Using Genetic Algorithms to Design Physical Singulating Sorting Machines

Problem: What is the best machine shape to sort different small parts? Can we simulate some vibrating plinko like mechanism by the millions, assess their ability to singulate parts using off the web components? Yes! This is my custom-built Genetic Algorithm (GA) and Machine Learning pipeline designed to procedurally generate, simulate, and optimize physical sorting mechanisms for small parts (like agricultural seeds, manufacturing components, and Lego bricks). Here we are dropping bricks (unseen) down an inclined ramp with different obstacles to help distribute the pieces. Standard optimization algorithms usually fail when applied to chaotic 2D rigid-body physics (using matter.js)—a one-millimeter adjustment can change a perfect, high-throughput funnel into a catastrophic jam! In this video, I walk through how I solved the "black box" problem of evolutionary stagnation by building a three-pillar architecture: 1️⃣ The Simulator: A multi-threaded Web Worker pool running headless physics simulations to evolve complex, heterogeneous genomes (including latent structural genes). 2️⃣ The Cartographer: A Python/Dash dimensionality reduction pipeline (PCA) that maps the AI's strategies into a 3D interactive fitness landscape. 3️⃣ The Validator: A rigorous Monte Carlo stress-tester to eliminate statistical "flukes" and prove mechanical reliability. If you're interested in AI, physics simulations, or generative engineering, let me know what you think in the comments! 🛠️ Tech Stack Used: JavaScript, matter.js, Web Workers, Python, scikit-learn (PCA), Plotly Dash.