Skip to content

VALIDATION.md

Latest commit

 

History

History
85 lines (57 loc) · 5.38 KB

VALIDATION.md

File metadata and controls

85 lines (57 loc) · 5.38 KB
Raw

Validation record

This document records the verification scope of this source snapshot. It is not a claim about untested configurations.

Environment

  • Host: JAIST Linux CPU/GPU server
  • Host application: ARCH with HDF5 and Helmholtz EOS
  • Integrated revision: Shiro-Akane/ARCH@23cfe9d
  • Reference source: supplied Timmes public_iso7.f90, public_aprox13.f90, public_aprox19.f90, public_aprox21.f90, and public_nse.tbz
  • Thermodynamic states: T = 1e9, 2e9, 5e9 K; rho = 1e6, 1e8 g cm^-3

Species registry

Selected network Expected Observed
iso7 7 7
aprox13 13 13
aprox19 19 19
aprox21 21 21

Names and ordering were checked against the corresponding network headers.

Direct original-Fortran parity

The comparison includes the species RHS, complete analytic Jacobian (including the Timmes temperature row and column), backward-Euler LHS = I - dt * J, nuclear-energy derivatives, and Helmholtz cv.

Network RHS relative error Jacobian relative error LHS relative error Helmholtz cv relative error
iso7 2.383748949565e-15 2.940236373588e-15 2.856278354823e-15 5.705278821529e-16
aprox13 3.596869184109e-14 2.850246815720e-13 2.850309327978e-13 1.331631293002e-15
aprox19 9.991734850962e-15 5.467476955907e-15 5.492012250384e-15 3.753488177566e-16
aprox21 3.997143540823e-14 1.092264838562e-14 1.091781737534e-14 3.947012873172e-16

All reported quantities are below the required 1e-12 threshold. The largest value is the aprox13 LHS difference, 2.850309327978e-13.

Temperature derivative convention

The ODE uses the original Timmes self-heating convention:

dT/dt = enuc / cv
J(T,j) = (d enuc / dX_j) / cv
J(T,T) = (d enuc / dT) / cv

The species temperature column and nuclear-energy temperature row use analytic reaction-rate derivatives; no temperature finite difference is used. Matching the original Timmes Jacobian, the implementation does not add quotient-rule derivatives of cv to the ODE Jacobian.

Generic online NSE

src/physics/nse/nse_solver.h implements the Timmes Saha equilibrium equations as NSESolver<NetType>. The solver reads NUM_SPECIES, AION, ZION, total binding energies, and statistical weights from the selected network; no isotope count or 47-species data table is hard-coded in the algorithm.

The validation covered:

  • eight direct comparisons with the original 47-species Timmes program over T = 2.5e9--1e10 K, rho = 1e6--1e9 g cm^-3, and Ye = 0.47--0.55;
  • 48 fixed-state solves across all four translated networks;
  • 84 variable-Ye states for each of aprox19 and aprox21 over Ye = 0.40--0.60;
  • an aprox19 projection coupled to the real Helmholtz EOS at rho = 4.322e7 g cm^-3, T_old = 4.67e9 K;
  • a 64 x 16 high-temperature CellularDet step at one and sixteen OpenMP threads.

The worst single-species absolute mass-fraction difference from the original 47-species program was 4.897193761622e-13. The aggregate L1 difference in that state was 1.363608769782e-12; this aggregate must not be misreported as a sub-1e-12 quantity. The worst variable-Ye mass and charge residuals were 6.49e-16 and 6.98e-13, respectively. The real-EOS relative energy-closure residual was 7.5678598227e-14. All 26 datasets from the one- and sixteen-thread CellularDet outputs were bitwise identical.

This is network-constrained NSE. iso7 and aprox13 contain only Z/A = 0.5 nuclei and therefore admit the restricted equilibrium only at Ye = 0.5. They must not be treated as replacements for the full 47-species equilibrium set.

OpenMP determinism

A 64 x 16 aprox13 case ran for three burning steps with NSE bypass disabled and a 3e9 K hot state. One-thread and sixteen-thread output had the same 20 HDF5 datasets and was bitwise identical at steps 0, 1, 2 and 3. Newly produced Ne20 confirmed execution of the network, ODE and DenseLU burn path.

Parallel execution is across independent cells with dynamic OpenMP scheduling. Network, Jacobian, ODE and small dense-LU loops use SIMD where appropriate; the implementation does not create nested thread teams for an individual 8--22 equation matrix.

MPI and CUDA scope

  • The network functions use process-local inputs and stack-local work arrays, so they can be called inside an MPI-decomposed host application. MPI communication and domain decomposition are not implemented by this module.
  • The device-qualified and cross-architecture validated CUDA network is currently aprox13.
  • For translated aprox13, CPU-vs-CUDA comparisons over the same six states produced RHS 5.207022670405e-13, Jacobian 5.094162745733e-13, LHS 5.094258977149e-13, and Helmholtz cv 7.834375032555e-15.
  • CUDA verification of iso7, aprox19, and aprox21 is not claimed in this snapshot.

Exclusions and physical scope

  • Weak-rate hooks irpen, irnep, and irn56ec
  • sneut5 neutrino losses
  • Tabulated reaction-rate mode
  • A resolved, converged cellular-detonation solution

The first three exclusions match the documented analytic-rate validation scope. Local RHS/Jacobian/LHS parity does not by itself validate a long-time multidimensional detonation; ZND/CJ initialization or adequate mesh/AMR convergence remains necessary.