| --- |
| title: SimQuantum Tuning Lab |
| emoji: ⚛ |
| colorFrom: gray |
| colorTo: green |
| sdk: streamlit |
| sdk_version: 1.57.0 |
| app_file: app.py |
| pinned: true |
| license: mit |
| short_description: Autonomous quantum dot tuning agent — AMD MI300X + Qwen2.5 |
| --- |
| |
| # ⚛ SimQuantum Tuning Lab |
|
|
| **Autonomous quantum dot tuning · AMD Developer Hackathon 2025** |
|
|
| SimQuantum is a 6-stage POMDP agent that autonomously tunes a double quantum dot |
| device to the **(1,1) charge state** — one electron per dot — required for spin |
| qubit operation. |
|
|
| ### What runs here (CPU demo) |
| - ✅ CIM physics simulator — real semiconductor physics |
| - ✅ Charge stability diagram visualisation |
| - ✅ Particle filter belief state |
| - ✅ 5-model CNN ensemble classifier (91.4% val acc, 51k training diagrams) |
| - ✅ Bayesian optimisation planner |
| - ✅ HITL safety critic |
| - ⬡ Dr. Q (Qwen2.5-1.5B) — connect your own vLLM endpoint in the sidebar |
|
|
| ### Full GPU stack |
| The complete system runs on **AMD Instinct MI300X** via ROCm with |
| Qwen2.5-1.5B-Instruct served via vLLM. To connect Dr. Q, set the |
| vLLM endpoint in the sidebar to your MI300X instance. |
|
|
| ### Architecture |
| ``` |
| BOOTSTRAPPING → COARSE_SURVEY → HYPERSURFACE_SEARCH → CHARGE_ID |
| ↓ ↓ ↓ ↓ |
| DQC gate 32×32 sweep 16×16 local scan CNN ensemble |
| + OOD detector |
| ``` |
| Navigation (→ VERIFICATION) is Phase 3 — active development. |
|
|
| ### Links |
| - GitHub: https://github.com/k1151msarandega/Agentic-Semiconductor-Quantum-Device-Tuning |
| - AMD Developer Hackathon: https://lablab.ai/ai-hackathons/amd-developer |
|
|
