Describe the phenomenon
Begin with a natural-language request for the physical system, controls, and visual result you want to explore.
AI + WebGPU for mathematical exploration
Create Math and Physics Simulations with AI and WebGPU
Describe a phenomenon—from orbital mechanics and electromagnetism to fluid dynamics and quantum physics—and turn it into an interactive GPU simulation in your browser.
GPUSims handles WebGPU resources, pipelines, rendering, and boilerplate so that AI can concentrate on the mathematical model and simulation logic.
Try a prompt:
Create a schema for an N-body simulation.
From an idea to a running model
Make advanced simulations more approachable
AI can generate simulation code for fields such as classical mechanics, electromagnetism, fluid dynamics, statistical physics, and quantum mechanics. GPUSims provides the structure that helps the AI produce code that can actually be compiled, executed, inspected, and improved.
Generate the simulation
AI creates a declarative schema and implements the mathematical logic inside an app-generated WGSL skeleton.
Run, inspect, and learn
Execute the simulation, examine numerical and visual output, correct problems, and generate an explanatory article.
Why the architecture works
Separate WebGPU orchestration from simulation logic
WebGPU setup is powerful but complex. Instead of asking AI to build every buffer, binding, pipeline, and renderer from scratch, GPUSims divides the work into four clear stages.
-
Schema
Declare the blueprint
AI defines GPU resources, shader passes, user controls, and execution order in a JSON-like schema.
-
Skeleton
Generate safe WGSL structure
GPUSims converts the schema into shader skeletons with matching structs, bindings, and entry points.
-
Logic
Implement the mathematics
AI fills in the core equations and algorithms without having to recreate the surrounding WebGPU infrastructure.
-
Execution
Run in the browser
GPUSims compiles the shaders, executes the GPU pipeline, renders the result, and exposes controls and readback values.
The key idea:
The app manages the machinery. AI focuses on translating mathematics and physics into executable shader logic.
Two ways to build
Stay in control or automate the whole pipeline
Interactive
Guided AI Workflow
Move step by step between GPUSims and your AI chat. This workflow is transparent, editable, and especially useful for learning.
- Copy the GPUSims system prompt into the AI chat.
- Ask the AI to generate a simulation schema.
- Paste the schema into GPUSims and generate the WGSL skeleton.
- Ask the AI to implement the skeleton, then run the simulation.
Advanced · Local
Automated Python Agent
Start with one prompt. A local Python agent coordinates the AI API, Selenium, Chrome, and GPUSims to complete the development cycle.
- Generate the schema and WGSL implementation.
- Enter the code and launch the simulation automatically.
- Return compilation or runtime errors to the AI for correction.
- Inspect numerical readback values and captured images.
- Create a thumbnail, metadata, and explanatory Markdown article.
Requires a configured local Python environment, browser automation, and AI API credentials.
Execution is only the beginning
Debug code and question the result
Code-error correction
If a schema or shader fails, copy the reported error to the AI. It can revise the schema or WGSL code and continue the iteration.
Fix this error.
Numerical inspection
Readback values can be returned from GPU memory and examined to catch mistakes that are not obvious from the animation alone.
Visual inspection
Capture the rendered canvas and ask the AI to evaluate whether the visible behavior is consistent with the intended model.
Is this result physically plausible?
AI-assisted checks are useful for iteration, but they are not formal scientific verification. Complex or high-stakes models still require knowledgeable human review.
A visual path into difficult subjects
Learn through simulation, explanation, and questions
Abstract equations become easier to approach when you can observe and manipulate their behavior. After a simulation is complete, AI can write an article explaining the visible phenomenon, mathematical background, and code implementation.
The article is a starting point rather than the end: continue asking questions about any equation, algorithm, or unexpected result until the model becomes understandable.
Generated learning article
- Observable phenomena
- Mathematical background
- Numerical method
- WGSL implementation
- Questions for further exploration
A virtual laboratory for AI
Generate, execute, observe, and correct
GPUSims gives AI more than a text-only coding task. The simulation produces compilation messages, numerical values, and images that can be fed back into an iterative loop of hypothesis, execution, observation, and correction.
Technical scope
WebGPU strengths and limitations
Designed for browser-based GPU computing
WebGPU enables parallel computation and high-performance visualization without requiring a native desktop application.
Primarily single precision
Simulations generally use 32-bit floating-point values (f32), not the 64-bit precision common in scientific computing.
No traditional numerical libraries
Libraries such as LAPACK are not available directly, so required algorithms must be implemented for the GPU architecture.
A bridge toward CUDA concepts
The same foundations—parallel execution, workgroups, buffers, and memory access—provide useful preparation for more advanced GPU computing.
Begin with one idea
What physical world would you like to build?
Start with a familiar system, inspect what the AI creates, and use the simulation as a path into deeper mathematics and physics.