Skip to content

Environmental dynamics: dry aerosols and humid air with ions

The dry study and humidity study are separate physical experiments. Each published example uses one stable website/results/EXAMPLE_KEY/ folder; superseded output remains in Git history. Failed refreshes or replots preserve the previous successful bundle.

The dry study starts with 2,490,368 unit-count aerosols: 19 conditions × 1024 independent boxes × 128 particles. Sixteen geometric altitudes span 0–86 km; three additional chambers use 2, 5 and 10 atm at 288.15 K. The ground-level case supplies the 1 atm chamber control. A hydrostatic US-standard-atmosphere approximation supplies altitude T/P, including molecular-scale temperature at 86 km. Number density stays fixed at 1e11 m^-3 to isolate transport effects; this is not an atmospheric abundance profile. The run advances 1000 s in 0.5 s steps, with no water or ions. Bars are ordered by altitude and pressure.

The humidity study starts with 2,621,440 resolved particles in 16,384 boxes: 128 aerosols, 16 positive ions and 16 negative ions per box. Each box reserves 32 additional slots. T is 278.15/298.15 K, P is 70000/101325 Pa, and initial RH is 40/60/80/90%. Aerosol number density is 5e13 m^-3; each ion sign starts at 6.25e12 m^-3. A paired source supplies 1e14 pairs m^-3 s^-1. This is an illustrative dense laboratory-style population, not ambient ion abundance. The run advances 3 ms in 1 microsecond steps. Water vapor is finite and unreplenished; prescribed T/P do not respond to latent heat. Vapor diffusivity is explicitly held at 2e-5 m²/s because the prepared condensation interface accepts species constants rather than per-box transport laws.

Both studies use matched lognormal dry radius samples (35 nm geometric mean, 1.6 geometric standard deviation, clipped 15–150 nm), density 1400 kg/m³ and initial aerosol charges -1,0,0,+1. The humidity model adds κ=0.3 water activity, Kelvin curvature and Buck vapor pressure. Condensation, charged aerosol coagulation and ion attachment are active together.

The opt-in mobility attachment law uses ion friction e/Z and Z=Z_ref sqrt(T/298.15) 101325/P, assuming constant collision cross section and fixed ion composition. Reference mobilities are illustrative 1.4/1.9e-4 m²/(V s), masses 150/90 u and contact radius 0.25 nm. It omits image forces, three-body trapping, hydration and composition changes. Opposite-ion recombination remains an explicitly fixed 1.7e-12 m³/s. LF2013 defaults and its reference-condition restriction remain separate and unchanged.

Physical-limit tests exposed an inherited inversion of the Coulomb enhancement ratio in the shared CPU/GPU diffusive Knudsen number. The corrected definition is sqrt(kBTm)/(fa) * Gc/Gk. With the dimensional fit, it approaches 4πa(kBT/f)Gc in the continuum and πa² sqrt(8kBT/(πm))Gk in the ballistic limit. The old CPU/GPU parity checks could not detect a shared formula error. Independent charged-limit tests now cover attraction, neutrality and repulsion; tiny-Knudsen comparisons use zero absolute tolerance.

The changed rates also exposed a CPU sampler defect: a mixed radius bin could use an ion as both donor and receiver and join two repelling aerosols through a spurious chain. Actual pair radii now determine donor orientation before uniqueness, and self-pairs are excluded. The existing no-spurious-calcite regression is retained. A prior integration expectation of such chains was corrected to require zero calcite-calcite events in that fixed-seed case.

Hardware: NVIDIA GeForce RTX 5060, 8 GiB; Warp 1.11.1 and Python 3.12.12. CUDA runs use the existing environment and WARP_CACHE_PATH=/tmp/particula-charged-warp. Restricted sandbox CUDA skips are not counted as hardware evidence.

Dry refinement uses 64 independent boxes per condition, seed 513, 1000 s, and dt=0.5/0.25 s. Every individual paired-box approximate 95% interval for the number-loss difference must lie within ±2 percentage points of initial number. All 19 conditions passed after the normalization correction. Intervals are individual, not a simultaneous 95% guarantee. The largest absolute interval endpoint was 1.4491 percentage points. docs/data/gpu/environment-dry-refinement.json retains all nineteen comparisons.

The initial humidity pilot used 16 boxes per condition and seed 412. Halving dt from 1 to 0.5 microseconds over 3 ms changed water inventories by at most 0.27014% and final vapor by 0.000225%. These passed 1%/0.1% criteria. Maximum positive/negative attachment differences were 5.8594%/7.8125% of the initial per-sign ion count; recombination differed by 2.3438%. The predeclared 2% population criterion therefore failed. This small stochastic pilot is inconclusive, not a passing timestep validation.

A separately frozen confirmation uses 256 boxes per condition and seed 413, with the same timestep, horizon and tolerances. It passed: water differed by at most 0.26361%, vapor by 0.0003033%, and aerosol number by 0.01526% of initial aerosol count. Positive attachment, negative attachment and recombination differed by 0.95215%, 1.90430% and 0.80566% of initial per-sign ion count, respectively. The negative-attachment comparison is close to the 2% criterion; passing it does not establish precision substantially below that threshold. The records are retained in docs/data/gpu/environment-humidity-pilot.json and docs/data/gpu/environment-humidity-refinement.json. The population comparison includes Monte Carlo variability; it is a sensitivity check, not a confidence interval or proof of asymptotic convergence. Reproduce the fixed checks with:

Terminal window
.venv/bin/python -m scripts.gpu_studies.qualify_environment dry /tmp/dry.json
.venv/bin/python -m scripts.gpu_studies.qualify_environment humidity /tmp/humid.json

The full CUDA run completed with seed 411 and 3 ms of physical evolution. The original publication run-20261114T205354713341Z recorded revision 5011becc0 and source/input hashes. All eight figures were inspected. Starting with 2,097,152 aerosols plus 524,288 ions, it recorded 334 aerosol mergers, leaving 2,096,818 aerosols (0.01593% number loss). The maximum per-box species-mass residual was 8.593e-14 and total charge was conserved exactly at every observation.

There were 46,354 positive and 58,889 negative attachments, 6,861 recombination events and 12,625 injected ion pairs. Attachment therefore changed aerosol charge much more often than aerosol mergers changed particle number. The attachment balance transfers net -12,535 elementary charges to aerosols, compensated by the remaining free ions. Faster negative-ion depletion is consistent with the prescribed larger negative mobility; this does not isolate mobility from the other coupled model processes.

The pooled median diameter rose from 74.453 to 85.209 nm. Water uptake drives this bulk size change: only 0.01593% of aerosols were lost to mergers. Final median diameters at initial RH 90% were about 101.84 nm at 278.15 K and 103.19 nm at 298.15 K, compared with about 74.26 nm at RH 40%. The finite vapor reservoir leaves final RH near 89.28% and 89.77%, respectively, in the 90% cases. Size distributions approach a plateau while attachment and aerosol mergers continue. Charge-change panels reveal redistribution that is difficult to see in the nearly flat pooled neutral/singly charged fractions.

Each condition recorded only 12–31 aerosol mergers. The bars follow physical T/P/RH order; their noisy ranking does not establish significant environmental differences or an ion-induced coagulation enhancement. That causal comparison would require matched no-ion controls and uncertainty estimates.

The refresh took 7450.46 s, including 7431.77 s synchronized stepping and 2.05 s preparation. It overlapped other CUDA work, so this is provenance, not a standalone performance benchmark. Saved source, declared helper inputs and figure hashes match; only the final successful generation is retained. The extracted download independently passed a real-CUDA smoke run with 10,240 particles over 100 microseconds and all eight plotting paths. That smoke exercised both attachment polarities, recombination and the ion source, with exact charge conservation and a 2.442e-15 maximum mass residual.

The full dry CUDA run completed with seed 512, 2,490,368 initial aerosols and 1000 s of physical evolution. Its original publication was run-20261114T180838997295Z, with revision 8aa63a3fb and executable/input hashes in the manifest. All six figures were inspected. It recorded 449,485 mergers, leaving 2,040,883 aerosols: an 18.04894% total number loss. The largest per-box material residual was 1.554e-15; every box conserved net charge exactly. The pooled number-median diameter increased from 70.121 to 75.853 nm.

The altitude response was nonmonotonic. Number loss rose from 9.875% at ground level to 24.392% at 50 km, then fell to 21.464% at 86 km. This is consistent with the model approaching its low-pressure collision limit while the reference temperature decreases above the stratopause. It is not evidence that ambient mesospheric aerosol concentrations or lifetimes follow these values: initial number density was deliberately held constant and omitted processes remain absent. The separate isothermal chamber series decreased monotonically from 9.875% loss at 1 atm to 6.647%, 4.482% and 3.853% at 2, 5 and 10 atm.

The time histories add information beyond those endpoints. The pooled collision rate declines as the particle population is depleted, while the size quantiles increase. Absolute neutral and singly charged counts decline; small populations with charge magnitude at least two emerge. A nearly unchanged neutral fraction therefore does not imply an unchanged neutral population. These are descriptive fixed-seed observations, not significance tests or a neutral-control comparison.

The complete refresh took 5357.70 s, including 5345.54 s of synchronized stepping and 0.92 s preparation. Other CUDA studies ran concurrently, so these times are execution provenance rather than standalone throughput measurements. Maintained, saved and downloadable Python sources match; ZIP helper bytes match the declared inputs. The extracted ZIP passed a separate CUDA smoke run with 4864 aerosols over 5 s, exercising all six plotting paths outside the checkout.

The implementation at 54ad1a812 passed the host CPU/CUDA package suite: 10,093 passed, 13 skipped, 57 deselected and one expected failure, with warnings treated as errors and 91.60% coverage. The skipped and deselected cases are not claimed as executed CUDA evidence. Independent charged continuum and ballistic limits supplement CPU/GPU parity; LF2013 default behavior is retained.

The examples/tooling run passed 516 tests with one skip. Its isolated wheel installation test initially failed because the validation wrapper exported the checkout through PYTHONPATH. Rerunning the seven bundle tests without that override passed, including the wheel and offline source-distribution checks. Ruff lint/format, mypy (589 files) and documentation structure checks passed. The final website build rendered 459 pages and verified 589 HTML pages, 272 API modules, 1860 legacy anchors and 130 redirects with zero errors. All three study downloads match their maintained Python sources and declared helpers; separately extracted bundles passed the CUDA smoke checks described here and in the charged-coagulation validation record.

The publication pipeline now stores one current bundle per example directly in website/results/gpu-dry-environment-sweep/ and website/results/gpu-environment-sweep/. All 79 registered example bundles were migrated without changing their selected manifest bytes; Git retains historical versions. Only the two environmental examples currently opt into numeric-data capture and replotting.

Their data.json contains compact group summaries, observation times, size histograms, charge counts and quantiles, not individual particle states. The pipeline caps serialized JSON at 10 MiB per example. Oversized or nonfinite data fails publication without replacing the previous successful bundle. This cap applies to saved plot data, not simulation RAM or PNG files.

For plotting-only changes, run:

Terminal window
CUDA_VISIBLE_DEVICES='' .venv/bin/python scripts/docs_examples.py replot gpu-dry-environment-sweep
CUDA_VISIBLE_DEVICES='' .venv/bin/python scripts/docs_examples.py replot gpu-environment-sweep

Replotting verifies the data hash, scientific-source fingerprint and library hash before rendering on CPU. It preserves the execution timestamp, revision, device and dependency versions, saves the original executed source, and records rendering provenance separately. Scientific changes require a full refresh. Cumulative merger plots use dots at recorded observations connected by straight lines; they do not claim to resolve individual event times. Interval-rate plots retain steps because each value is an average over its measurement interval.

The old publications had not saved complete numeric histories, so a one-time data-saving CUDA refresh was necessary. The dry refresh at revision 54edfe13b7c99ec01b3a5f853e3ee45e6454ab61 took 3790.63 s and reproduced the 449,485 mergers and 1.554e-15 mass residual. Its 11 observations occupy 95,371 bytes. The subsequent CPU-only replot took 10.09 s and preserved the exact data bytes, all six figure hashes and original execution provenance.

The humidity capture at the same revision took 6054.08 s and reproduced 334 aerosol mergers, 105,243 attachments and 6,861 recombinations, with maximum mass residual 8.593e-14. Its 18 observations occupy 157,620 bytes. CPU replotting took 11.41 s and preserved exact data bytes, all eight figure hashes and original execution provenance. Both cumulative and overview figures were inspected; the remaining figures match the original publications byte for byte. These capture times include overlapping CUDA work and are not benchmark results.

The saved numerical SHA-256 digests are:

Example data.json SHA-256
Dry 8aab526a58e745252e1baa1ff9c13471219320a3495dabfc89a47c5856961924
Humidity a36f8b3dc9a1d8d5979e050442aa6fb2886ffb43b27b09733645059eefd044b1

The final infrastructure/example regression suite passed 527 tests with one skip and warnings treated as errors. Its integration checks cover an actual kernel refresh followed by replot with no additional simulation invocation, size-limit failures preserving the old bundle, data mutation rejection, scientific-input invalidation, stable filenames and publication rollback.

Each observation checks every box’s species inventories, including water gas, injected ion material and neutral-product exports, at 2e-11 relative tolerance. Total aerosol-plus-ion charge is exact. Aerosol charge alone can change. Aerosol spectra exclude free ions and neutral products. Histograms use dN/(N_initial dlog10 D) so their integral shows number loss. Charge counts and histogram counts reconcile independently with the represented population.

Aerosol N0−N counts aerosol-aerosol mergers. Free-ion balances separately count positive/negative attachment and recombination, checked against ion material attached to aerosols. Recombination creates an inert neutral particle; periodic cleanup exports its mass. Interval collision rates are averages, not exact event times. Pooled diameter quantiles and per-condition size/charge histories show dynamics beyond the final bars. Printed interpretations use actual run values; rankings do not establish statistically significant environmental responses, atmospheric validation or CPU/GPU speedup.

Explore this topic: Scientific and numerical correctness