An Edge-AI-Driven Crystal Structure Exploration Tool for Secondary School Chemistry: Design and Implementation

Abstract

Crystal structure is one of the most abstract topics in secondary school chemistry. Textbookspresent three-dimensional lattices as flat diagrams, and physical models cannot be modified, sostudents often memorize conclusions such as "NaCl has a coordination number of 6" withoutunderstanding how structure determines properties. Professional crystallography software is powerfulbut far too complex for students, while online viewers are mostly read-only and cannot answer"what-if" questions. This project presents Crystallographer, an offline crystal exploration tool built on a four-layerarchitecture: (1) a natural-language understanding layer powered by on-device Gemma 4 E2B, whichtranslates student requests into standard tool calls; (2) a device-independent crystal structurefile layer; (3) a deterministic display and computation layer for 3D views, plane projections,cross-sections, density, packing fraction, coordination number, radius ratio and lattice energy;and (4) a teaching layer with guided task cards. The model is converted locally from HuggingFacesafetensors to GGUF Q4_K_M (3.19 GiB) together with a 940 MB multimodal projector (mmproj), andserved by llama.cpp with Vulkan. Using the OpenAI-compatible function-calling protocol, an agentwith 21 crystal tools builds or edits structures, validates them (electroneutrality, overlap,density, coordination) and self-corrects from the feedback. The system also embeds an offline MP-20 database of 44,217 inorganic crystals searchable bychemical formula, and a camera mode that recognizes Chinese exam questions from photos, reproducesthe unit cell and solves the problem with deterministic tools. Benchmarks on three Vulkan devicesshow that single-GPU inference reaches 1929 t/s prompt processing and 66 t/s generation, whereascross-device layer splitting is significantly slower for small models; the default strategy istherefore performance-first with multi-device options. A deterministic engine verified by 38 unittests matches literature values (e.g., NaCl density 2.163 g/cm3, coordination number 6, radiusratio 0.564), and Playwright end-to-end tests cover the full natural-language workflow.