Changelog

All notable changes, newest first. Versions follow semantic versioning; until 1.0 a minor release may change results and file formats, and says so here.

0.5.0 — 2026-10-13

This release changes results. Runs that used the Skyfield or SPICE ephemeris had the Sun in the wrong place; runs with the synthetic Sun had a Sun that rose and set like at the equator. Re-run anything you made with 0.4.

Geometry

• Sun direction on the terrain. The Skyfield path imported a lunar frame that Skyfield does not provide and fell back to a rotation built from active instead of passive matrices; the Moon’s pole came out about 45° from where it is. The local frame was north/east/up, which is left-handed and mirrors every azimuth, and the Sun vector was never rotated from local east/north into the DEM’s grid axes. All Sun sources now return a body-frame vector and one module (artemis_thermal/core_engine/frames.py) rotates it onto the mesh, using the DEM’s own CRS. The IAU 2015 lunar orientation, now with its thirteen periodic terms, agrees with SPICE MOON_ME to better than 0.003°.
• Polar stereographic convention. polar_stereographic.py used y = −ρ cos λ, the mirror image of the PROJ CRS our DEM tool writes. It now matches PROJ, and a test checks the two against each other.
• Lunar curvature. Terrain is lowered onto the sphere (terrain.curvature, on by default): 115 m at 20 km from the DEM centre, which matters when the Sun is 1.5° above the horizon.
• Synthetic Sun. The sub-solar point now circles the Moon once per synodic month at a fixed latitude (ephemeris.synthetic_subsolar_lat_deg, default −1.5°). At the pole that keeps the Sun near the horizon all day, as it is.

Physics

• Sublimation is the Hertz–Knudsen mass flux with the Murphy & Koop (2005) vapour pressure. The old function returned a number flux labelled kg m⁻² s⁻¹ (off by 26 orders of magnitude) and the Langmuir form used a natural-log constant inside a log₁₀ formula; one test asserted that ice at 200 K lasts a million years. It lasts about 50 days.
• Cold traps are classified from the hottest temperature a face reaches during the run, with the thresholds from the configuration, instead of from the last time step with hard-coded thresholds.
• Initial condition. Each column starts at the temperature that radiates away its day-mean absorbed sunlight (solver.thermal.initial_condition: equilibrium). Short runs no longer report the 200 K starting guess as their maximum.
• Newton iteration now stops when the whole column has converged, not only the surface node.
• Research-mode checks also run inside the runner, so --ephemeris synthetic can no longer bypass the SPICE requirement. A failing roughness model stops a research run instead of being silently replaced.

Speed and memory

• BVH construction is compiled: 583k triangles in 1.1 s instead of 338 s. Shadow maps are bit-for-bit identical to 0.4.
• Ray kernels no longer allocate per ray, and view-factor post-processing is compiled. Reciprocity is now measured as the mismatch between the power each facet sends into the terrain and the power it receives (0.16 at 128 rays, 0.08 at 512 on a test bowl). The old per-entry check with a 10⁻⁵ tolerance could never pass on Monte Carlo data, and printed a warning on every run.
• A run no longer keeps one full-mesh copy per output step (11 GB for a month at 1M faces). It keeps running minimum, maximum and mean maps and streams float32 snapshots to disk inside output.max_snapshot_mb.
• The memory check before a run covers mesh, BVH, thermal state, work arrays and view factors (resources.max_memory_mb, 1.5 GB by default, replacing solver.thermal.max_state_memory_mb).
• The thermal kernel precomputes density and conductivity profiles and updates columns in place; per-step logging moved to DEBUG.

Desktop app

The Streamlit dashboard is replaced by a PySide6 app (artemis ui): simulations run in a separate process with progress, estimated time and cancel; results show as rasters with a time scrubber; clicking a cell reports slope, aspect, temperatures, sunlit share, mean flux on a Sun-tracking panel and the longest dark spell; two runs on the same grid can be compared as a difference map. Light and dark themes, bundled fonts, no network access.

Command line and files

• One package. Everything now lives under artemis_thermal (artemis_thermal.simulation.runner, artemis_thermal.core_engine.frames, …). 0.4 installed seven top-level packages with names such as gui, simulation and validation that would collide with other projects. Update imports by prefixing artemis_thermal..
• artemis has subcommands: run (the default), ui, render, configs, warmup. It works from any directory; bundled configurations are found by name (--config research_shackleton).
• New run-folder format (version 2, documented in Outputs), read lazily by artemis_thermal.simulation.io_manager.ResultsBundle. Folders written by 0.4 are recognised and the app asks for a re-run.
• New configuration shackleton_preview: real LOLA terrain and SPICE Sun at laptop cost.
• Report figures are drawn from rasters and share their colours with the app.
• tools/fetch_research_data.py pins kernel hashes, resumes interrupted downloads and rejects mismatches. The de440s.bsp that 0.4’s downloader produced was 1.3 MB larger than the file NAIF serves (the old resume code appended to a partial file); its ephemeris data were identical, but the hash was not. The pin is now the NAIF file’s hash. tools/prepare_lola_dem.py falls back to the S3 mirror of the USGS product.
• rasterio, skyfield and spiceypy moved to the dem and ephemeris extras; tqdm, xarray, plotly and kaleido are no longer needed.

Removed

The Streamlit dashboard, artemis_thermal/thermal_solver/regolith_properties.py (unused, and a copy of the solver’s own property functions), tools/download_sample_data.py (downloaded a shaded-relief image and treated it as elevation), unused ray-tracer and ephemeris helpers, unused configuration keys, and tests that did arithmetic on literals without calling the code.

0.4.0 — 2026-08-03

• Fail-closed research mode: provenance-verified LOLA DEM, NAIF DE440 kernels, minimum spin-up.
• Disk-integrated projected irradiance per solar sample.
• Geothermal heat applied once, at the bottom boundary.
• Dense batched thermal state (about 5× faster than per-face objects).
• Checksums for every saved result file.

0.3.0 — 2026-04-07

• Hapke directional-hemispherical albedo, Bandfield roughness emissivity.
• Sparse Monte Carlo view factors for one bounce of terrain infrared.
• Volatile stability classification.
• Streamlit dashboard (replaced in 0.5.0).

0.1.0 — 2026-02-12

First public version: synthetic crater, BVH shadows, Crank–Nicolson regolith columns. (This version was tagged “v1.0” in its commit message; it was not a stable release.)