Skip to content

CAPABILITIES_AND_USECASES.md

Latest commit

 

History

History
31 lines (19 loc) · 3.72 KB

CAPABILITIES_AND_USECASES.md

File metadata and controls

31 lines (19 loc) · 3.72 KB
Raw

SCPN Control: Architectural Capabilities & Use-Case Scenarios

1. Core Architecture: Neuro-Symbolic Actuation

scpn-control is a control research layer for tokamak plasma scenarios. It processes diagnostic inputs, computes stability and transport quantities, and evaluates candidate control actions behind fail-closed admission gates. It does not have live tokamak actuator authority.

Technical Specifications

  • Strict Fallback Degradation: Compute dispatch attempts scpn-control-rs (Rust) first, falls back to sc-neurocore (JAX/GPU acceleration) where available, and finally to strictly typed pure NumPy/Python. A missing native backend degrades the compute path; it does not crash the loop.
  • Control-Loop Latency: Measured latency artefacts (per-kernel and end-to-end percentiles, with recorded host, affinity, and isolation context) live in validation/reports/ and are regression-gated in CI. All committed figures carry an evidence class and are local-regression evidence, not facility or production timing claims; see VALIDATION.md for the current claim boundary.
  • Stochastic Petri Nets (SPNs): Supervisory logic is expressed as SPNs executed via populations of Leaky Integrate-and-Fire (LIF) neurons, providing graceful behaviour under diagnostic noise and sensor dropouts.

2. Dynamic Domain Actuation

The architecture is domain-configurable: the controller updates its $K_{nm}$ coupling matrix online via the AdaptiveKnmEngine, and layer adapters map domain telemetry onto the phase-coupling model.

  • Plasma Control Loop: Uses $\beta_N$, Mirnov RMS, and Greenwald limits in research scenarios. Singular Value Decomposition (SVD) and $H_\infty$ control synthesis produce candidate magnetic coil voltage changes ($\Delta V_{coil}$) for Neoclassical Tearing Mode (NTM) suppression studies; these are not live actuator commands.

3. Gyrokinetic Transport (v0.17.0)

Three gyrokinetic transport paths are implemented:

  • Path A — External GK Coupling: Interfaces to five first-principles codes (TGLF, GENE, GS2, CGYRO, QuaLiKiz) via subprocess with automatic input deck generation and output parsing; fail-closed when the external executable is absent. Transport solver mode: transport_model="external_gk".
  • Path B — Native Linear GK Solver: Self-contained ballooning-space eigenvalue solver (Miller geometry, Sugama collision operator, Gauss-Legendre velocity grid), benchmarked against Cyclone Base Case reference ranges (Dimits et al. 2000). Quantitative cross-code agreement with external GK codes is an open validation target; see VALIDATION.md.
  • Path C — Hybrid Surrogate+GK Validation: A QLKNN-style surrogate runs at microsecond speed; Mahalanobis + ensemble OOD detection triggers GK spot-checks, with a correction layer (multiplicative/additive/replace with EMA smoothing) and online retraining.
  • SCPN Phase Bridge: An opt-in bridge maps GK growth rates and diffusivities onto the 8-layer UPDE Kuramoto coupling matrix (P0↔P1 microturbulence↔zonal flows, P1↔P4 zonal↔transport barrier, P3↔P4 sawtooth↔barrier). The bridge currently records coupling metadata; closed-loop actuation through the phase layer is a research lane, not a shipped control path.

4. Use-Case Scenarios

  • Nuclear Fusion (Digital Twin & Flight Sim): A full tokamak_flight_sim.py and tokamak_digital_twin.py allow continuous integration testing of real-time control algorithms against a synthetic plasma model before any hardware exposure.
  • Industrial Control Research: The same SPN/phase-coupling architecture is applicable to nonlinear multi-actuator control problems (manufacturing, chemical reactors, grid balancing) as a research substrate; no production deployment in these domains is claimed.