v0.7.0 — real Born excitation, parameter-free, sub-keV CSDA closed

v0.7.0 — real Born excitation, parameter-free, sub-keV CSDA closed

This release came out of a Geant4-DNA expert review of the project. It addresses
the three top concerns at once, and the headline is a physics win where it matters
most: the low-energy end of the track.

The audit (reviewer #1: physics-list consistency)

A source audit (E29) settled whether the cascade and chemistry were validated
against the same Geant4 physics list. They are: both dnaphysics (cascade
oracle) and chem6 (chemistry oracle) register G4EmDNAPhysics_option2 —
no seam. And for option2, G4EmDNABuilder never enables the Emfietzoglou
models (the energy limit is 0; only option4 enables them), so the reference
excitation is Born, across the whole electron range.

The fix (reviewer #2: remove the σ_exc fudge)

The codebase had been approximating Born excitation by flat-scaling an
Emfietzoglou table (SIGMA_EXC_SCALE = 0.39). But the Emfietzoglou/Born ratio is
energy-dependent (~2.5× at keV, ~10× at ~10 eV), so one scalar left
low-energy secondaries over-excited ~4×. v0.7.0 loads the real
sigma_excitation_e_born cross section and removes the scalar — the
track-structure physics is now parameter-free (with RECOMB_BOOST already
gone since v0.5.0).

The result (reviewer #3: the sub-keV CSDA)

	100 eV	300 eV	500 eV	1 keV	10 keV
before (scaled-Emf)	0.782×	0.852×	0.894×	0.933×	0.997×
v0.7.0 (Born)	0.956×	0.986×	0.994×	0.987×	0.997×

The chronic sub-keV deficit — the project's weakest spot for its whole history,
and the energy range where track-structure codes earn their keep (the track ends,
the dense terminal clusters break DNA) — is closed. All 8 energies are now
0.956–1.005×. Cascade ions 0.937→0.942×; chemistry RMS flat at 7.0% (the G(H)
overshoot is gone); energy conservation 99.89%; primary bit-exact; 46/46 tests.

Honest notes

• The SSB indirect/direct ratio drifted 2.72→3.26 with the new physics. P_indirect
  is a calibrated fit; we report the drift honestly rather than re-tune it to
  the band — it demonstrates exactly why we treat DNA damage as methodology, not
  validated absolute physics.
• Scope: electrons, 100 eV–20 keV, low LET. Protons/heavier ions and realistic
  chromatin geometry are future work.

v0.6.1 — σ_exc → Born level (clean win)

v0.6.1 — σ_exc → Born level (clean win)

A one-parameter refinement that the v0.6.0 full electron cascade unlocked.

What changed

SIGMA_EXC_SCALE lowered 0.5 → 0.39 (≈ Born level). The 0.5 value had been
tuned for the truncated cascade — it over-excited to compensate for the
missing tertiary radicals. Once v0.6.0 tracked the full cascade and supplied
those radicals, ≈Born excitation became better on every axis, with nothing
regressing — and it shrinks the Emfietzoglou-vs-Born divergence, so the model
is more Geant4-faithful:

metric	v0.6.0 (σ=0.5)	v0.6.1 (σ=0.39)
CSDA @ 100 eV	0.736×	0.782×
CSDA @ 1 keV	0.912×	0.933×
CSDA @ 10 keV	0.994×	0.997×
cascade ions	0.931×	0.937×
chemistry RMS vs chem6	7.6%	6.8%
G(H) overshoot	1.085×	1.055×
G(e⁻aq)	0.887×	0.899×
SSB ratio	2.53	2.72 (in band)

How it was found

By directly testing — and refuting — my own prior conclusion. E26 had inferred
the residual cascade deficit was an immovable "σ_exc Emfietzoglou floor," and
E7e had read that σ_exc couldn't be lowered without breaking the chemistry. Both
were true for the truncated cascade and false for the full one — a direct
flip-and-measure test (E28) showed σ_exc=0.39 improves everything. Verify against
ground truth, don't trust the inference.

Validated end-to-end: CSDA (8 energies), cascade ions, chemistry G-values @ 1 µs,
SSB ratio, 0.1 ps initial yields. All artifacts under experiments/results/2026-06-09/.

v0.6.0 — full electron cascade

v0.6.0 — full electron cascade

The secondary shader now tracks the full electron cascade (tertiary / gen3+), instead of absorbing tertiary electrons in place. This resolves the long-standing cascade-ion deficit and is a clean win on every axis:

metric	v0.5.0	v0.6.0
cascade ions @10 keV vs Geant4	0.766×	0.931×
chemistry RMS vs chem6	19.7%	7.6%
G(H₂) vs chem6	0.74×	0.99×
G(H₂O₂) vs chem6	0.69×	0.93×
SSB indirect/direct	2.32	2.53 (band 2–3)

• The chem6 1 µs gap — the H₂/H₂O₂ deficits that stood through the entire project history — is largely closed.
• Primary track is bit-exact vs Geant4 (195.4 ionisations/primary vs 195.6, by trackID in the 6.8 GB ntuple).
• Validated across all 8 ESTAR energies; the browser build is bit-identical to the native runtime; energy conserved (99.9%).

How

secondary.wgsl emits tertiary electrons into sec_buf (G4DNABornAngle direction sampling, OH+H₃O⁺ products, deferred eaq); dispatch.ts grows the Phase B wavefront in chunks so they get tracked.

Honest note

The investigation included a self-caught error: a normalization bug in the analysis (run_irt invoked with the wrong n_therm) briefly made the fix look like it broke chemistry, and two experiments (E23–E24) chased a phantom before verify-before-asserting caught it (E25). The fix was always a clean win.

Full details in CHANGELOG.md.

v0.5.0 — parameter-free pipeline

v0.5.0 — parameter-free pipeline

The last unphysical tuning fudge (RECOMB_BOOST) is removed (set to its neutral 1.0). The pipeline is now parameter-free — the only remaining non-unity scalar, SIGMA_EXC_SCALE = 0.5, is a documented Emfietzoglou-vs-Born physics divergence, not a tuning knob.

The flip was validated end-to-end on all three gates:

• Cascade ions recover 0.677× → 0.766× (32% → 23% deficit). Source archaeology found RECOMB_BOOST=2.0 was destroying ~⅓ of the autoionisation H₃O⁺ ions via boosted geminate recombination; at the physical 1.0 they survive. [E7d]
• Chemistry stays parameter-free at +1.4 pp RMS and improves OH/eaq/H (removes the H overshoot). [E10r]
• SSB indirect/direct ratio holds in PARTRAC's 2–3 band at 2.32 (was 2.46) with no P_indirect recalibration. [E13d]

Removing the fudge improved the cascade-ion deficit and made the chemistry parameter-free — the former "two-knob tradeoff" was an artifact of the fudge fighting the physics, not fundamental.

Production, README §Numbers, the paper, and the live site (webgpudna.com) all now report the same parameter-free config.

Also in this release

• E12-local-exact: exact voxel dose (C=991) confirms the C≈981 proxy to 1% — L5 absolute-yield vindication now rests on measured dose.
• Free-compute infrastructure (FREE_COMPUTE.md): IRT chemistry + SSB/DSB validation run GPU-free on GitHub Actions 16 GB runners. Kaggle/Colab GPU path tested and closed (compute-only, no Vulkan). Oracle Always Free runbook for Geant4.

Full details in CHANGELOG.md and the corrected paper (paper/main.pdf).

Validation inputs (rad_buf dumps for CI chemistry)

Pre-chemistry rad_buf dumps + 10 keV dose grid (WebGPU Phase A+B output) for the CI chemistry/SSB validation. Shaders: parameter-free production (σ_exc=0.5, RECOMB_BOOST=1.0) as of 2026-06-08. N=4096 primaries. Regenerable via experiments/lib/regenerate-dumps.mjs. rad_E10000_recomb1.bin is the same 1.0 dump kept for the E13d artifact provenance.

v0.4.1 — paper + parameter-free chemistry

Paper + parameter-free chemistry. Adds the arXiv-style preprint and resolves
the chemistry stage's reliance on an unphysical knob.

Added

• paper/ — arXiv-style LaTeX preprint (main.tex, refs.bib, main.pdf)
  with six figures generated from the committed artifacts (figs/make_figs.py).
  All references verified against the publisher (exact pages + DOIs).
• E10r — parameter-free chemistry measurement. Re-ran the IRT at
  RECOMB_BOOST = 1.0 (regenerated dump): the unphysical knob is not
  load-bearing — removing it shifts the 5-species RMS vs chem6 @1µs by only
  ~1.4 pp (18.3% → 19.7%), improves G(OH)/G(eaq), and removes the G(H)
  overshoot. The paper now reports the parameter-free yields as primary.

Changed

• Corrected gMicroMC/MPEXS attribution: both use a Smoluchowski
  reaction-radius contact test, not the Brownian bridge; the bridge is
  re-attributed to Clifford et al. 1986.
• README documents the real deploy path (Cloudflare Pages via wrangler,
  project webgpudna) and the parallel Vercel mirror.
• GEANT4_DIVERGENCES.md B2 row updated with the measured non-load-bearing
  result; removed the empty tests/integration/ directory.

v0.3.0 — Research-grade closure release

Live: https://webgpudna.com — every number on the site is now backed by a falsifiable JSON artifact under experiments/results/, with README § Numbers as the single source of truth.

Highlights

• L5 marquee closure — SSB indirect / direct ratio 0 → 2.96 (in PARTRAC's published 2-3 band) via 4 stages:
  1. Split SSB_R_DAMAGE_NM into direct=0.29 + indirect=1.0
  2. Instrument public/irt-worker.js to score every OH-death event + 1 μs survivors
  3. Calibrate SSB_P_INDIRECT 0.4 → 0.05 (the IRT-time accumulator visits ~10× more events than the 1 μs-only scan, so the per-event probability needed scaling down)
  4. Thread the per-bp hit mask into the DSB clusterer so co-located strand-0/1 indirect hits cluster correctly
• Joint physics fix in src/shaders/ — SIGMA_EXC_SCALE = 0.5 (partial Emfietzoglou rollback) + RECOMB_BOOST = 2.0 (approximates Geant4's time-integrated e-h recomb). E10i measures RMS dev vs chem6 dropping 30.3% → 19.0%, CSDA @ 100 eV lifting 0.587× → 0.74×, G(H₂) closure 0.51× → 0.78×.
• E5b surfaces the energy-dependent CSDA deficit — 0.587× @ 100 eV → 0.992× @ 20 keV. The previously isolated "0.988× @ 10 keV / 3.59σ" was the tail of a much larger sub-keV deficit, not an isolated artifact.
• E10e/f/g/h decompose the 0.1 ps H₂/H₂O₂ deficit:
  • Cross-event recomb: REFUTED (3.5% of gap)
  • Per-primary IRT partitioning: REFUTED at 0.1 ps; CONFIRMED for 96% of 1 μs gap
  • Recomb-rate sensitivity: linear sweep places the missing factor at ~25%
  • Time-integrated recomb with proper H₂Ovib branching: 30% → 22% RMS dev but G(eaq) takes collateral damage (two-knob structural limit)
• README restructured — every quantitative claim moved into a single § Numbers section at the bottom. CLAUDE.md, index.html, and the OG image are summaries that link back; no number lives in two places.

Validation status

Per-level results (all rows backed by JSON artifacts in experiments/results/):

• L0 (env): 2/2 pass — B0 browser env, B1 harness liveness
• L1 (cross sections): 9/9 pass — E1/E1b/E1c, E2/E2b, E3/E3b, E4/E4b
• L2 (track structure): 3 pass / 2 honest negative / 1 partial — E5/E5b/E6/E6b/E7/E8
• L3 (pre-chem): 1 honest negative — E9
• L4 (chemistry): E10/E10b pass, E10c/E10d/E11 honest negatives, E10e refuted, E10f mixed, E10g/E10h/E10i quantify the path forward
• L5 (DNA damage): L5 indirect SSB gap closed in E13c (ratio 2.96 in PARTRAC band); E12 within factor-5 pass band
• L6 (performance): 455× vs Geant4 11.4.1 single-thread, 280× vs Geant4 MT-8 (production-realistic), 40× kernel-fusion factor

46 / 46 unit tests pass.

Changelog

See CHANGELOG.md for the full list of changes, including:

• 4 new experiment IDs (E5b, E10e, E10f, E10g, E10h, E10i)
• Joint shader fix (SIGMA_EXC_SCALE, RECOMB_BOOST constants)
• L5 indirect-SSB scoring refactor in public/irt-worker.js
• Docs consolidation into README § Numbers
• Build fix: pin fetch-demo.mjs to demo-v1 tag

What's next

PHYSICS_DIAGNOSIS.md tracks the remaining open gaps:

• H₂O+ tracking with proper time-integrated recomb (~3 hr) — replace the RECOMB_BOOST constant with the physical model
• W_sec distribution shifter (~2 hr) as a third knob
• E14 vs molecularDNA full chromatin geometry (~1 day, deferred)

🤖 Generated with Claude Code

v0.1.0 — WebGPU Geant4-DNA in a browser tab

First versioned release of webgpu-dna — a WebGPU port of Geant4-DNA
running entirely in the browser.

Live: https://webgpudna.com

Highlights

• CSDA range 0.985× of Geant4-DNA at 10 keV (N = 4096 primaries)
• Energy conservation 100.0%, ions per primary ratio 1.00×
• Karamitros 2011 9-reaction IRT chemistry in a Web Worker
• 21×21 parallel B-DNA fibers × 3 μm — 3.89 Mbp scoring grid
• 46 unit tests across 7 files (Vitest)
• Mobile-friendly WGDNA-4D Gaussian-splat viewer with touch + pinch

What's in this release

See CHANGELOG.md for the full list including the physics
engine, chemistry pipeline, DNA scoring, validation harness, G4EMLOW
converter, and the 4D viewer.

Get started

```bash
git clone https://github.com/abgnydn/webgpu-dna
cd webgpu-dna
npm install
npm run dev # → http://localhost:8765
```

WebGPU support required (Chrome/Edge 113+, Safari 26+, Chrome 121+ Android,
Firefox 141+ Windows / 145+ macOS Apple Silicon).

MIT licensed. Author: @abgnydn.

This release supersedes the existing `demo-v1` tag (which is a snapshot of
the WGDNA-4D demo .bin data, hosted out-of-repo to avoid git bloat — that
tag stays around as a download mirror).

WGDNA-4D demo snapshot v1

Bundled demo snapshot for the /splat.html 4D viewer.

Generated by the Geant4-DNA WebGPU validation harness:

• 4096 primaries @ 10 keV
• 50,000 radicals subsampled per checkpoint
• 8 checkpoints across 1 ps → 1 microsecond
• 21x21 B-DNA fiber grid included

Hosted as a release asset so the 8.4 MB binary stays out of the git history.
The viewer fetches the latest copy via:
https://github.com/abgnydn/webgpu-dna/releases/latest/download/wgdna-default.bin