ACCESS-AIS3 | Development record

Where the project stands

The present-day state is no longer effectively static: the main initialization and inversion failures have been isolated, repaired, and tested. The current production state uses Schoof friction at 70.43 m/yr grounded RMSE and is verified converged in practice. A reproducible, fully converged Budd state reaches 98.9 m/yr, while a promising 60.4 m/yr Budd result still needs validation through the production pipeline. The latest pipeline update validates the SSA historical run at 1.402 m/yr dH/dt RMSE; the HO historical run continues through chunked warm restarts.

70.43 m/yrcurrent production grounded RMSE
Schoof, converged in practice

Validated

HO thermal spin-up

Full higher-order thermal solve converged and produced AIS3_thermal_steadystate.nc, with post-solve temperature safeguards and rheology recomputation.

Paused

HO friction inversion

Chunked continuation reached a stable checkpoint, but chunk 4 failed five times with two failure signatures. The continuation is paused pending a better diagnosis.

Validated + active

Historical pipeline

SSA historical tuning is complete at 1.402 m/yr dH/dt RMSE. The HO track is still running in six-year warm-restart chunks; the melt calibration is finalized at gamma_0=300 with basin refits.

Results worth carrying forward

9.01%of grounded vertices carry 90.6% of total squared error

17×real spatial viscosity range restored from RACMO temperature

25 / 25Budd sensitivity runs converged cleanly after the boundary-condition fix

1.402 m/yrvalidated SSA historical dH/dt RMSE against MIPKIT

Configuration	Grounded RMSE	What it means
Null model	310.6 m/yr	Baseline to beat
Original setup	323.5 m/yr	Model was effectively static
Friction-only, 200 iterations	110.6 m/yr	83% within 50 m/yr
Schoof, AIS3_inverted.nc	70.43 m/yr	Current production state; converged in practice
friconly_nfix, Budd p=q=3	98.9 m/yr	Best validated, reproducible self-consistent SSA result
Budd p=q=1, regularised	60.4 m/yr	Best RMSE found; not yet validated through the production step
Forward-only damage-relaxed shelf B	67.3 m/yr	Not self-consistent; does not survive re-optimization

What changed the diagnosis

Friction left its dead zone

The Budd initial coefficient was moved from 100 to 1.8, where the inversion has useful sensitivity. The former results were discarded because the model was not moving.

Rheology became spatially physical

Uniform placeholder B was replaced with a RACMO surface-temperature-derived field. The temperature remains a proxy for depth-averaged rheology, but it restores the missing spatial variation.

Grounding-line drag and front BCs were repaired

The drag mask is vertex-based, and the floating calving front is cleared back to Neumann conditions after extraction. This avoids stripping drag from grounded ice and avoids an artificial velocity clamp at the front.

The mountains exposed a geometry error

Experiment scripts rebuilt the surface after applying a thickness floor, fabricating steep slopes and driving stress. Production now preserves the observed surface and absorbs the floor into the base.

Historical runs became restartable

The HO historical run now advances in six-year warm-restart chunks, recomputes geometry to enforce base-above-bed consistency, and leaves extra wait-time margin for hugemem queue congestion. The SSA track is complete and validates at 1.402 m/yr dH/dt RMSE.

Pipeline map

Stage	Step name	Status	Output
1. Thermal spin-up	`ho_thermal_steadystate`	Validated	`AIS3_thermal_steadystate.nc`
2. HO friction inversion	`ho_friction_inv`	Paused after repeated chunk-4 failures	`AIS3_ho_friction_inv.nc`
3. Melt calibration	`melt_gamma_tuning`	Finalized; naive RMSE invalid	`gamma_0=300` with basin refits
4. Relaxation	`ho_relaxation`	Completed solve; output salvage/qualification required	`AIS3_ho_relaxed.nc`
5. Historical tuning (HO)	`historical_dhdt_tuning`	In progress; chunked warm restart	`AIS3_historical_1995_2019.nc`
5. Historical tuning (SSA)	`historical_dhdt_tuning_ssa`	Complete and validated	`1.402 m/yr RMSE`
6. Future projections	`projection_ssp`	Scaffold	`AIS3_projection_{gcm}_{scenario}.nc`

Scaffolded stages intentionally retain explicit TODOs for unresolved science choices. Their intermediate outputs must not be treated as validated results.

Selected figures

Observed velocity, modelled velocity, and residual maps for PIG/Thwaites, Siple Coast, and Totten — **Best regional inversion benchmark** Budd p=q=1 with regularisation reaches 60.4 m/yr domain grounded RMSE in this benchmark. It is the best RMSE found, but has not yet been reproduced through the production inversion step. Open full resolution

Antarctic basal temperature, surface temperature, vertical temperature gradient, basal speed, and temperature distribution — **Validated HO thermal steady state** Continental temperature structure, basal speed, and temperature distribution from `AIS3_thermal_steadystate.nc`. Open full resolution

Observed and modelled surface speed, residual, and inverted basal friction over the PIG and Thwaites subdomain — **HO friction inversion validation** Observed versus modelled surface speed, residual, and inverted basal Schoof friction for the PIG/Thwaites test domain. Open full resolution

Inverted floating-ice rheology B, modelled velocity, observed velocity, and residual across Antarctica — **Floating-ice rheology inversion** Corrected temperature-dependent shelf rheology B alongside modelled and observed velocity and the residual field. The plotted state reports 56.1 m/yr RMSE. Open full resolution

Observed and modelled Antarctic velocity, grounded residual, and inverted friction coefficient for the Budd p equals q equals 1 benchmark — **SSA friction inversion benchmark** Budd p=q=1 with light regularisation, showing observed and modelled velocity, grounded residual, and friction coefficient C. This benchmark reaches 60.4 m/yr grounded RMSE but still needs production-step validation. Open full resolution

Read the source record

Inversion worklogDiagnostics, repairs, benchmarks, dead ends, regional failures, and the current candidate.

Post-inversion pipelineThermal spin-up, restart strategy, resource limits, and the six-stage pipeline status.

Run guideWhich script runs each pipeline stage on Gadi, its inputs, outputs, resources, and known failures.

Boundary-condition notesBoundary-condition choices and the extraction workflow used by production runs.

Active investigation notesBudd versus Schoof, effective pressure sources, and current friction experiments.