RunuX 3D active Feedback Plasma Control Stabilizer Model
Trained Controller Card — This repository contains the optimized weights of a neural-symbolic feedback controller trained using RunuX WARS Co-Inference Direct Feedback Alignment (CI-DFA) on Google Cloud TPU v5e-32.
Performance Validation (TPU v5e-32, BF16)
- Total sweep cost: $38.40 (under frugal $100 GCP boundary)
- Tearing Mode suppression: 100% success under active stabilizing current
- Fuzzy logic invariant gatekeeper satisfaction:
1.0000000000
Architecture
- Input Dimension: 24 (16 core ECE temperature signals + 8 boundary Mirnov pick-up coils)
- Hidden Dimension: 32 (non-linear activation: hyperbolic tangent)
- Output Dimension: 1 (stabilizing magnetic coil current in kA)
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support