Conversation
Three coupled defects of the DCMIP2016 Kessler kernel, found by the SCM calibration audits, are fixed together because they share the same bookkeeping: 1. Sedimentation geometry: the upwind rain-flux divergence divided by the distance between cell centers (and a half cell at the top) instead of the finite-volume cell thickness, so the column rain budget did not close on stretched grids (every junction cell) and the top cell lost rain at twice the correct rate. Fluxes ρqʳ𝕎ʳ are now divided by Δzᶜᶜᶜ, with zero inflow at the top and the CFL bound on every cell's thickness. 2. Density basis: on the anelastic core the kernel moved ρᵣ rʳ 𝕎ʳ in mixing ratio space at fixed total density and converted back with the changed rᵗ, rescaling vapor and cloud in every cell rain passed through and reporting a surface flux that was not the rain removed. Sedimentation now updates the prognostic partial density ρqʳ directly; the Kessler processes act on dry-air mixing ratios with ρᵈ = ρᵣ - ρᵗ (anelastic) or the prognostic ρᵈ (compressible), so the prognostic water budget closes to the substep-mean surface flux exactly. 3. Phase-change thermodynamics: T was incremented by ℒˡᵣ Δrˡ / cᵖᵈ and θˡⁱ re-derived from it, a percent-level inconsistency with the formulation's own θˡⁱ ↔ T relation that acted as a net θˡⁱ sink. Sedimentation is now applied at fixed T and phase change at fixed θˡⁱ, with the DCMIP single Newton step linearized by ∂T/∂rˡ at fixed invariant (`phase_change_temperature_slope`), matching `SaturationAdjustment`. The static-energy formulation is rejected with a clear error; the parcel path uses the same slope for its own invariant. Tests: a new `dcmip2016_kessler_conservation.jl` (budgets on uniform, junction and geometric grids, top cell, vapor untouched by sedimentation, all-process water closure, θˡⁱ conservation for condensation, cloud and rain evaporation across 1000-470 hPa, finite-difference check of the slope, Float32 and Float64), and the column reference implementation in `dcmip2016_kessler.jl` rewritten to the documented algorithm on a stretched grid, with a compressible water-budget check. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…Kessler tests - The constructor docstring now states that negative water partial densities are clipped to zero on entry (unchanged behaviour), that clipping creates the clipped mass, and that the closure holds for non-negative inputs only. A new test pins the water created by clipping to exactly the clipped mass. - The conservation tests read the model temperature after `set!`; `update_state!` diagnoses it from the scheme's diagnostic mass fractions, which it refreshes only after reading them, so the helper now runs two passes before recording the pre-step temperature (this is what made the latent-heating-sign checks fail on CI). - The reference-comparison test places rain in the bottom cell as well, so the surface flux it asserts on is actually exercised within one 10 s step. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ce; retune the clipping test - `test/dcmip2016_kessler.jl`: the DCMIP2016 Fortran translation that the previous "Julia vs Fortran" test used is restored verbatim as `dcmip2016_fortran_kessler!` (documented as the legacy algorithm, not the current one), together with its per-cell adjustment block `legacy_kessler_adjustment`. Two new test sets compare the current kernel with it only where the physics is unchanged: the process formulas (autoconversion, accretion, single-step saturation adjustment, rain evaporation) agree to 1e-14 when the current step is given the legacy latent-heating slope ℒ/cᵖᵈ, and the sedimentation of dry-air-carried rain on a uniform grid away from the top cell (compressible core, dry column, evaporation off) gives identical rain, θˡⁱ and surface flux. Whole-column equality is not claimed where the geometry, density basis or temperature update were deliberately corrected; those are covered by `kessler_column_reference!` and the conservation tests. - `test/dcmip2016_kessler_conservation.jl`: the negative-input clipping test uses a dry raining column and a tolerance scaled by the inventory round-off, so the created mass is resolvable at Float32 (the previous relative tolerance on the created mass itself failed by 1e-4 on CI). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Codecov Report✅ All modified and coverable lines are covered by tests. 📢 Thoughts on this report? Let us know! |
Documenter cannot resolve `@ref` links to `LiquidIcePotentialTemperatureState` and `StaticEnergyState`, which have no docstrings, so the docs build failed; refer to them as plain code instead. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…the clipped-relative bound - `test/dcmip2016_kessler.jl`: a new test set measures, on the former "Physical fidelity" column (anelastic, uniform 100 m, all processes, one 10 s step), how far the corrected kernel departs from the legacy Fortran translation where the physics was deliberately corrected, and bounds it: vapor < 2e-3, cloud < 3e-2, rain < 1e-1 of the species maxima; θˡⁱ by less than 2 % of the step's latent heating (measured 0.56 %) and by more than 1e-3 K (so the deliberate departure cannot silently vanish); the legacy convention's water residual is nonzero while the current one is at round-off. - `test/dcmip2016_kessler_conservation.jl`: the clipping test keeps the dry raining fixture but returns to the bound relative to the created mass (`≈ clipped rtol=budget_rtol(FT)`), which the dry fixture satisfies with margin (independent measurement: Float32 error 9.3e-9 against a 1e-6 bound); the injected negatives are doubled and the signal-to-round-off margin `clipped / (eps(FT) W₀ Nz) ≥ 1000` is asserted explicitly. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR fixes three conservation issues in Breeze’s DCMIP2016 Kessler implementation. Rain sedimentation now uses finite-volume cell thicknesses and includes the top cell in the CFL bound. It transports rain partial density directly, leaving vapor and cloud water unchanged and making the reported surface precipitation flux consistent with the water removed from the column.
Condensation and evaporation now preserve the liquid–ice potential temperature used by the model, with a temperature response and saturation-adjustment slope consistent with that formulation. Sedimentation retains the existing fixed-temperature convention, and accretion uses the post-sedimentation rain. Unsupported atmospheric use with StaticEnergyFormulation now produces an explicit error.