeling competence across 7 ninth-grade biology classes (N = 204).
Pre-test Mean6.92
Post-test Mean8.32 (+20%)
110 complete pre/post datasets: mean score rose from 6.92 to 8.32 (max 24) after four 45-minute lessons.
Epistemological Shift2019
Medium overall effect size on students' epistemological understanding of scientific models.
Nature of modelsd = 0.74
Multiple modelsd = 0.28
Driven by substantial progress on 'nature of models' (d = 0.74) and 'multiple models' (d = 0.28).
Ages 7–18 Cohort2015
Complete learner datasets in the Dutch Science Museum planetary modeling study.
Valuing Scale3.36 / 4.0
3-Body Orbit Built204 / 248 (82%)
204 learners modeled Sun–Earth–Moon orbits; rated 3.21–3.36/4 on valuing and 2.71–3.05/4 on perceived competence.
36 → 87 Utterances2018
Design-based research iterations investigating evolutionary reasoning in lower secondary biology.
Task-first framing36 utterances
Tool-first & simplified87 (2.4×)
Demonstrated that simplifying habitat/behavior controls and introducing tool mechanics prior to inquiry more than doubled peer reasoning.
From Published Literature to 2026 SimSketch Innovations
IEEE TLT · 2013doi:10.1109/TLT.2013.9
Empirical Insight: Freehand externalized representations lower syntax barriers and let learners construct, test, and revise multi-agent models.
2026 Feature: 19-Behavior Sandbox, Pressure Ink & 2D Rigid Physics
Expanded from 8 Java primitives to 19 typed behaviors, perfect_freehand calligraphic contours, raycasted 2D shadows, and 3x sub-stepped Newtonian collisions.
JSET · 2015doi:10.1007/s10956-014-9540-6
Empirical Insight: Physical, astronomical, and mechanical phenomena benefit from domain-specific representations beyond generic sprite motion.
2026 Feature: GearSketch Mechanical Studio & MathAnim 2D
Dedicated studios synthesize true involute gears, coaxial axles, and roller chains from circles, alongside a 60 FPS function, Riemann integral, and Euler ODE studio.
JSET · 2018doi:10.1007/s10956-017-9707-z
Empirical Insight: Unconstrained complexity overwhelms novices; progressive parameter exposure and live population statistics elevate causal reasoning.
2026 Feature: 3-Tier Progressive Ecology Laboratory
Stages complexity across Basic (survival), Intermediate (mating & carrying capacity K), and Advanced (metabolism & photosynthesis) with real-time population curves.
Springer Ch. 8 · 2019doi:10.1007/978-3-030-30255-9_8
Empirical Insight: Diagnostic classroom video logs showed students need active Socratic prompting to connect simulation outcomes back to biological hypotheses.
2026 Feature: Socratic Modeling Assistant & Multimodal Vision
Integrates Gemini, ChatGPT, and Claude to pose guiding causal questions rather than handing out pre-built models, plus notebook Photo Import and 60s time-travel.
Empirical metrics and effect sizes reported above originate strictly from the cited 2013–2019 publications on SimSketch. The 2026 multi-domain and AI capabilities represent the modern re-architecture and ongoing research direction.
2026 Capabilities · Multi-Domain Workspace
Four Specialized Simulation Studios Added in 2026
Expanding beyond a single generic drawing pad, the 2026 re-architecture pairs open-ended multi-agent sketching with dedicated ecological, mechanical, and mathematical laboratories.
Interactive Studio Preview · Switch tabs or use arrow keys
SimSketch — Sandbox Mode
⌘Z Undo·[ Time-Travel·Flutter 3.12+
Behavior Palette19 behaviors available
Hover or focus any chip to inspect its rule & parameters.
InkLasso
Fox:
Rabbit:
Plant:
Ready · Space to run
Inspector PanelFox_01
Live Ring
2D Rigid-Body Physics
Restitution65% (Bounce)
Surface Friction30%
Solver3x Sub-Step
60s Snapshot BufferJSON Ready
60s Time-Travel & Cross-Mode Undo
Circular snapshot buffer captures state every 1.5s for step-backward inspection (`[`) alongside full `⌘Z / ⌘Y` stacks.
Portable JSON Scene Serialization
Export complete vector strokes, behavior graphs, and parameters as `.json` files or persist locally via `shared_preferences`.
6-Target Native & Web Parity
Single Dart codebase targeting Web (`/simsketch`), macOS, Windows, Linux, iOS, and Android with touch-pinch & stylus support.
2026 Capabilities · Multimodal Foundation Models
Multimodal AI Co-Creation & Socratic Inquiry
Earlier classroom studies highlighted that responsive Socratic scaffolding is crucial when learners interpret unexpected simulation results. In 2026, SimSketch introduces vision-language import, generative sketching, and a Socratic dialogic tutor to support inquiry without giving away answers.
AI Photo Import
Extracts hand-drawn paper sketches via multimodal vision, isolating distinct contours into live, centered agents.
Paper Sketch Input3 Vector Units
Socratic Modeling Assistant
Coaches learners via Socratic “why” questions to reason through causal chains before spawning the ecosystem.
“What happens to the rabbit population if the foxes reproduce too quickly?”
Socratic Tutor“What do you think would happen to the grass first?”
Google GeminiOpenAI ChatGPTAnthropic Claude
Provider-agnostic by design. Keys stay on the device.
2026 Engineering · System Architecture
Client-First Modular Architecture, Zero Intermediary Backend
Every stroke normalizer, kinematic solver, and differential compiler executes natively inside the Flutter/Dart client. Multimodal LLM calls route directly from device memory through an abstract ILlmService contract.
Flutter / Dart Client
lib/features/*drawing_engineisolatedperfect_freehand ink contours, bounding-box auto-centering & undo stacks
simulation_engineisolatedModelUnit, 19 behaviors, 3x sub-stepped 2D physics & raycasted shadows
gear_engineisolatedInvolute tooth synthesis, BFS phase alignment, coaxial axles & chains
math_engineisolated60 FPS zero-dependency expression AST compiler & Euler ODE stepper
pedagogy_llmisolatedAiObjectFactory JSON normalization, Vision import & Socratic prompts
LLM Abstraction Layer (ILlmService)lib/core/api_interfaces
Swappable multimodal repositories emitting validated ModelUnit stroke & behavior schemas:
GeminiFlash / Pro Vision
OpenAIGPT-4o Multimodal
ClaudeSonnet Vision
shared_preferencesLocal-first persistence for user API keys, workspace preferences, 60-second circular snapshots, and portable JSON scene files—never touching a third-party server.
- Flutter
- Dart
- Gemini API
- OpenAI API
- Anthropic Claude API
- perfect_freehand
- shared_preferences
- JSON scene serialization
- Web / macOS / Windows / Linux / mobile
“The sketch is the interface. The simulation is the payload. The AI is the co-author.”
2026 Re-Architecture · Design Principles
Modular Architecture, Reproducibility & Local Privacy
Uniting four simulation studios and three foundation-model providers inside one cross-platform environment required strict module isolation, typed domain models, and client-only credential storage suitable for classroom and personal devices.
lib/features/*Feature-driven modular architecture
Isolated drawing_engine, simulation_engine, gear_engine, math_engine, and pedagogy_llm packages keep four distinct studios maintainable under one workspace.
ILlmServiceProvider-agnostic LLM abstraction
Google Gemini, OpenAI ChatGPT, and Anthropic Claude sit behind a single typed Dart interface, decoupling multimodal inference from studio canvases.
ModelUnit · GearUnitTyped state and immutable models
ModelUnit, GearUnit, GearChain, and MathStudioState enforce compile-time safety across spatial transforms, kinematics, and JSON serialization.
Cross-Mode QAManual & exploratory testing across four studios
Structured cross-mode regression checklists verify that Sandbox behaviors, time-travel snapshots, and undo stacks remain intact as GearSketch and MathAnim evolve.
Docs & CitationsComprehensive technical documentation
Ships with complete setup guides, peer-reviewed academic citations, keyboard & gesture shortcut tables, and per-engine architectural specifications.
shared_preferencesPrivacy-by-default client-side AI
User API keys are stored exclusively on-device via shared_preferences and connect directly to provider endpoints—never routed through an intermediary backend.