AEGIS-Qwen: Domain-Specific LLM for Casimir Physics
Fine-tuned Qwen2-7B-Instruct for the AEGIS anti-gravity research platform.
Training
- Base model: Qwen/Qwen2-7B-Instruct
- Hardware: AMD Instinct MI300X (192GB VRAM)
- Method: LoRA (r=16, alpha=32) on all attention + MLP layers
- Data: 20 domain-specific instruction pairs covering Casimir forces, Drude-Lorentz metamaterials, negative-index physics, and experimental design
- Epochs: 4
- Final loss: 1.162 (from 1.877)
- Token accuracy: 72.6%
Domain Coverage
- Casimir force computation (sphere-plate, plate-plate)
- Drude-Lorentz metamaterial electromagnetic response
- Negative-index (NI) band identification
- ZPE-gravity coupling constant (κ) detection thresholds
- A/B experimental comparison methodology
- Apparatus optimization for κ_min sensitivity
Built For
AMD Developer Hackathon 2025 — Track 2: Fine-Tuning on AMD GPUs
Platform
This model powers the AI assistant in the AEGIS Complete Platform, a computational physics research tool for designing Casimir force experiments in negative-index metamaterials.
