| # Model Card: Lithotheque Edge AI Models | |
| ## Model Details | |
| * **Developer:** GeoStratum | |
| * **Model Versions:** `VisionAiManager` (MobileNetV5-300m / MobileNetV4-Large) and `LocalAiManager` (Metrology). | |
| * **Model Type:** Convolutional Neural Networks (CNN) optimized for Edge Inference (INT8/INT4 Quantization). | |
| * **License:** Proprietary (GeoStratum). | |
| * **Training Data:** Hybrid corpus of proprietary metrology images and open-source MIT-licensed geological datasets (Udayl, Stealth Technologies). | |
| ## Intended Use | |
| * **Primary Use Case:** Offline, on-device geological and lithological classification for field geologists, students, and enthusiasts. | |
| * **Secondary Use Case:** Scale detection and topographical metrology mapping via reference objects. | |
| * **Target Environment:** Mobile devices operating in remote areas without internet connectivity. | |
| ## Out-of-Scope Use | |
| These models are designed for educational and preliminary field classification. They **MUST NOT** be used for: | |
| * **Critical Engineering Decisions:** Assessing rock mechanics, stability, or load-bearing capacities for civil engineering or construction. | |
| * **Financial/Mining Exploration:** Making high-stakes financial decisions regarding resource extraction or ore grade estimation. | |
| * **Safety:** Detecting hazardous environments or toxic minerals for safety compliance. | |
| *Always consult laboratory analysis (e.g., XRD, XRF) for definitive geological validation.* | |
| ## Metrics & Performance | |
| * **Accuracy (Standardized Multi-Source Benchmark):** | |
| * **MobileNetV5-300M (Premium / Standard Tier):** Up to **88.33% Top-1 / 92.67% Top-3** in baseline FP16/BF16, **83.33% Top-1 / 89.33% Top-3** in **INT8 (W8A8)**, and **80.33% Top-1 / 87.67% Top-3** in **W4A8 Full Integer**. | |
| * **MobileNetV4-Large (Balanced / Legacy Tier):** **78.78% Top-1 / 91.37% Top-3** in baseline FP16/BF16, **77.84% Top-1 / 90.42% Top-3** in **INT8**, and **73.91% Top-1 / 87.15% Top-3** in **W4A8** (reaching >91% Top-1 on in-situ daylight field subsets). | |
| * **Latency:** Sub-3 ms on dedicated NPU hardware for MobileNetV4-Large (~400 FPS steady-state), and 12–18 ms for 1024×1024 Scale Detection. | |
| * *For comprehensive hardware profiling, please refer to the main `README.md`.* | |
| ## Limitations & Bias | |
| * **Occlusion & Surface Conditions:** Accuracy degrades if the rock sample is heavily covered in mud, moss, or artificial coatings. Freshly broken surfaces yield the most accurate predictions. | |
| * **Lighting Dependency:** Extreme low-light conditions or harsh direct camera flash may alter texture perception, slightly increasing the margin of error. | |
| * **Micro-crystalline vs. Cryptocrystalline:** Some visually identical rocks require chemical or microscopic analysis for definitive differentiation. The model predicts the highest mathematical probability based strictly on visual morphological features. | |