GlazeSmith : AI Digital Twin for Ceramic Glazes

GlazeSmith is a physics-grounded digital twin for ceramic glaze chemistry: instead of stopping at a prediction on a page, it renders the predicted fired outcome live on a rotatable 3D vase, so a potter can inspect a glaze before ever loading a kiln. Under the hood, a deterministic physics engine computes the unity molecular formula and coefficient of thermal expansion via Appen coefficients, producing a real stress-delta against the target clay body rather than a guess. Three gradient-boosted XGBoost classifiers, trained on an AMD ROCm GPU instance, predict surface finish, transparency, and color family from the same recipe. A K-NN layer retrieves the nearest real-world glazes from a reference corpus, and a Pareto optimizer proposes adjusted candidates. The generated material texture is then mapped directly onto the 3D vase geometry - a true visual twin of the predicted result, not a flat swatch. A conversational agent, powered by Google's Gemma served through Fireworks AI, lets users ask about their recipe in plain language, and verifies every suggested change by re-running it through the same physics and ML pipeline rather than asserting an answer from memory.