ALADIN EGLM v02: Div-B Guardian for a Constrained-Transport Cylindrical GPU MHD Driver — Exact Monitor, Hodge Projection and GLM Cleaning, Verified on CPU and GPU
Year: 2026
Type: article
Abstract: EGLM v02 is the div-B guardian of the ALADIN native-cylindrical GPU MHD Driver (Z-pinch and J×B research). The Driver advances B with upwind constrained transport (UCT-EMF v02), which preserves the discrete divergence to round-off by construction, so GLM cleaning is not run inside the normal step. EGLM v02 provides what a constrained-transport Driver needs:• monitor() — every step: the exact discrete divergence (net magnetic flux out of each cell), identical to the quantity constrained transport preserves; for the run log and the MakeResults evidence bundle.• project() — on demand, exact: discrete Hodge projection with the kernel's own operators (FFT in φ and z, direct tridiagonal solve in r). One call drives div B to round-off and leaves every edge circulation — the current J — unchanged.• clean() / glm_step() — the author's GLM splitting (half step on B, full step on ψ, damping, half step on B) for iterative cleaning.• eglm_sources() — the EGLM momentum and energy sources −(div B)B and −B·∇ψ (Dedner et al. 2002).One CUDA/C++ source compiles for GPU (NVRTC/CuPy) and CPU (g++/OpenMP), FP64 and FP32, on the same grid contract as UCT_CURL v1.3 and UCT-EMF v02. Certification campaign EGLM-01 (frozen source, predeclared thresholds, CPU leg + NVIDIA L4 GPU leg, compute-sanitizer): 18 tests — 17 VERIFIED, 1 LIMITATION by design, 0 BLOCKER; memcheck 0 errors, racecheck 0 hazards.• Independent NumPy oracle: divergence, ψ update, face gradient, sources agree to ≤1.2e-15.• Monitor equals UCT-EMF's own divergence to 4e-16; reads round-off on constrained-transport fields.• Mimetic: the gradient correction leaves all edge circulations unchanged (5.6e-15); div(grad) equals the cylindrical 7-point Laplacian.• project(): div B from ~1e-2 to 5e-16 in one call on annulus and axis grids, FP32 3e-7; recovers B0 exactly from B0 + ∇χ (6e-16); leaves clean fields unchanged.• GLM splitting follows the Störmer–Verlet map to 6e-16; damping exact. Addendum after review (kernel and frozen run unchanged): (A1) E12 restated per grid — the iterative GLM clean() converged to 1e-12 up to 24×32×16 but not on the 48×64×8 axis grid within 30000 iterations (2.8e-8); iterations grow ~N^1.6–1.9 and the divergence history is non-monotone (damped-wave behaviour); classified as a convergence-rate scaling restriction, not a correctness failure; clean() is not recommended on large or axis grids. (A2) project() on grids up to 128×64×64 VERIFIED on CPU and GPU: FP64 ≤6.2e-16, FP32 ≤2.9e-7, circulations kept to 3.5e-16. (A3) compute-sanitizer on 12 large and odd grid configurations: memcheck 0 errors, racecheck 0 hazards (scope: executed workloads). The GPU leg of the addendum is included as a verbatim console transcript. Predecessor: EGLM (Zenodo record 24291219). The bundle contains the kernel, the oracle, the campaign, both legs' raw results verbatim, sanitizer logs, merged results, the addendum, the verification report (PDF) and the original EGLM. code/uct_emf_v02.py is an unmodified copy of UCT-EMF v02 (Zenodo 24394846), included only as a test dependency. Kernel source SHA-256: d4753fd84a078a70fd53c0b0095eb3d96086b7cdad51073f33c0e132db854713 AI assistant Team: SuperGrok, Claude, ChatGPT
Author: Bucurenciu, Mihai Alexandru
Language: English
Subfield: Nuclear and High Energy Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234176006
