Add restart capability to the fire behavior model by masih-e · Pull Request #39 · NCAR/fire_behavior

masih-e · 2026-08-26T20:15:01Z

Description of changes:

The CFBM has had no way to stop a simulation and resume it.
This feature adds a restart capability. A run can write its dynamic fire state to a
NetCDF file at a fixed interval, and a later run can be initialized from one of
those files instead of a cold start. It covers standalone runs, serial and MPI,
and runs coupled to WRF.

Two options are added to the &time namelist block:

Option	Type	Default	Meaning
`restart`	logical	`.false.`	Resume from `fire_restart_<start datetime>.nc`
`restart_interval`	integer	`-1`	Restart write interval in seconds; `-1` disables writes

Both default to the previous behavior, so existing namelists are unaffected.

Changes by file:

state/state_mod.F90 carries the Read_restart and Write_restart
type-bound methods, Build_restart_file_name, and the MPI helpers they need.
A restart read validates the grid metadata stored in the file against the
current configuration before touching any state, so a mismatched restart file
fails early with a message instead of producing a silently wrong run.
io/netcdf_mod.F90 adds attribute writers (Add_netcdf_att) and
MPI-aware readers and writers for distributed 2D and 3D fields.
driver/initialize_mod.F90 routes the real-case restart path through a
new Init_fire_state_from_restart, leaving the existing geogrid
initialization untouched. The WRF-coupled path reads the restart after the
fire components are initialized.
driver/fire_behavior.F90 reads the restart state at startup in place of
Save_state, and writes a restart every restart_interval seconds.
restart_interval must map to a whole number of time steps or the run stops
with a message.
io/namelist_mod.F90 adds the two options, their broadcast, and a range
check.
physics/fmc_wrffire_mod.F90 drops its unused state_mod and
namelist_mod imports. This one is load-bearing rather than cosmetic:
state_mod now uses fmc_wrffire_mod for the moisture restart fields, so
leaving those imports in place would create a circular module dependency.

Type of change

New feature (non-breaking change which adds functionality)

Tests conducted:

All testing on Derecho (NCAR) using the repository environment
env/derecho/gnu-12.2.0: GCC 12.2.0, cray-mpich 8.1.25, NetCDF-Fortran 4.9.2,
HDF5 1.12.2, ESMF 8.5.0, CMake 3.26.3.

Builds. Three configurations, all clean with no errors or new warnings:
serial (--mpi-off), MPI (default), and --esmx (which includes NUOPC).

Registered ctest suite, run on the --esmx build so every registered case
is reachable:

test7	test8	test7esmf	test8esmf	test7esmx	test8esmx	testx
pass	pass	pass	pass	pass	pass	pass

7 of 7. The NUOPC cases matter here because the cap uses both initialize_mod
and state_mod, and this PR touches both.

Restart regressions. These run a case straight through, then run it again
stopping and restarting partway, and compare the two. Restart test scripts are
not included in this PR (see below), but the results are:

Serial build, 20 of 20 checks:

test7_restart	test8_restart	test9	test9_restart
5/5	5/5	5/5	5/5

MPI build on a compute node, 30 of 30 checks:

ranks	test7_restart	test8_restart
1	5/5	5/5
2	5/5	5/5
4	5/5	5/5

The multi-rank runs are the ones that exercise the distributed restart read and
write and the halo exchange following a restart read. Restarting at 2 and 4
ranks reproduces the continuous run, so the domain decomposition is handled
correctly on the restart path.

57 passing checks in total across the three build configurations.

Not covered. Two gaps, stated explicitly:

The WRF-coupled restart path (Init_fire_state_within_wrf calling
Read_restart) is exercised by a separate WRF harness, and not tested here.
Rank counts above 4 have not been tried.

On the test scripts. The restart test scripts (test7_restart.s,
test8_restart.s, test9.s, test9_restart.s, the test9 case directory, and
the tests/CMakeLists.txt registration) are deliberately held out of this PR
pending the separate discussion about how tests should be organized. They are
ready and can be added here or in a follow-up PR, whichever is preferred.

Documentation:

Documentation has been updated for these changes
No documentation update needed

Documentation was left out to keep the diff focused on
the code. Happy to add it to this PR if preferred, otherwise it will follow
separately.

Contributors:

@masih-e (Masih Eghdami:masih@ucar.edu)
@mefrediani (Maria Frediani:frediani@ucar.edu)
@pedro-jm (Pedro Jimenez:jimenez@ucar.edu; reviewer)

LLM/coding assistant technology used: the restart implementation was developed
with assistance from Codex GPT-5.5 and Claude Opus 4.7. Claude Opus 5 assisted
with the pre-review cleanup of this branch.

Further information:

state/state_mod.F90 grows from 1063 to 1965 lines in this PR, which is more
than that file should carry. A restart-specific module is worth doing as a follow-up. This
change is not possible with the current code because the restart calls depend on the fire
state type.

Lets a simulation write its dynamic fire state to a NetCDF restart file at a fixed interval and resume from one of those files, for standalone runs (serial and MPI) and for runs coupled to WRF. Namelist (time block): restart - resume from fire_restart_<start datetime>.nc restart_interval - restart write interval in seconds, -1 disables state_mod: Read_restart and Write_restart type-bound methods, plus Build_restart_file_name and the MPI helpers they need. A restart read validates the grid metadata stored in the file against the current configuration before touching any state, so a mismatched restart file fails early instead of producing a silently wrong run. netcdf_mod: Attribute writers (Add_netcdf_att) and MPI-aware readers and writers for distributed 2D and 3D fields. initialize_mod: Init_fire_state_from_restart takes the real-case restart path, so the existing geogrid initialization is left as is. The WRF-coupled path reads the restart after the fire components are initialized. fire_behavior: Reads the restart state at startup instead of Save_state, and writes a restart every restart_interval seconds. restart_interval must map to a whole number of time steps or the run stops with a message. fmc_wrffire_mod: Drops the unused state_mod and namelist_mod imports. state_mod now uses fmc_wrffire_mod for the moisture restart fields, so these imports would otherwise form a circular module dependency. The code was created with assistance from Codex GPT-5.5 and Claude Opus 4.7. Co-authored-by: mefrediani <frediani@ucar.edu>

Move restart-specific state code and helpers into a new state_restart_submod.F90 submodule.

masih-e · 2026-09-25T15:00:55Z

Updated the PR with the state restart refactor. Restart-related implementations were moved out of state_mod.F90 into a state submodule without changing the public state_mod interfaces or restart logic.

Move Distribute_global_var2d out of netcdf_mod and into mpi_mod so the domain-distribution helper lives with the other MPI utilities. Keep the restart-specific variable names at the restart read call sites, but use generic naming and messages in the MPI helper itself.

Move standalone restart metadata initialization into state initialization and remove the initialize_mod restart wrapper. Move restart output interval validation and write timing into state_fire_t%Handle_restart so the main driver loop stays focused on high-level model actions.

Move the standalone restart read/save decision into Init_fire_state so the main fire_behavior driver stays focused on top-level model control. Move the per-step model-clock consistency check into state_fire_t%Check_model_clock. Keep it separate from restart scheduling because the clock check is a general model consistency check, not restart-only logic.

masih-e · 2026-09-25T21:52:07Z

Updated with review cleanups:

moved generic 2D distribution helper into mpi_mod
moved standalone restart initialization/read-save flow out of the main driver
moved restart output scheduling into state_fire_t%Handle_restart
moved model clock consistency check into state_fire_t%Check_model_clock

mefrediani · 2026-10-16T00:00:49Z

@masih-e I checked the restart branch against the changes we've made recently for the regression. This is what I found:

For the current conflict, you can accept both (keep the comment and the call to check_model_clock)

However, I found two substantive interactions with the recent merges:

Observed-perimeter activation is not restored. The new fire_perimeter_ignited flag defaults to false, and Add restart capability to the fire behavior model #39 neither saves nor restores it. After restarting an already ignited perimeter case, the first advance would reinstall lfn_hist, overwriting the
evolved front. That activation state must survive restart. Activation logic (

fire_behavior/physics/fire_model_mod.F90

Lines 144 to 150 in 26d57b6

    
           ! Install an aligned perimeter at the inclusive interval end. It begins 
        
           ! propagating next step; zero-time perimeters were installed at initialization. 
        
           ! Mid-step activation would age fuel without proportional propagation. 
        
           ! Check_nml rejects it pending a separate review of split advances. 
        
           if (config_flags%fire_is_real_perim .and. .not. grid%fire_perimeter_ignited) then 
        
             ignition_time = anint (config_flags%fire_ignition_start_time1 / grid%dt) * grid%dt 
        
             if (ignition_time <= time_end) call grid%Ignite_fire_perimeter (ignition_time)

).

Restart I/O has incompatible fuel-array types. Recent changes made fuel_frac and fuel_frac_burnt_dt REAL64, but Add restart capability to the fire behavior model #39’s restart read, write, and halo helpers still accept default real. The combined source therefore has argument-kind mismatches. Restart
storage must also preserve double precision; converting to single precision would discard the accuracy that the recent fix introduced. Current declarations
(

fire_behavior/state/state_mod.F90

Lines 66 to 69 in 26d57b6

    
           ! Retain small consumption increments when subtracting nearly-one fractions. 
        
           real(REAL64), dimension(:, :), allocatable :: fuel_frac ! "fuel remaining" "1" 
        
           real, dimension(:, :), allocatable :: fire_area ! "fraction of cell area on fire" "1" 
        
           real(REAL64), dimension(:, :), allocatable :: fuel_frac_burnt_dt ! "fraction of fuel burnt on current dt" "-"

), restart reader
(

fire_behavior/state/state_restart_submod.F90

Lines 605 to 616 in e73f23c

    
           subroutine Read_restart_field (file_name, var_name, var) 
        
             implicit none 
        
             character (len = *), intent (in) :: file_name, var_name 
        
             real, dimension(this%ifms:this%ifme, this%jfms:this%jfme), intent (in out) :: var 
        
             call Get_netcdf_var_mpi (file_name, this%cfbm_comm, this%nx, this%ny, & 
        
                 this%ifps, this%ifpe, this%jfps, this%jfpe, var_name, var(this%ifps:this%ifpe, this%jfps:this%jfpe)) 
        
           end subroutine Read_restart_field

).

Beyond the two restart defects identified earlier, there are configuration and numerical changes to account for:

Change	Consequence for restart
[#47: timing and perimeter initialization](#47)	Preserve the corrected interval calculation and restore `fire_perimeter_ignited`. New namelist checks also apply to restart runs: valid ignition configuration, timestep-aligned perimeter activation, and `interval_atm` equal to an integer multiple of `dt`.
[#56: fuel and coordinate precision](#56)	Restart read/write and MPI helpers need to preserve the two `REAL64` fuel arrays. The new projection coefficients are reconstructed by the projection constructor, so they do not require additional checkpoint fields.
[#57: standalone wind interpolation](#57)	Atmospheric fields now follow the revised horizontal/vertical interpolation order. Restart already saves `uf`, `vf`, and roughness; I found no additional missing wind-state field. Continuity should nevertheless be checked across the first atmospheric update after restart.
[#49: interpolation precision, wind bounds, Strict builds, harness](#49)	Preserve these fixes and compare continuous/restarted runs using the same build configuration. The current regression matrix explicitly excludes restart continuity.
[#55: NUOPC wind imports](#55)	Changes coupled wind mapping; I found no additional standalone restart-file requirement.

Thus, the concrete source incompatibilities remain the missing perimeter flag and the fuel-array precision mismatch. The other changes require appropriate configuration and validation.

Write and read fire_perimeter_ignited as an integer restart attribute and broadcast it on every rank.

Read and write fuel_frac and fuel_frac_burnt_dt as NetCDF doubles. Halo exchange for the variables as doubles.

With `fmoist_freq > 1`, the first scheduled moisture update could miss the step-1 initialization, leaving class moisture and forcing history undefined. Initialize once from the first valid atmospheric state independently of cadence, apply zero elapsed moisture time during initialization, and restore that state on restart. For example, with `dt=4` and `fmoist_freq=3`, initialization is at 4 s, the update at 12 s uses 8 s elapsed time, and later updates use 12 s intervals. Time-based updates retain their existing phase. Invalid or unavailable current/previous forcing fails explicitly. The NUOPC rainfall accumulator is initialized before its first import. Checkpoints now record initialization status/time and moisture activation/cadence. Loading restores the lifecycle and validates class state, forcing history, and matching settings. Prescribed legacy checkpoints remain readable. PR #39 time-based checkpoints are accepted when their metadata establishes compatible initialized state; ambiguous legacy step-based checkpoints fail with a specific diagnostic. Configuration documentation and twelve focused lifecycle scenarios cover the correction. No new namelist options, ordinary diagnostic fields, or humidity-formula changes are included. Refs #42. Its remaining extinction, controls, and diagnostic work remains open. ### Review base and dependencies - This draft targets `develop` at `ead119eca6dfb576d181a5560273d70035f780a3`, which includes the squash merge of #39. The original local restart merge `bc205bc` and upstream `ead119e` have identical file trees but different ancestry. - Head `a87b23dae1c8b4a282ed503202cbb34579f0536f` contains six commits and six files, without historical restart commits in its commit range. The final atmospheric defaults use the original patch bounds (`ifps:ifpe`, `jfps:jfpe`) in `Set_vars_to_default`; `state/state_mod.F90` now matches develop exactly and drops out of the feature diff. - The review follow-up sets the four atmospheric fill values in `Set_vars_to_default`, initializes the precipitation-rate accumulator in NUOPC `Realize` while preserving restored rainfall, and groups external/intrinsic imports before CFBM modules. The component member remains `initialized`; its NetCDF attribute remains `fmc_initialized`. - P00 / #61 at `ed18f91` remains an explicit validation dependency absent from this feature's ancestry. The remaining foundation merge order is this PR, then #61. Update and validate P00 separately before its merge. - The earlier replayed implementation `dec7eb2` was byte-for-byte identical to validated `3e472d8`. Initialization-placement validation used separate local integration `d144a898a436efb7ef411c781368e6853c0415b3`, combining P00 and P02 at `439474f`; all seven feature files matched that revision exactly. The subsequent patch-bound restoration changes only the four atmospheric default assignments. Original branches and validation evidence remain preserved. ### Validation Earlier validation at combined revision `554d150` used GNU Strict builds on Derecho with GNU 12.2.0, Cray MPICH 8.1.25, NetCDF 4.9.2, and ESMF 8.5.0. Serial, OpenMP, MPI, hybrid, and ordinary NUOPC/ESMX execution passed. - Twelve focused initialization/forcing/timing scenarios per build; 45 Python tests; formatting and diff checks passed. - All 15 restart cases passed, including the five delayed-moisture cases previously failing in P00. There were 65 continuous/segmented model runs and 30 loader probes. - All 320 file comparisons and 12 cross-layout comparisons had zero maximum absolute error; 450 checkpoint checks passed. Loader probes comprised five accepted schema-compatible resumes and 25 expected rejections. - Existing PR suite: 36 model runs, 40 execution comparisons, and 36 approved-reference comparisons passed. Largest cross-driver absolute difference was `fgrnhfx = 0.125 W m-2`, consistent with the existing baseline. No reference or tolerance changed. - Original PR #39-era checkpoints generated with P00 were also checked: prescribed/time-based resumes agreed exactly with corresponding P02 checkpoints, and ambiguous step-based state was rejected. PBS jobs: `7890398.desched1`, `7890541.desched1`, `7890596.desched1`. Full retained evidence: `/glade/derecho/scratch/frediani/tmp/cfbm-P02-20261021T034953Z/`, including `attempt-02/results/restart-z8yehdcl/summary.json`, `attempt-02/results/pr-hw75p1af/summary.json`, and `legacy-originals/summary.json`. P01/P02/P04 combined acceptance remains pending, including shared clock precision and fractional/long-duration timing. Direct WRF-host execution/restart, NUOPC/ESMX restart, and transfer of checkpoints between MPI layouts remain untested here. Ordinary NUOPC/ESMX execution and comparisons among independent MPI layouts do not establish those restart paths. GitHub CI must pass on the final PR revision before merge; its normal PR suite does not run P00's optional restart suite. ### Initialization-placement follow-up validation Feature `439474f` matches tested P00/P02 integration `d144a898a436efb7ef411c781368e6853c0415b3` exactly. Four fresh GNU Strict builds (serial, OpenMP, MPI/NUOPC/ESMX, hybrid/NUOPC/ESMX) passed with the environment above. - All 15 restart cases pass: 65 continuous/segmented runs, 30 loading probes, 320 zero-error file comparisons (315 continuity and five accepted legacy resumes), 12 zero-error cross-layout comparisons, and 450 checkpoint checks. The restart suite's 67 CTest invocations include the twelve focused P02 lifecycle scenarios in each build. - The existing 36-run PR matrix, 40 execution comparisons, and 36 approved-reference comparisons pass. Maximum cross-driver difference remains `fgrnhfx = 0.125 W m-2` (maximum percentage error `0.0006368705741347137%`), within unchanged tolerances. - A connected, nonzero precipitation-rate ESMX_Data check passes with one MPI rank, four MPI ranks, and four ranks by four OpenMP threads. All 33 output-file comparisons against the previously validated executable are exact. Thirty accumulated-rain checks have maximum absolute error `1.862645149230957e-09` in model accumulation units. - Formatting and regression both passed in [GitHub workflow 39690399727](https://github.com/NCAR/fire_behavior/actions/runs/39690399727) on `439474f`. Local formatting and whitespace checks pass. No approved reference or regression tolerance changed. PBS job `7900048.desched1` completed the builds and both suites successfully, then exited 1 because the scratch rate-check script looked for its completion marker in stdout instead of the ESMF log. After correcting only that checker, job `7900164.desched1` reran the six rate-check model executions and exited 0. Failed-attempt evidence and original scripts remain preserved. Evidence: `/glade/derecho/scratch/frediani/tmp/cfbm-P02-init-review-20261021T225746Z/`, including `attempt-01/results/restart-hl_i2yr_/summary.json`, `attempt-01/results/pr-jpmyi5r2/summary.json`, and `attempt-01/rate-probe-02/summary.json`. The combined P01/P04, coupled-restart, and direct WRF-host validation gaps above remain. ### Patch-bound restoration The four atmospheric defaults retain their original `ifps:ifpe, jfps:jfpe` patch sections. The temporary explicit memory-bound spelling in `f6a6206` is superseded by `a87b23d`. The final assignments match the original expressions, and `state/state_mod.F90` matches develop exactly. Defaults remain in `Set_vars_to_default`; the NUOPC cold-start rainfall initialization in `Realize` is unchanged. Whitespace checks pass. The detailed HPC results above are from `439474f`; the optional restart suite has not been rerun after restoring patch bounds. GitHub CI on `a87b23d` is pending.

The unrestricted Rothermel moisture cubic produced nonzero reaction at extinction and negative reaction above it, affecting `iboros`, base spread and flame diagnostics before the later spread clamp. This change sets moisture damping to exactly zero at or above extinction, preserves the polynomial below extinction, and leaves physical moisture uncapped. An explicit moisture threshold is retained because a lower clamp on the polynomial alone can leave positive roundoff at exact extinction. The moisture-damping comment identifies `etam`. The independent empirical chaparral spread law remains unchanged. Refs #42; this addresses its extinction scope only. Validation follows the source and lifetime of each quantity: - `Check_nml` requires finite, nonnegative prescribed `fuelmc_g`. Zero is valid; units are water mass per dry fuel mass. - `fuel_t%Check_fuels`, called once during Anderson initialization, requires finite, positive `fuelmce` for burnable categories. - Moisture initialization validates the fuel loads and fixed averaging weights. The component then checks each evolving class after its update, before averaging. Valid classes with valid weights do not require a repeated check of the computed mean. Empty categories retain zero weights. - The restart reader checks effective and class moisture separately through the FMC component. `Check_fmc` is a common type-bound procedure of `fmc_t`, inherited by the moisture implementation. A source comment explains why `Check_fmc` remains: `fmc_g` is read independently from the checkpoint and used directly, and prescribed-moisture restarts have no class state. This is an input check, including for supported legacy files; an unchanged checkpoint from the same validated revision would already satisfy it. Computed-class checks execute after the OpenMP loop on the initial thread, avoiding concurrent fatal diagnostics across tiles. Spatial checks cover owned burnable cells. ROS owns the extinction calculation. No namelist option or checkpoint schema is added; P02 initialization and restart compatibility checks are retained. Full live/dead SB40 combustion and gridded-input readers remain separate work. The review base is `develop` at `622e191`, including merged #39 and #62. Head `af2ca03` contains a 12-file P03 diff. P00/#61 is a test dependency only in separate integration `3b56260`; its commits are absent from this feature branch. Production files, focused Fortran tests and configuration documentation in that integration match the PR head exactly. Validation of `af2ca03` / integration `3b56260`: - [GitHub workflow 39997883832](https://github.com/NCAR/fire_behavior/actions/runs/39997883832) passes formatting and regression on the published head. - Derecho PBS job `7922406` passes four GNU Strict builds: serial, OpenMP, MPI and hybrid, including ordinary NUOPC/ESMX execution. All 35 focused scenarios pass in every build, covering extinction boundaries and diagnostics, namelist/table inputs, class state, initialization loads, and independently restored effective moisture. The bound checker is also exercised without initialized class arrays, matching prescribed-moisture restarts. - All 378 PR CTests, 36 model runs, 40 execution comparisons and 36 approved-reference comparisons pass. Maximum reference error is zero; maximum execution difference is 0.125 W m-2 in `fgrnhfx`, within the existing tolerance. - All 15 restart cases and 30 loading probes pass. The 315 continuity file comparisons have zero error; 450 checkpoint checks and 12 independent layout comparisons also pass. - An additional 22 checkpoint mutation/control probes pass in serial and hybrid configurations. These include invalid `fmc_g` with unchanged valid classes, invalid classes, and accepted zero/wet states. - A separate Fortran probe with the GNU Strict arithmetic flags confirms that, at a moisture ratio of one, the polynomial and `max(etam, 0.0)` both return `2.384185791015625e-7` in single precision. This small positive residual explains retaining the explicit extinction condition. The below-extinction arithmetic is unchanged. No reference or tolerance was changed. Direct WRF-host execution/restart, NUOPC/ESMX restart, cross-layout checkpoint transfer and combined P01/P04 acceptance remain untested. Legacy readability does not establish numerical equivalence across the extinction correction. Evidence, exact commands, source snapshots and PBS logs are retained under `/glade/derecho/scratch/frediani/tmp/cfbm-P03-type-bound-review-20261023T174535Z`.

Adds continuous-versus-restarted regression coverage to the existing harness and runs it in GitHub CI alongside the existing PR suite. The cases exercise prescribed and evolving moisture, delayed moisture updates, checkpoints around observed-perimeter activation, and preservation of REAL64 fuel state. All 15 restart cases pass on the merged P02 initialization correction, including the five cases that failed on the original restart foundation. ## Scope and review base - Base: `develop` at `622e191fac104e1a73ae99a769e5bd3dfee2946c`, including merged #39 and #62. - Head: `56b832037a8e73d198d395e1c194afbdb46ed205`. - The 23-file diff contains the restart harness, tests/documentation, suite help and CI selection. Production code and P02's focused unit tests are identical to develop. No scientific tolerance or reference changes. - Merge `ce551ee` incorporates develop into the published P00 history. The moisture component and restart-reader conflicts retain develop's merged P02 implementation. This preserves the original P00 commits and the common ancestor used by stacked #65. - The harness includes the initialization metadata checks and 30 restart-loading probes previously validated in the P00/P02 integration. These check supported legacy inference and explicit rejection of ambiguous or damaged state. CI builds serial, OpenMP, MPI and hybrid configurations once, then runs both `pr` and `restart`. Both suites execute even if one fails; either failure fails the regression check. The existing test-evidence artifact retains both result directories. The approved `github-pr` baseline remains unchanged, and optional candidate generation still uses only the `pr` results. The configuration guide includes a restart YAML example and explains suite selection, segment timing and comparison policy. ## Validation Derecho PBS `7915414.desched1` exits zero after fresh serial, OpenMP, MPI and hybrid GNU Strict builds. MPI/hybrid builds include NUOPC and ESMX. Environment: GNU 12.2, Cray MPICH 8.1.25, NetCDF 4.9.2, ESMF 8.5, shared NPL Python. - All 67 restart CTests pass: 15 restart cases, 48 P02 lifecycle checks, and four Python harness registrations. - All 65 continuity model invocations and 30 restart-loading probes complete as expected. The 315 continuity file comparisons plus five accepted-legacy checkpoint comparisons are bitwise equal, with zero maximum error. - All 450 checkpoint validity checks and 12 cross-layout comparisons pass; cross-layout error is zero. - All 238 existing PR CTests pass, including 36 model runs. All 40 execution comparisons and 36 approved-reference comparisons pass. Reference error is zero; maximum cross-driver error is 0.125 W m-2 in `fgrnhfx`, within the unchanged tolerance. - All 45 Python tests pass; YAPF 0.43.0, shell syntax and whitespace checks pass. - Local workflow checks confirm both suites execute and the final status is correct when neither, either, or both suites fail. The documented restart configuration matches the production YAML; its 28 s split retains the next output at 36 s. - [GitHub workflow 39970138677](https://github.com/NCAR/fire_behavior/actions/runs/39970138677) passes on implementation revision `5de15ed`: formatting, all four builds, all 238 PR CTests and all 67 restart CTests. Downloaded reports confirm all 15 restart cases, 320 zero-error file comparisons and 30 loading probes pass in CI. Both suite summaries and eight JUnit reports are retained in the uploaded test-evidence artifact; the 36 approved-reference comparisons also have zero error. The PBS run used clean candidate `9a64864`, whose complete tree equals merge `ce551ee`. Later `c8aeeaa` changes documentation, and `5de15ed` changes CI selection and documentation; production code, model cases, harness and approved baseline remain identical to the PBS-validated revision. PBS evidence is preserved under `/glade/derecho/scratch/frediani/tmp/cfbm-P00-develop-20261022T230324-0600/`; CI-update evidence is under `/glade/derecho/scratch/frediani/tmp/cfbm-P00-restart-ci-20261023T082618-0600/`. Repository validation history retains the original failures and records the successful PBS rerun. Documentation-only follow-up `2c6b5b9` adds the [test-count breakdown](https://github.com/NCAR/fire_behavior/blob/test-restart-continuity/tests/regression/README.md#how-test-counts-are-reported): checkpoint validity, file comparisons, the per-build PR CTest table, and restart invocation counts. Counts and links were checked against the downloaded reports from workflow 39970138677. Its [CI run 39974535865](https://github.com/NCAR/fire_behavior/actions/runs/39974535865) uses the unchanged workflow; no executable, harness, case, or baseline changed. Documentation follow-up `b2da936` replaces development labels in the user-facing regression guides with descriptions of moisture initialization, legacy restart metadata, and coverage limits. `tests/AGENTS.md` requires this distinction and keeps development provenance in validation records. Follow-up `56b8320` consolidates all test guidance there, retains a pointer at `tests/regression/AGENTS.md`, and requires informative unit-test comments describing the behavior, setup and expected result. The historical validation file is unchanged. Guide-label, relative-link, whitespace, and pinned YAPF checks pass. Workflows 39974535865 and 39977209659 on the preceding documentation revisions passed both formatting and regression checks. ## Coverage limits Restart comparisons use the standalone driver and the same revision/execution layout. NUOPC/ESMX integrations pass in the existing PR suite, but coupled restart, native WRF-host restart, cross-layout checkpoint transfer, and combined P01/P04 behavior remain separate validation work. Synthetic legacy checkpoints test schema compatibility, not numerical equivalence across scientific revisions. No production correction for #42 is included in this test PR. --------- Co-authored-by: masih <masih@ucar.edu>

masih-e and others added 2 commits August 26, 2026 13:57

Refactor restart code into state submodule

cf7696c

Move restart-specific state code and helpers into a new state_restart_submod.F90 submodule.

masih-e added 3 commits September 25, 2026 12:26

masih-e added 5 commits September 29, 2026 13:45

Clean up restart helper dependencies

5225310

Move FMC restart writes into FMC model

922aa41

Make restart submodule signatures explicit

3d75991

Merge develop into restart-pr

8aba8fc

Clean up restart initialization interface

e73f23c

pedro-jm previously approved these changes Oct 2, 2026

View reviewed changes

masih-e added 3 commits October 20, 2026 12:16

Merge remote-tracking branch 'origin/develop' into restart-pr

e5b75c5

Preserve perimeter activation state across restart

3291f13

Write and read fire_perimeter_ignited as an integer restart attribute and broadcast it on every rank.

Preserve double precision in fuel restart fields

59b9081

Read and write fuel_frac and fuel_frac_burnt_dt as NetCDF doubles. Halo exchange for the variables as doubles.

masih-e dismissed pedro-jm’s stale review via 59b9081 October 20, 2026 19:32

This was referenced Oct 21, 2026

Add focused restart regression coverage #61

Merged

Initialize fuel moisture independently of update cadence #62

Merged

Enforce nonnegative reaction quantities at moisture extinction #63

Closed

pedro-jm approved these changes Oct 21, 2026

View reviewed changes

mefrediani merged commit ead119e into NCAR:develop Oct 21, 2026
2 checks passed

mefrediani mentioned this pull request Oct 21, 2026

Enforce nonnegative reaction quantities at moisture extinction #64

Merged

Navigation Menu

Repository navigation

Add restart capability to the fire behavior model - #39

Add restart capability to the fire behavior model#39
mefrediani merged 13 commits into
NCAR:developfrom
masih-e:restart-pr

masih-e commented Aug 26, 2026 •

edited

Loading

Uh oh!

masih-e commented Sep 25, 2026

Uh oh!

masih-e commented Sep 25, 2026

Uh oh!

mefrediani commented Oct 16, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

masih-e commented Aug 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description of changes:

Type of change

Tests conducted:

Documentation:

Contributors:

Further information:

Uh oh!

masih-e commented Sep 25, 2026

Uh oh!

masih-e commented Sep 25, 2026

Uh oh!

mefrediani commented Oct 16, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

masih-e commented Aug 26, 2026 •

edited

Loading