SCPN-Fusion-Core

Evidence-bounded neuro-symbolic tokamak control. Native plasma-solver kernels, a Rust real-time control backend, and fail-closed fusion-validation infrastructure — for equilibrium, transport, gyrokinetic, MHD, runaway-electron, impurity, and free-boundary campaigns.

PyPI Package Physics Solvers Evidence Boundary
277Python Modules
13Rust Crates
425Test Files
0.52 µsRust PID Kernel
118Core Physics Modules
54Control Modules
7MHD Criteria
8SPARC Equilibria

Install

# Core framework
pip install scpn-fusion

# Grad-Shafranov equilibrium kernel
scpn-fusion kernel

# Tokamak flight simulator
scpn-fusion flight

Control-First Pipeline

From a control idea to a reproducible evidence package — without pretending that partial research kernels are full production reference solvers.

Petri-Net Contract
inspectable spec
→
Controller
SNN / NMPC / H∞
→
Plasma Solver
GS / transport / GK
→
Fail-Closed Gates
accepted / blocked
→
Evidence Package
reports + checksums

Key Capabilities

Native Plasma Solvers
Grad-Shafranov equilibrium, 1.5D transport (+ QLKNN and FNO surrogates), MHD stability (7 criteria), neoclassical, disruption chain, ELM/MARFE/L-H transition, runaway electrons, pellet injection, and 3D equilibrium.
Gyrokinetic Three-Path
Three transport lanes with explicit fidelity limits: external-code bridges, a bounded native GK research solver (linear eigenvalue + nonlinear 5D operator contracts), and neural surrogates. Full GENE/CGYRO/GS2 parity stays fail-closed.
Neuro-Symbolic Control
Controllers compile from inspectable Petri-net contracts: SNN compiler, NMPC-JAX, H∞, μ-synthesis, gain-scheduled, fault-tolerant, safe RL (PPO), free-boundary tracking, burn control, RZIP, RWM feedback, EKF state estimation.
Rust Real-Time Backend
13 Rust workspace crates. PID kernel measured at 0.52 µs P50; full-order GS solve 413 µs SOR (33×33). PyO3 bindings, ML inference, native math.
Fail-Closed Evidence Gates
Production-parity lanes stay fail-closed until external same-case evidence exists. Every public claim carries an accepted/blocked row status, a reproducibility command, and a checksum.
Reference-Code Bridges
Adapter layer and same-case harnesses for GENE, CGYRO, GS2, DREAM, Aurora, STRAHL, FreeGS, GEQDSK, and IMAS/OMAS — for reproducible, comparable validation.

Evidence Boundary

What is validated today — and what remains blocked

This is a control-algorithm development and validation framework with explicit fidelity boundaries — not a replacement for TRANSP, JINTRAC, GENE, CGYRO, GS2, DREAM, Aurora/STRAHL, or EFIT, and not IEC 61508 / SIL certified.

Validated today
Local controller contracts, reduced-order replay lanes, selected Grad-Shafranov / operator-source checks, native kernel benchmarks, formal proof slices, fail-closed benchmark gates, the accepted public FreeGS same-case free-boundary strict-parity lane, and 1 accepted public reference artefact.
Blocked / fail-closed
Full GENE/CGYRO/GS2 nonlinear turbulence parity, full Vlasov-Maxwell parity, DREAM kinetic parity, independent mechanistic Aurora/STRAHL recycling validation, EFIT-grade free-boundary reconstruction, full 3D nonlinear MHD, and production MPI / multi-GPU scaling.
Full Evidence Boundary → Benchmarks

Explore

Physics Solvers →
Grad-Shafranov, transport, gyrokinetic three-path, MHD stability, neoclassical, runaway electrons, free-boundary.
Benchmarks →
Rust kernel latencies, surrogate accuracy, disruption-rate campaign, SPARC GEQDSK, confinement data.
Evidence →
Accepted vs blocked lanes, fail-closed gates, reproducibility commands, and claim-evidence map.
277 Modules →
9 capability layers: physics, control, phase dynamics, diagnostics, engineering, nuclear, compiler, I/O, Rust.
Licensing →
Dual licensing: AGPL-3.0 open path and commercial path for proprietary reactor programs.
Changelog →
Version history and recent validation milestones.