Skip to content

Charged coagulation and atmospheric ion validation

The corrected CPU/GPU charged Knudsen normalization passed a fresh twenty-case CUDA qualification on 2026-11-14. The frozen specification records revision 54ad1a812, 8192 independent boxes per case and root seed 320. Its helper hash still matches the maintained source. All sixteen individual conservative 95% probability-difference intervals lie inside the predeclared [-0.01, 0.01] tolerance; the largest absolute endpoint is 0.00714698. All four baseline interval widths satisfy the separate 0.01 sampling-precision target. These are individual comparison coverages, not a simultaneous 95% statement.

Ion source [pairs cm^-3 s^-1] Omitted coefficient assumption Reached Captured Unresolved Baseline interval width
2 0 m³/s 142 8050 0 0.00577974
2 1e-15 m³/s 142 8050 0 0.00577974
10 0 m³/s 154 8038 0 0.00600904
10 1e-15 m³/s 154 8038 0 0.00600904

Every case resolved all original lineages. The largest per-species mass residual over the campaign was 3.109e-15, and every box conserved exact net charge. Each doubled-capacity case preserved the entire recorded time/outcome series exactly, checked independently after execution. Maximum observed occupancy was 37 slots; diagnostic samples do not bound occupancy between observations. The complete specification, twenty comparisons, case summaries and capacity checks are saved in docs/data/gpu/charged-survival-corrected-qualification.json.

The current package regression and environmental qualification are recorded in environmental validation. Concurrent CUDA jobs mean campaign timings are not standalone performance measurements.

The five-case CUDA publication used the separate root seed 321 and 8192 boxes per case. Its final generation is run-20261114T203855083494Z, with revision 799e75d92 and source/helper hashes in the manifest. The previous generation was removed after this successful refresh; only the final output remains.

Ion source [pairs cm^-3 s^-1] Omitted coefficient assumption Reached Captured Success fraction Exact 95% interval
0 Strict lookup, no ions 163 8029 0.01990 [0.01698, 0.02316]
2 0 m³/s 152 8040 0.01855 [0.01574, 0.02172]
2 1e-15 m³/s 152 8040 0.01855 [0.01574, 0.02172]
10 0 m³/s 170 8022 0.02075 [0.01778, 0.02408]
10 1e-15 m³/s 170 8022 0.02075 [0.01778, 0.02408]

All lineages resolved by 3730–3750 s. About 98% were captured before reaching 50 nm; the cumulative-success curves rise only near the growth time of the surviving minority. The q=2 point estimate is below the no-ion control and the q=10 estimate is above it. However, the individual conservative 95% ion-minus-control intervals are [-0.00827, 0.00559] for q=2 and [-0.00626, 0.00797] for q=10 under either omission assumption. Each includes zero, so this realization does not resolve an ion effect. Matching outcomes under both coefficient assumptions do not establish identical probabilities or turn those assumptions into guaranteed survival bounds. The one-percentage-point numerical criterion does not establish smaller effects.

Every box conserved exact net charge; maximum species-mass residuals were about 2.89e-15. Observed peak occupancy was 33, 35 and 37 slots for q=0, 2 and 10, respectively, against capacity 128. The final three-panel figure was visually inspected; source/helper/figure hashes match, with one generation retained. The complete refresh took 2030.59 s while the humidity study also ran on the GPU; this is execution provenance, not a throughput comparison.

The independently extracted download passed all five small CUDA smoke cases with eight boxes and a 100 s horizon; Python source and helper bytes match. That smoke test checks execution and inventory/lineage bookkeeping, not survival precision.

The following results predate the charged Knudsen normalization correction. They are evidence for their recorded revisions, not the current results above.

Development evidence, 2026-11-13, branch codex/gpu-charged-coagulation-lucky-few, based on 19a8fb11d. These results concern the implementation on this branch; no release is claimed. Sections retain chronological checkpoints. The earlier qualification and publication run describe their original revisions of the Lucky Few example.

The five existing charged models can run directly or through prepared Warp execution. A bounded parallel reduction supplies the exact pair maximum to the existing stochastic sampler. Prepared CUDA capture and same-device checkpoint continuation retain the selected model. Existing Brownian defaults remain.

The CPU Dyachkov dimensionless formula now evaluates differences with expm1 and an equivalent rearrangement to avoid cancellation near neutrality. The analytic neutral limit is checked independently; CUDA agreement does not define that limit. CUDA decimal fit coefficients are retained as explicit fp64 constants rather than rounding through float32 literals.

IonPairSource injects explicit positive/negative particles with unit counts, independent persistent source randomness, a fixed-capacity rejection boundary, and cumulative external-mass accounting. IonNeutralCleanup exports only neutral particles made entirely from both reserved ion materials, with persistent species mass and particle ledgers. Neither establishes ion collision physics or the completed Lucky Few survival study.

The LF2013 helper rejects unsupported domains by default. Optional lower/upper values for omitted repulsive fluxes are explicit coefficient assumptions, not bounds on survival probability or an extrapolated fit. Same-sign charged seeds with diameter 3 nm fall below the corresponding fit’s lower radius, approximately 5.308 nm. On 2026-11-14 the user selected both coefficient assumptions for the survival sensitivity demonstration; neither is a fitted best estimate.

Host CUDA: NVIDIA GeForce RTX 5060, 8151 MiB; Warp 1.11.1; CUDA toolkit 12.9, driver 13.4. The restricted sandbox cannot see the driver; CUDA qualification was performed with host access. No CPU fallback was used for CUDA cases.

Commands below used WARP_CACHE_PATH=/tmp/particula-charged-warp where relevant.

Check Result
.venv/bin/pytest particula/gpu/kernels/tests/coagulation_test.py particula/execution/tests/prepared_test.py particula/execution/tests/prepared_restart_capture_test.py -q --no-cov -Werror 942 passed, host CPU/CUDA
.venv/bin/pytest particula/gpu/dynamics/tests/charged_models_test.py particula/execution/tests/prepared_charged_test.py particula/dynamics/ions/tests particula/dynamics/coagulation/tests/charged_dimensionless_kernel_test.py -q --no-cov -Werror 43 passed, host CPU/CUDA
.venv/bin/pytest particula/execution/tests/prepared_ions_test.py particula/gpu/kernels/tests/ion_source_test.py particula/execution/tests/prepared_restart_capture_test.py particula/execution/tests/prepared_charged_test.py -q --no-cov -Werror 62 passed, host CPU/CUDA
.venv/bin/pytest particula -Werror --cov=particula --cov-report=term-missing --cov-fail-under=80 10,047 passed, 14 skipped, 1 xfailed; 92.04% coverage; host CPU/CUDA
.venv/bin/pytest examples_tests scripts/tests -q --no-cov -Werror 459 passed, 1 skipped; host execution
.venv/bin/python scripts/docs_examples.py refresh gpu-charged-models Complete CUDA run; saved figure visually inspected
.venv/bin/python -u scripts/docs_examples.py check Passed; all saved bundles current and CPU-policy examples executed; CUDA/manual policies require explicit refresh
.venv/bin/ruff check particula/ examples_tests/ scripts/ --no-fix Passed
.venv/bin/ruff format particula/ examples_tests/ scripts/ --check Passed
.venv/bin/mypy particula/ --ignore-missing-imports Passed
.venv/bin/python scripts/check_docs_structure.py 0 errors
npm --prefix website test 3 passed
npm --prefix website run build Passed; 0 link errors

The full package run at 80e271627 includes paired-source integration, composition and zero-production tests. Subsequent cleanup changes are covered by the focused checks below; that earlier full run does not qualify them. The existing functional and object charged examples were also refreshed; all four saved figures were visually inspected. Notebook execution required host access; restricted notebook processes were stopped after stalling during kernel startup. A restricted tooling run also hit the read-only Warp cache; the successful host rerun above supersedes it.

Hosted CI is separate from these local results. All six workflows passed at 8d8f476fd; newer increments require their own hosted results. No CPU speedup or field-validated atmospheric survival prediction is established by this record.

Cleanup tests check that neutral aerosols remain, charged ion-only particles remain, each exported species is accounted for, repeated cleanup is a no-op, and ledger/state continuation survives checkpoint and native CUDA capture. Invalid reserved indices reject at preparation; evolving fractional counts fault before modifying the affected box. The source and cleanup retain distinct families so the example must explicitly choose their ordering.

Host command: WARP_CACHE_PATH=/tmp/particula-charged-warp .venv/bin/pytest particula/execution/tests/prepared_ion_cleanup_test.py particula/execution/tests/prepared_ions_test.py particula/execution/tests/prepared_restart_capture_test.py -q --no-cov -Werror: 58 passed, including actual CUDA. Ruff and mypy passed (571 source files).

SeedCohort stores unresolved original-lineage counts and persistent background labels separately from particle inventories. It consumes only applied compacted collision pairs and observes size after each physical process. Outcome counts partition the original cohort into reached, captured and unresolved lineages.

The affected prepared suite passed 171 tests on host CPU/CUDA:

Terminal window
WARP_CACHE_PATH=/tmp/particula-charged-warp .venv/bin/pytest \
particula/execution/tests/prepared_cohort_test.py \
particula/execution/tests/prepared_test.py \
particula/execution/tests/prepared_charged_test.py \
particula/execution/tests/prepared_ions_test.py \
particula/execution/tests/prepared_ion_cleanup_test.py \
particula/execution/tests/prepared_restart_capture_test.py \
-q --no-cov -Werror

Tests cover two seed lineages in one merged particle, merger-induced threshold crossing, capture precedence, irrevocable previous success, unresolved censoring, real condensation growth, actual CUDA capture, restart before/after merging, corrupt checkpoint outcome rejection and conflicting ion material maps. Ruff, formatting and mypy passed (575 source files). These tests establish bookkeeping and integration; survival-probability convergence is still pending.

Ion attachment numerics and condensation exclusion

Section titled “Ion attachment numerics and condensation exclusion”

The LF2013 device helper evaluates the published degree-23 polynomials in a centered logarithmic-radius basis. Coefficients are transformed once using 70-digit decimal arithmetic and uploaded as fp64 values. This is an algebraic basis change, not a new fit or interpolation. Tests independently evaluate the original decimal coefficients at nine radii in each valid interval, covering all 201 aerosol charges and both ion polarities. CPU and CUDA agree with those references within 1e-10 relative tolerance. Default unsupported rates return a fault sentinel; optional lower/upper bounds remain explicit. Device helpers classify free ions, inert neutral products, aerosols and unsupported ion-only clusters. They are not yet connected to the atmospheric collision scheduler.

IonCondensationExclusion gates ion-only proposals before gas-inventory limiting in every condensation substep. Tests verify exact ion/product masses, continued uptake by an aerosol carrying ion mass, particle/gas conservation, isothermal and latent-heat capture/restart, and configuration rejection. Mixed physical material densities use the CPU surface model’s constant Kelvin density rather than replacing physical densities; a native CPU comparison passes at 2e-12 relative tolerance. The former mixed-material rejection test now exercises the still-unsupported nonuniform surface-density case, while the new parity test covers the newly supported material-density case.

Host command (171 passed, actual CPU/CUDA):

Terminal window
WARP_CACHE_PATH=/tmp/particula-charged-warp .venv/bin/pytest \
particula/execution/tests/prepared_ion_condensation_test.py \
particula/execution/tests/prepared_test.py \
particula/execution/tests/prepared_cohort_test.py \
particula/execution/tests/prepared_restart_capture_test.py \
particula/gpu/dynamics/tests/ion_attachment_test.py \
particula/dynamics/ions/tests/attachment_test.py \
-q --no-cov -Werror

The direct condensation regression, parallel reductions and autodiff were also checked after the shared proposal signature changed:

Terminal window
WARP_CACHE_PATH=/tmp/particula-charged-warp .venv/bin/pytest \
particula/gpu/kernels/tests/condensation_test.py \
particula/gpu/kernels/tests/condensation_autodiff_test.py \
particula/gpu/kernels/tests/condensation_parallel_test.py \
-q --no-cov -Werror

Result: 324 passed, 3 skipped. Ruff, formatting and mypy (581 source files) passed; the website built with zero link errors.

Atmospheric collision integration and ion balance

Section titled “Atmospheric collision integration and ion balance”

The follow-up to 8d8f476fd routes ion pairs through the existing sampler and merge application. Both the exact pair maximum and acceptance probability use the same derived ion classification and cached LF2013 rates. Validation faults before collision RNG or physical merge mutation. Neutral products are inert; same-sign ions do not cluster. Source, cleanup, condensation exclusion and cohort material contracts are checked together. Derived scratch is recomputed at each step and does not introduce a new checkpoint payload.

Host CPU/CUDA evidence, 2026-11-14 UTC:

Command Result
.venv/bin/pytest particula/execution/tests particula/gpu/kernels/tests/brownian_parallel_test.py particula/gpu/dynamics/tests/ion_attachment_test.py -q --no-cov 1,721 passed before the final three composition/marker tests
.venv/bin/pytest --cov=particula --cov-report=term-missing --cov-fail-under=80 -Werror 10,110 passed, 14 skipped, 1 xfailed; 91.85% coverage; includes all 30 atmospheric tests and actual CUDA
.venv/bin/ruff check particula examples_tests scripts --no-fix Passed
.venv/bin/ruff format particula examples_tests scripts --check 643 files formatted
.venv/bin/mypy particula --ignore-missing-imports Passed, 586 files
.venv/bin/python scripts/docs_examples.py refresh gpu-ion-balance Complete RTX 5060 run; figure visually inspected
.venv/bin/python scripts/check_docs_structure.py Passed
npm --prefix website test 3 passed
npm --prefix website run build 453 pages; passed

The ion-balance example simulates 64 independent boxes per q=0/2/10 condition, 256 slots per box, timestep 1 s and duration 2000 s. A separate sparse master equation solution supplies the exact finite-population mean and variance; its truncated upper-state probability is checked below 1e-10. Maximum observed mean discrepancies were 1.58, 2.67 and 2.54 standard errors, below the fixed six standard-error acceptance criterion. Net charge was exactly zero. Per-species mass budgets included source input and neutral-product export, with maximum relative residual below 1e-14. Observed peak occupied slots were 40, 70 and 128. The current Python download matches the maintained source and compiles.

This qualifies the source/recombination example, not a survival conclusion. The 3–50 nm study and its timestep/cleanup/capacity/ensemble convergence remain outstanding. The production treatment of omitted LF2013 repulsive coefficients is still awaiting the user’s scientific choice; strict rejection is the default.

Periodic cleanup and documentation reachability follow-up

Section titled “Periodic cleanup and documentation reachability follow-up”

Hosted documentation CI at 0288c548b detected that the new ion-balance page was not linked from authored navigation. The earlier local check ran before that file was tracked and did not inventory it. Commit fc18f012d adds the owning GPU index and guide links; the committed-file structure check reports zero errors and its 13 tests pass. This corrects the earlier local-only structure result rather than treating the failed hosted build as successful.

IonNeutralCleanup.interval_steps makes refresh frequency independent of the physical timestep. A bounded per-box phase counter is checkpointed; cleanup runs on multiples of the configured interval. Inert neutral products continue to occupy slots between exports. Unit-count checks still run every call and invalid checkpoint phases or changed intervals reject. CPU/CUDA cleanup, atmospheric, cohort and restart regression passed 104 tests, including midperiod capture/restart and default every-step behavior. Ruff and mypy passed. The earlier full-package coverage result predates this periodic extension.

The subsequent combined study-helper/ion-exclusion/cleanup/atmospheric/cohort/ restart suite passed 125 tests on host CPU/CUDA. This caught and corrected a shared-plan assumption: condensation exclusion must not inherit the cleanup interval parameter. The complete ion-balance example was refreshed again and its default every-step results retained.

A preliminary neutral survival pilot used the explicit defaults in scripts/gpu_studies/lucky_few.py: 64 independent boxes, one 3 nm seed each, 32 initial 100 nm background particles per box, pulsed vapor replenishment, and 20,000 s duration. It recorded 2 reached, 62 captured and 0 unresolved lineages, with maximum mass residual 4.67e-15. The exact 95% success interval is broad (0.003807–0.108371); this is not a converged production estimate. The pending timestep/capacity/ensemble work and ion-domain decision remain in the plan. No ion-effect conclusion follows from this q=0 pilot.

Neutral growth and survival uncertainty controls

Section titled “Neutral growth and survival uncertainty controls”

Host CUDA checks at library revision 5aa5b47ff used the frozen physical parameters in the survival helper. With background particles disabled, four seeds per run all reached 50 nm at each of dt = 2, 1 and 0.5 s. First passage was first observed at 1390 s in all three runs, using 10 s diagnostic spacing. This satisfies the predeclared crossing-time spread criterion of the larger of 10 s and 1% of the finest-step crossing time. It does not resolve passage time within a diagnostic interval. Maximum mass residuals were 2.45e-15, 2.23e-15 and 3.45e-15, respectively.

Survival timestep qualification uses an absolute probability tolerance of 0.01, fixed before the 2048-box comparison. Each run has one initial seed per independent box. The exact intervals follow Clopper and Pearson (1934). Two exact 97.5% binomial intervals give at least 95% joint coverage by the union bound; subtracting their opposite endpoints gives a conservative interval for the probability difference. This does not assume independence between runs sharing random streams. The entire difference interval must lie inside [-0.01, 0.01] to establish the declared tolerance. An interval that crosses a tolerance boundary is inconclusive, even when point estimates agree. Any larger confirmatory ensemble must be declared before its outcomes are examined; pilot estimates alone are not a stopping rule for nominal confidence coverage.

All six hosted workflows passed on 5aa5b47ff: test, linting, documentation, assistant plugins, PyPI packaging and Conda feedstock checks. These are validation results, not a package release or merge.

The 2048-box pilot gave 39 reached at dt=1 s and 34 at dt=0.5 s, with no unresolved lineages. The conservative difference interval is [-0.016076, 0.011208], so the pilot does not resolve the declared tolerance. A fixed confirmation is declared before execution: 8192 independent boxes per timestep, new root seed 318, dt=1 and 0.5 s, all physical defaults and 20,000 s duration unchanged. Its intervals will be assessed once, separately from the pilot; neither pooling nor repeated nominal-95% stopping is used.

Neutral pilot Reached / 2048 Maximum mass residual Synchronized stepping
dt=1 s, capacity 128 39 6.22e-15 205.7 s
dt=0.5 s, capacity 128 34 5.56e-15 407.8 s
dt=1 s, capacity 256 39 6.22e-15 304.0 s

The capacity comparison preserves the complete cohort outcome time series exactly; maximum observed occupancy was 33 in all runs. These single-run stepping times exclude preparation, explicit gas forcing and analysis. They are workload observations on RTX 5060, not a matched CPU speedup or a repeated performance benchmark. Ion-production capacity and cleanup sensitivity are still outstanding.

The updated interval helper passed all 21 study-helper tests on host CPU/CUDA (.venv/bin/pytest scripts/tests/lucky_few_study_test.py -q -Werror --no-cov). Tests cover exact binomial endpoints, sign reversal of difference intervals, invalid confidence levels and controlled condensation with inventory closure. Ruff and formatting passed; documentation structure had zero errors and the website built 453 pages with no link errors. The full-package result above still predates periodic cleanup; a final complete run remains required.

Full-package follow-through after periodic cleanup

Section titled “Full-package follow-through after periodic cleanup”

At library revision d14f0899c (with documentation and study-helper review changes in the working tree), host CPU/CUDA validation completed:

Command Result
WARP_CACHE_PATH=/tmp/particula-charged-warp .venv/bin/pytest --cov=particula --cov-report=term-missing --cov-fail-under=80 -Werror 10,119 passed, 14 skipped, 1 xfailed; 91.85% coverage
.venv/bin/ruff check particula examples_tests scripts --no-fix Passed
.venv/bin/ruff format particula examples_tests scripts --check 646 files formatted
.venv/bin/mypy particula --ignore-missing-imports Passed, 586 files

This full-package result includes the periodic-cleanup implementation and supersedes the earlier package run that preceded it. The CUDA device was available on the host; it was not replaced with a CPU fallback. The test run overlapped the 8192-box half-step confirmation from 06:53 UTC on 2026-11-14. That confirmation’s stepping duration is therefore excluded from performance comparisons. Numerical outcomes and conservation assertions remain separate from timing claims. All six hosted workflows passed on d14f0899c; later uncommitted example/helper changes still require their own follow-through.

The ion-balance reference now rejects nonuniform/nonzero-origin time grids instead of silently evaluating different times, and rejects invalid chain parameters. Ten independent tests include analytic two-pair decay: rates 4 alpha/V then alpha/V yield P2=exp(-4 alpha t/V) and P1=(4/3)[exp(-alpha t/V)-P2], from which both moments are checked.

The complete example/tooling command, WARP_CACHE_PATH=/tmp/particula-charged-warp .venv/bin/pytest examples_tests scripts/tests -q -Werror --no-cov, passed 493 tests with one skip on the host. This includes controlled growth under both aerosol process orders on Warp CPU and actual CUDA. The full ion-balance CUDA refresh is run-20261114T070049797164Z; its inspected figure is byte-identical to the prior numerical figure. The executed source and built Python download both match the maintained source and compile. The changed example check passed, website tests passed (3), and the 453-page build reported zero link errors. CUDA-policy skips in the routine example checker are distinct from the deliberately refreshed CUDA workload.

The selected environment retains installed Particula distribution metadata 0.2.14 while the executed checkout reports source version 0.3.0. The manifest records installed distribution versions separately from the actual source revision and package-source hash; the environment was preserved.

Composed-system CUDA checks used 64 boxes, root seed 319, dt=1 s, cleanup every 10 s and all frozen physical defaults for 20,000 s:

q [pairs cm^-3 s^-1] Omission assumption Capacity Reached / captured / unresolved Maximum mass residual Observed peak slots
2 lower 128 1 / 63 / 0 4.00e-15 34
2 upper 128 1 / 63 / 0 4.00e-15 34
10 lower 128 1 / 63 / 0 3.78e-15 35
10 upper 128 1 / 63 / 0 3.78e-15 35
10 lower 256 1 / 63 / 0 3.78e-15 35
10 upper 256 1 / 63 / 0 3.78e-15 35

Both q=10 capacity comparisons preserved the complete cohort time series exactly. These are composed-path and capacity checks under each implemented assumption, not production model selection. The ensemble is too small to establish an ion effect or insensitivity of rare survival. Coefficient endpoints do not necessarily bound survival outcomes in this nonlinear coupled model. These runs and the example refresh overlapped the neutral confirmation; their timings are not performance qualification.

At this stage the survival manuscript’s plots had been inspected using the actual neutral pilot. The manuscript has since moved to docs/examples/gpu/prepared/gpu_lucky_few.py for the accepted two-assumption sensitivity demonstration. Charged qualification and publication evidence are recorded below as they become available.

The predeclared confirmation used 8192 independent boxes, root seed 318 and 20,000 s at the frozen physical defaults. Both cases used the helper loaded at d14f0899c, before the order-control option was added. Their specification recorded helper SHA256 14f936ca87a115653c73b3dc8d36af14793473b264cd7a142ee50eb77918361e.

Timestep Reached Captured Unresolved Exact 95% success interval Maximum mass residual
1 s 153 8039 0 [0.0158563, 0.0218466] 5.78e-15
0.5 s 159 8033 0 [0.0165326, 0.0226344] 5.33e-15

The conservative 95% difference interval (half-step minus default) is [-0.00616579, 0.00762977]. It lies inside the predeclared absolute tolerance [-0.01, 0.01], so the neutral timestep check passes at that resolution. This is not proof that numerical differences smaller than one percentage point are negligible, nor does it qualify charged ion-production conditions. Pilot and confirmation samples were not pooled, and the fresh-seed confirmation was assessed once. The half-step timing overlapped other CUDA validation and is excluded from performance comparisons.

Neutral process-order confirmation and reference review

Section titled “Neutral process-order confirmation and reference review”

The separately predeclared order check used ee8384eec, 8192 boxes, root seed 318, dt=1 s and 20,000 s. The script asserted that every configuration field except aerosol process order matched the saved confirmation baseline. With coagulation before condensation, it recorded 152 reached, 8040 captured and zero unresolved lineages. Its exact 95% success interval was [0.0157437, 0.0217152], and maximum mass residual was 6.67e-15.

The conservative 95% difference interval versus the default order is [-0.00694530, 0.00670130], inside the fixed [-0.01, 0.01] tolerance. This qualifies the neutral ordering comparison at that absolute resolution. Timestep and ordering intervals are separate, not simultaneous 95% bounds across every numerical comparison. Charged-production and cleanup-frequency qualification remain outstanding. Overlapping runs exclude timing claims.

Source review found and corrected one remaining incorrect D’yachkov citation in charged_kernel_wp. The publisher record identifies D’yachkov, the two Khrapaks and Morfill in Physics of Plasmas 14, 042102 (2007). The common pair scaling and hard-sphere fit are separately attributed to Gopalakrishnan and Hogan (2012), with Gatti/Kortshagen and Chahl/Gopalakrishnan references included for the other selected fits. After removing docstrings, the changed GPU module has an identical executable AST to ee8384eec.

The manuscript now states the implemented interaction boundary explicitly: aerosol–aerosol electrostatics enter through the product of the charges, with no additional image/polarization term. At fixed masses and radii, either zero aerosol charge removes that Coulomb correction. LF2013 ion attachment remains a separate collision law. These model boundaries limit interpretation of any future atmospheric-survival comparison.

Accepted two-assumption qualification campaign

Section titled “Accepted two-assumption qualification campaign”

The accepted study compares the documented omissions replaced by 0 and 1e-15 m^3/s. Settings were frozen before execution: 8192 boxes, root seed 320, 20,000 s horizon, dt=1 s, capacity 128, cleanup every 10 s, q=2/10 pairs cm^-3 s^-1. For each assumption and q, the baseline is compared with dt=0.5 s, reversed aerosol order, cleanup every 5 s and capacity 256. Each conservative 95% probability-difference interval must lie within [-0.01, 0.01]; each baseline exact 95% interval must have width <=0.01. These cover individual comparisons, not a simultaneous confidence guarantee across the matrix.

Run the fixed matrix with:

Terminal window
WARP_CACHE_PATH=/tmp/particula-charged-warp .venv/bin/python \
-m scripts.gpu_studies.qualify_lucky_few --output /tmp/lucky-few-qualification

The output directory must be new. The specification records the source revision and helper hash before any run; reduced results and checks are saved after each case. The script retains failed or inconclusive comparisons and exits nonzero if any acceptance criterion is unresolved.

stop_when_resolved=True ends a case only when every original lineage has an absorbing reached/captured outcome. It preserves survival at the configured horizon but does not predict later aerosol evolution. Stopped runs matched the full-run prefix and final outcomes on CPU/CUDA; a censored case still ran to its horizon. The focused helper suite passed 25 tests. The complete example/tooling suite passed 494 tests with one skip using host CUDA and the writable Warp cache. A preceding sandbox run failed only on its read-only default Warp cache.

The registered source is linked from the GPU index, prepared guide and website navigation. A preliminary build produced 454 pages and 584 HTML documents with zero link errors. The extracted helper ZIP imported its own local helper and executed all five small CUDA cases outside the checkout with PYTHONPATH removed. Its first smoke assertion used exact decimal-time equality; rerunning with the established 1e-14 time tolerance passed. This download smoke check is separate from the complete publication workload and its numerical qualification.

All 20 CUDA runs completed and all predeclared checks passed. The four baseline precision checks and 16 numerical comparisons use individual confidence intervals; they do not form a simultaneous 95% guarantee over the matrix. The accepted resolution is one percentage point in absolute survival probability. Smaller ion effects are not qualified by this criterion.

The campaign used library revision 58ab7e3e870b08932149baca4ae5c62a5ba99664 and the study helper now committed in f861f9fd4, SHA256 5f61eb0415b560e577b8f6d841053c3cf8eda7e9a4a3acb184812663d9ec8ca1. All cases resolved every lineage before the 20,000 s horizon. Their last diagnostics ranged from 3710 to 3790 s. Maximum per-species mass residual was 2.998e-15, exact net charge remained zero, and peak observed occupancy was 37 slots. Diagnostic occupancy is not an instantaneous upper bound.

q [pairs cm^-3 s^-1] Omitted coefficient [m^3/s] Baseline reached / 8192 Exact 95% success interval
2 0 140 [0.014395, 0.020136]
2 1e-15 140 [0.014395, 0.020136]
10 0 151 [0.015631, 0.021584]
10 1e-15 151 [0.015631, 0.021584]

Differences below are variant minus its matching baseline.

q Omitted coefficient [m^3/s] Numerical change Reached / 8192 Conservative 95% difference interval
2 0 half-step 135 [-0.007102, 0.005882]
2 0 reversed-order 140 [-0.006549, 0.006549]
2 0 half-cleanup 140 [-0.006549, 0.006549]
2 0 double-capacity 140 [-0.006549, 0.006549]
2 1e-15 half-step 135 [-0.007102, 0.005882]
2 1e-15 reversed-order 140 [-0.006549, 0.006549]
2 1e-15 half-cleanup 140 [-0.006549, 0.006549]
2 1e-15 double-capacity 140 [-0.006549, 0.006549]
10 0 half-step 153 [-0.006569, 0.007057]
10 0 reversed-order 152 [-0.006680, 0.006924]
10 0 half-cleanup 150 [-0.006902, 0.006658]
10 0 double-capacity 151 [-0.006791, 0.006791]
10 1e-15 half-step 153 [-0.006569, 0.007057]
10 1e-15 reversed-order 152 [-0.006680, 0.006924]
10 1e-15 half-cleanup 150 [-0.006902, 0.006658]
10 1e-15 double-capacity 151 [-0.006791, 0.006791]

All four doubled-capacity runs preserved the complete recorded outcome time series exactly. Nine earlier full-horizon validation histories also retained identical outcome counts after every lineage had resolved, supporting the absorbing-outcome stopping rule alongside the CPU/CUDA regressions.

Short CUDA regression and download-smoke jobs overlapped early campaign runs; these timings are not performance comparisons. The publication workload uses a separate fixed root seed, 321, with the qualified default settings. Both coefficient assumptions remain explicit sensitivity cases, not bounds on survival probability or an empirical atmospheric validation.

The complete five-case CUDA refresh is run-20261114T140715836273Z, executed from clean source revision 4a550421e7c0f01f4f0e6d2e9a9a75e8648c4b9f on the RTX 5060. It used 8192 boxes per case, root seed 321, and the qualified defaults. The saved source matches the maintained percent script byte for byte; both helper hashes match the manifest. The 300-DPI three-panel figure was visually reviewed with the actual five publication cases.

q [pairs cm^-3 s^-1] Omitted coefficient [m^3/s] Reached Captured Unresolved Exact 95% success interval
0 strict; no ions 163 8029 0 [0.016984, 0.023159]
2 0 154 8038 0 [0.015969, 0.021978]
2 1e-15 154 8038 0 [0.015969, 0.021978]
10 0 163 8029 0 [0.016984, 0.023159]
10 1e-15 163 8029 0 [0.016984, 0.023159]

The conservative 95% ion-minus-control intervals include zero for every comparison. This run does not resolve an ion-production effect. Matching coefficient-case outcomes do not establish identical survival probabilities or mathematical bounds. These are illustrative finite-box results, not field-validated atmospheric predictions.

All reported maximum mass residuals were 2.89e-15, exact charge checks passed, and observed peak occupancy was at most 37/128 slots. Final diagnostics were at 3710–3740 s, with no unresolved lineages. The complete refresh took 975.98 s (about 16.3 minutes). Printed stepping times exclude diagnostic transfers, plotting and analysis; preparation timing begins after CPU initial-state construction. These are execution diagnostics, not a matched CPU/GPU speedup measurement.

Final download checks compared the built Python and ZIP sources with the maintained source and both declared helpers byte for byte. The extracted ZIP passed all five small CUDA smoke cases outside the checkout with PYTHONPATH removed. The changed-example check passed; its CUDA policy skips are separate from the deliberate full CUDA refresh above. Website tests passed (3), and the final build produced 454 pages / 584 HTML documents / 270 API modules with zero link errors. Final hosted status is recorded on PR #1650, separately from local CUDA evidence.

Explore this topic: Scientific and numerical correctness