GlazeSmith is a hybrid computational engine for evaluating and optimizing ceramic glaze stability. It enforces a strict separation between deterministic thermodynamic math and machine learning, avoiding the pitfalls of asking neural networks to guess at fundamental physics.

The system processes formulations through a six-layer decoupled pipeline: deterministic physics for thermal expansion, XGBoost classification for surface and color, K-NN vector retrieval for real-world analogues, Pareto optimization, SDXL for visual representation, and an LLM-based interpretation loop.

This project was built for the AMD Unicorn Track Hackathon. It leverages AMD ROCm MI300X hardware to co-locate graph neural network inference, vector search, and image generation in VRAM, while using Fireworks AI for fast LLM communication.