Skip to content

Competitive Evidence — SCPN Control and Adjacent Fusion Systems

Evidence date: 2026-08-28. SCPN Control version: v0.23.0. The machine-readable source registry is docs/_data/competitive_evidence.json.

This page compares documented scope and evidence. It is not a product ranking or a claim that every relevant system has been surveyed. A blank or undocumented capability is reported as not assessed, never converted into “does not exist.”

Reading the comparison

Each external row is bound to an exact public release or a primary paper. Project documentation supports capability statements; it does not by itself establish accuracy, deployment readiness, or equivalence to another code.

No cross-project numeric result is admitted in this revision. A quantitative comparison requires the same problem, inputs, precision, tolerances, convergence criteria, warm-up and compilation treatment, sample definition, isolated hardware, and failure accounting. The previous page mixed local loopback, bare-kernel, solver-step, reconstruction-iteration, and full control-cycle timings; those values answer different questions and are not retained as a ranking.

Assessed systems and strongest documented evidence

System and assessed artifact Primary role Strongest evidence relevant to this comparison SCPN Control boundary
SCPN Control v0.23.0 snapshot f1f6e36d Neuro-symbolic plasma-control research toolkit and evidence-admission layer Executable classical, robust, predictive and neuro-symbolic control surfaces; formal/admission contracts; Python, Rust, TypeScript, Lean and C++ interfaces Research software: not a commissioned PCS; external-code parity, target-hardware timing and facility validation are separate fail-closed claims
TORAX v1.4.3 Differentiable JAX core-transport simulator Coupled heat, particle and current-diffusion PDEs; differentiable nonlinear solves; QLKNN, direct TGLF and IMAS core-source integration Reduced transport and TORAX-shaped interfaces are not an identical-input TORAX result
FUSE v1.2.0 Integrated fusion-device simulation and design Plasma physics, engineering, control, balance-of-plant and costing actors; constrained multi-objective optimisation; IMAS/OMAS interoperability SCPN Control supplies control and evidence components, not an equivalent whole-device design environment
FreeGS v0.8.2 Static free-boundary Grad–Shafranov equilibrium Coils, plasma profiles, machine circuits, inverse constraints and G-EQDSK in a transparent Python reference Code-to-code agreement is useful verification, not experimental truth
FreeGSNKE v3.0.1 Static and evolutive free-boundary equilibrium and circuits Fourth-order Newton–Krylov solves; active/passive structures and magnetic diagnostics; MAST-U/EFIT++ validation literature Equivalent evolutive and shot-matched validation is not currently admitted in SCPN Control
DREAM v26.5 Fluid-kinetic disruption and runaway-electron simulation Coupled nonlinear fluid/kinetic equations with specialised runaway and shattered-pellet-injection physics SCPN Control's disruption and mitigation models are control-grade models, not a DREAM-equivalent kinetic solver
PROCESS v3.4.2 Reactor systems design Integrated reactor physics, engineering, costing and optimisation with versioned/Zenodo release lineage SCPN Control is not a reactor design or costing system
OMAS v0.97.3 ITER IMAS data-model interoperability Multiple file/database formats mapped to the IMAS model and used as an OMFIT integration substrate IMAS-facing contracts do not yet demonstrate comparable ecosystem breadth
TCV deep-RL magnetic control, doi:10.1038/s41586-021-04301-9 Learned magnetic control demonstrated on a tokamak Peer-reviewed 10 kHz closed-loop TCV experiments, zero-shot simulation-to-hardware transfer and 19-coil actuation across several plasma configurations SCPN Control has no commissioned tokamak actuation result; simulation and offline replay are different evidence classes
EFIT-AI GPU reconstruction, doi:10.1145/3624062.3624607 Performance-portable inverse equilibrium reconstruction Named NVIDIA, AMD and Intel accelerator evaluation of GPU-offloaded EFIT reconstruction kernels No admitted identical-input, full-output, named-GPU comparison exists in SCPN Control

Where external evidence is currently stronger

These are evidence differences, not statements about the worth of either project.

  • Facility operation: the TCV deep-RL work reports actual tokamak actuation; SCPN Control reports simulation, local runtime and offline evidence only.
  • Evolutive free-boundary validation: FreeGSNKE documents self-consistent circuit/plasma evolution and peer-reviewed MAST-U/EFIT++ validation. SCPN Control does not currently admit equivalent evidence.
  • Core-transport depth and integration: TORAX documents a dedicated, differentiable transport stack with direct TGLF and expanding IMAS support.
  • Integrated design breadth: FUSE and PROCESS cover whole-device engineering, costing and optimisation scopes outside SCPN Control's role.
  • Disruption/runaway fidelity: DREAM is purpose-built around fluid-kinetic runaway-electron physics and has a dedicated validation/publication lineage.
  • Data and workflow adoption: OMAS/OMFIT expose a broader public interoperability and module ecosystem than SCPN Control currently demonstrates.
  • Accelerator reconstruction evidence: the EFIT-AI study reports named multi-vendor GPU measurements. SCPN Control's equilibrium timing evidence does not satisfy that matched boundary.

What SCPN Control demonstrates today

Within this assessed set, the sources document a different emphasis for SCPN Control: neuro-symbolic controller construction, executable safety/admission contracts, formal traceability, polyglot runtime surfaces, and evidence objects that keep simulation, external-code, hardware and facility claims separate. That observation is limited to the sources below and is not a market-wide novelty claim.

Current package counts are generated rather than duplicated here. See the capability manifest. Scientific and deployment limits are maintained in production readiness and validation deficiencies.

Quantitative comparison status

The admitted cross-project quantitative comparison set is currently empty. Repository-local Python/Rust or controller-baseline comparisons remain useful for implementation decisions when their own provenance is complete, but they do not rank SCPN Control against TORAX, FUSE, FreeGSNKE, DREAM, TCV control or EFIT-AI.

When a matched external case becomes available, the public result must preserve all runs, including failures, and report the exact code versions, inputs, hardware/load, warm-up, precision, tolerances, samples and admission decision.

Primary sources