PANDA

Current release: v1.7.21 — WSL2 Results-Folder Focus Fix, Take Two. The v1.7.19 fix (launching explorer.exe directly instead of xdg-open) wasn't sufficient on its own: Windows' foreground lock blocks SetForegroundWindow from a background WSL-interop-launched process, and Explorer just flashes the taskbar entry instead of raising an already-open window for the same folder. PANDA_GUI.py's "Open Results Folder" button now routes through a small PowerShell helper that finds/opens the window via the Shell.Application COM object and explicitly forces it to the front via user32.dll's ShowWindow/SetForegroundWindow -- confirmed working. Also: run.mac's dummy 28nm-node test geometry (.04/.04/.3 um sensitive, 3 um dead layer) had criticalCharge corrected from an unrealistic 150 fC (far above anything this device ever deposits) down to 1 fC, with Results/Current refreshed to match.

PANDA — Protons And Neutron charge Deposition in mAterials — is a Geant4 Monte Carlo simulation for calculating charge deposition from energetic particles (protons, neutrons, ions) in semiconductor structures, for Single Event Effect (SEE) analysis.

The core physical quantity computed is collected charge per event; upset probability, cross-section spectra, and other SEE metrics are derived from it in post-processing (PANDA_Analyze.py), keeping particle transport and interpretation cleanly separated.

Compared against CREME-MC (see compare_creme_panda.py) and against McNulty et al.'s published proton-induced experimental/CUPID data (see compare_mcnulty_panda.py). Includes optional cross-section biasing (SEEBiasingOperator) to efficiently sample the rare nuclear-recoil tail, with correct raw-charge/event-weight separation so biased and unbiased runs produce statistically consistent spectra. The dead+sensitive stack sits inside a configurable bulk "surrounding volume" (matching the sensitive material, /sim/surroundingXY//sim/surroundingThickness), with optional biasing of secondary neutrons produced there too (/sim/secondaryNeutronBiasFactor), so nearby nuclear reactions can contribute recoils into the sensitive volume.

Primary particles

Set via /sim/particle in a macro, or the Particle dropdown in PANDA_GUI.py: proton, neutron, alpha, deuteron, triton, He3, e-, or one of six heavy-ion primaries -- C12, F19, Cl35, Ni58, I127, Au197 (ground-state, most-abundant-stable-isotope-per-element, matching common heavy-ion SEE test cocktails). Cross-section biasing (/sim/biasCrossSectionFactor) is validated for every species except e- (see PANDA.cc's wrapped-process list). Unlike SRIM -- which computes a bare ion's stopping power/range but has no device or charge-collection model -- PANDA reports cross-section-vs-LET through its own geometry, charge-collection-efficiency, and angle-of-incidence models for these ions directly, covering the direct-ionization-dominated regime of SEE testing alongside its existing proton/neutron nuclear-recoil focus. See Documentation/PANDA_MASTER_DESIGN's Cross-section biasing exception (Section 5) for the ion-creation ordering constraint this needed and how it was resolved.

Nuclear-recoil tail beyond SRIM: SRIM only computes a bare ion's stopping power along its own track, so it has no way to predict the population of secondary recoil fragments produced when a heavy ion undergoes a nuclear reaction (ionInelastic) inside the device -- fragments that can carry higher LET than the primary ion's own direct ionization. PANDA's differential charge spectrum shows this directly: a 200 MeV Cl35 beam through a 1 um dead layer + 1 um Si sensitive volume, with sensitiveXY set to 0.1 um (representative of a 65 nm SRAM cell's sensitive volume), produces a sharp direct-ionization peak matching the SRIM-predicted LET (2.694E+00 MeV*cm^2/mg -> ~28 fC expected from LET x thickness, matching the observed peak just above 20 fC), plus a separate, much rarer tail extending out past 100 fC from nuclear-recoil events that a SRIM-only hazard assessment would never predict. Despite the initial concern that such a small lateral sensitive volume might be an edge case for the geometry/charge-collection model, the resulting spectrum was well-behaved and physically sensible at this deep-submicron scale. This is the same underlying mechanism (nuclear-reaction-produced recoils, not direct ionization) behind PANDA's proton/neutron SEE-susceptibility work -- heavy ions just add a high-LET direct-ionization peak on top of it.

Materials

The sensitive volume and dead layer/electrode each have an independently selectable material, set via /sim/sensitiveMaterial and /sim/deadMaterial in a macro, or the matching dropdowns in PANDA_GUI.py (next to the Sensitive/Dead Thickness fields):

• Sensitive volume: Si, GaAs, Ge, SiC, GaN
• Dead layer/electrode: SiO2, Al2O3, TiO2, Si

The sensitive volume's material also selects the pair-creation energy and carrier mobility/saturation velocity used to convert deposited energy to charge -- see DetectorConstruction::GetSensitivePairCreationEnergy() and neighbors for the per-material constants and sources. Defaults are Si/Si, matching PANDA's original silicon-only behavior.

Per-hit recoil/LET export

Set /sim/logRecoilHits true (default false) to additionally write Results/Current/recoil_hits.csv, one row per LET-worthy hit in the sensitive volume (species, Z, A, LET in MeV·cm²/mg, position, EventWeight), filtered to recoils only (excludes proton/e-). This is a finer-grained companion to events.csv's per-event Proton_keV/Electron_keV/PrimaryIon_keV/Recoil_keV sums -- useful for studying the recoil-species/LET spectrum directly, e.g. checking where the highest LET a given recoil species reaches actually lands (a common reference point: silicon recoils are often cited as topping out around LET~12, informing the assumption that heavy-ion hardness above LET=20 implies proton/neutron-SEE immunity). Off by default since per-step file writes add real overhead most runs don't need. A hit's LET is computed from edep/path length accumulated across a track's consecutive steps inside the sensitive volume (EventAction::AccumulateRecoilHit()), only exported once that running path reaches a validated 100nm minimum -- geometry-independent, since it no longer depends on how finely Geant4 subdivided the real path into individual steps (see Documentation/PANDA_VALIDATION_SUMMARY.md Section 2). A track whose total in-volume path never reaches 100nm is genuinely too short a sample and is excluded; events.csv's PrimaryIon_keV/Recoil_keV columns (not gated by this floor) still capture its energy either way.

events.csv's PrimaryIon_keV column (added alongside the existing Proton_keV/Electron_keV/Recoil_keV) separates the primary beam particle's own track (any species other than proton/e-, i.e. alpha and every heavier ion) from genuine secondary nuclear-recoil deposits, which previously shared one Recoil_keV bucket purely because both are "not proton, not e-". For a 200 MeV Au197 primary this is the majority of Total_keV (~88%) -- that's the primary ion's own expected electronic stopping (SRIM's "Ions" curve), not a nuclear-recoil effect, and it's why Recoil_keV alone should not be read as a recoil-population indicator for non-proton/e- primaries.

Run python3 PANDAEX_Analyze.py to analyze it: a per-species summary table (count, weighted count, Z, A, max/mean LET, saved as PANDAEX_recoil_species_summary.csv) and a differential LET spectrum plot (PANDAEX_LET_spectrum.png, overall plus the top species by hit count). Falls back gracefully if recoil_hits.csv doesn't exist.

Angle of incidence (approximate)

Set /sim/incidentAngle (degrees, default 0) to approximate a beam tilted off the sensitive volume's normal. Rather than rotating the beam or geometry (which would also need to reshape the dead layer, surrounding volume, and secondary-neutron-biasing region), this divides the sensitive volume's effective thickness by cos(incidentAngle) -- the same chord-length-elongation model used for tilt-angle corrections in heavy-ion SEE testing (LET_eff = LET(0 deg)/cos(theta)), valid for a broad-uniform-beam test condition (beam footprint much larger than the device). The lateral (XY) footprint is intentionally left unchanged -- no physically-grounded correction for it was found; see Documentation/PANDA_MASTER_DESIGN for why. Valid range [0, 90) deg; anything outside that raises a fatal exception rather than silently producing an infinite or negative thickness.

See Documentation/PANDA_VALIDATION_SUMMARY.md for the full validation summary -- comparisons against McNulty et al./CUPID and CREME-MC/MRED (both against independently-specified geometry), plus every geometry/physics robustness check run against PANDA itself. (Hitachi HM68512 open proton-SEU data is also discussed there, but reclassified as a capability demonstration rather than a validation, since its device geometry and critical charge were never published.)

Known limitations

• CREME-MC comparison shoulder (resolved 2026-07-19): the previously reported ~1-2 order of magnitude gap was a comparison-methodology bug in compare_creme_panda.py -- it compared CREME-MC's ideal-100%-charge-collection reference against PANDA's collected-charge curve rather than the deposited-charge curve (PANDA's own 100%-collection-equivalent quantity). Corrected, the two curves agree closely (log-RMSE ~0.29 decades) across the full range, including the region previously described as an unexplained "hump." See Documentation/PANDA_VALIDATION_SUMMARY.md Section 1.3.

• McNulty et al. comparison under-prediction (reframed 2026-07-19): comparing PANDA's integral cross-section-vs-deposited-energy curve against McNulty et al.'s 148 MeV proton data for an 11.7x11.7x13.5 um silicon sensitive volume (compare_mcnulty_panda.py) originally showed PANDA systematically under-predicting, growing smoothly from ~1.4x at 1 MeV deposited to >100x by 25-28 MeV (log-RMSE ~1.1 decades), traced to the dead+sensitive stack sitting in a vacuum world with no material for nearby nuclear reactions to occur in -- the actual subject of the 1989 paper this comparison was aimed at (El-Teleaty, McNulty, Abdel-Kader, Beauvais, Nucl. Instr. and Meth. B40/41, 1300-1305). Adding a 100 um silicon surrounding volume plus biasing secondary neutrons within it improved log-RMSE to ~0.91 decades -- a real ~18% reduction -- but a gap remained at high deposited energy (e.g. ~48x at 25 MeV, down from ~133x). Two candidate causes were tested directly and ruled out: widening the beam to cover the surrounding volume made agreement slightly worse (log-RMSE 0.426 -> 0.455 decades over a matched range), and growing the surrounding volume to 300/500 um showed no improving trend (0.900 -> 0.887 -> 0.945 decades). Root cause: heavy recoil fragments (the only secondaries that deposit meaningfully) have very short range, so only reactions happening close to the sensitive volume can contribute -- a structural feature of the physics, not a fixable sampling gap. Combined with CUPID's liquid-drop nuclear model being fundamentally different from PANDA's Geant4 cascade physics (not two approximations of the same model), agreement within an order of magnitude with matching curve shape is judged a genuine validation by the standard used for cross-code nuclear-model comparisons elsewhere in the field, not an open defect. See Documentation/PANDA_MASTER_DESIGN Known Limitations for the full before/after.

• Heavy-ion test energy should stay realistic (>~1 GeV for Au-mass ions) (found 2026-07-22, mechanism resolved 2026-07-22): 200 MeV Au197 (specific energy ~1 MeV/nucleon) produces an anomalously narrow, near-delta-function differential charge spectrum (relative std 2.96%) rather than genuine Vavilov/Landau straggling. Confirmed by rerunning the identical geometry at 3000 MeV (~15 MeV/nucleon, within real cyclotron heavy-ion SEE test-beam ranges): the spectrum becomes a properly-spread peak (relative std 8.30%), well clear of the max-possible-charge edge. Separate from the mean-LET model swap in v1.7.10/v1.7.11. The mechanism is ordinary Landau/Vavilov collision-counting straggling (Weeden-Wright et al., "Effects of Stopping Ions and LET Fluctuations on Soft Error Rate Prediction," IEEE TNS, 2015 -- from the MRED/CREME-MC developers themselves), not charge-state fluctuation (considered and found unsupported in that source). The specific magnitude is explained by the Vavilov shape parameter kappa ~ z^2/beta^4 (not dE/dx): a ~215x swing in kappa between 200 MeV and 3000 MeV Au197, driven by velocity alone, correctly predicts the observed ~2.8x change in relative spread -- see Documentation/PANDA_VALIDATION_SUMMARY.md Section 2 for the full derivation. Not a PANDA-specific bug -- just a reminder to pick physically-achievable test energies.

• CREME-MC thin-film direct-ionization shoulder, magnitude (found 2026-07-23/24, called good enough 2026-07-25): a 200 MeV proton through a 0.5um sensitive Si / 0.5um SiO2 dead layer stack showed PANDA's cross section running up to ~75x higher than CREME-MC's in a narrow 0.1-0.5 fC window, traced to genuine, correctly-functioning delta-ray transport (83% of shoulder events carry an explicit knock-on-electron hit) -- not a PANDA defect, but the cross-code magnitude gap itself had no confirmed explanation after ruling out differing EM-model choice and production-cut granularity (both checked directly, both wrong). Follow-up at an even smaller 65nm-node-representative footprint (0.25x0.25x0.5um Si, 2026-07-25) tested MRED's "Secondary Electron Computation Mode: detailed" directly -- it moved the wrong direction, fixing a large excess specific to CREME-MC's own default mode rather than closing the gap toward PANDA. A units error in that same comparison (CREME-MC's .agr export x-axis is energy-equivalent MeV, not fC) was found and corrected, revealing a much more modest, bounded ~1.3-6x gap across the shoulder and recoil tail, not orders of magnitude. A 10x-finer production-cut rerun confirmed PANDA's own numerics are fully converged at this geometry, and a direct measurement from PANDA's own recoil_hits.csv supports the leading remaining explanation (a straggling-model difference): the dominant heavy recoil species have their accumulated path length cut ~4-5x at this thin geometry, pushing them into a smaller-kappa, more model-sensitive straggling regime, while lighter species are unaffected. The two remaining candidates (MRED's exact nuclear-reaction model configuration and cross-section libraries) are Vanderbilt-proprietary and untestable. Residual gap treated as intrinsic PANDA/MRED model-family variation, the same category as the McNulty and Ni58/I127 entries above -- not an open PANDA defect. See Documentation/PANDA_VALIDATION_SUMMARY.md Section 1.5 (full writeup) and Section 2 (production-cut convergence row).

• Ni58/I127 ICRU73-vs-SRIM gaps are genuine model variation, not a bug (confirmed 2026-07-24): fixing Au197's Fe-scaling issue (this release, see above) required directly testing whether the same fix helps Ni58 (Z=28) and I127 (Z=53) too -- it doesn't. Both stayed on Fe-scaling deliberately after native data was confirmed to make Ni58 worse (+14.7% -> +17.8% vs SRIM) and flip I127's sign at similar magnitude. Their residual SRIM gaps (Ni58 ~14.7%, I127 ~6.4%, see Documentation/PANDA_VALIDATION_SUMMARY.md Section 1.4) are ordinary Geant4-EM-vs-SRIM stopping-power model variation for ions reasonably close to Fe in Z, not something a scaling-table fix can close further.

Build

mkdir -p build && cd build
cmake ..
make -j$(nproc)

Run

From the project root (not from inside build/):

./build/PANDA run.mac
python3 PANDA_Analyze.py

Linting

ruff (Python) and clang-tidy (C++) checks live in scripts/pre-commit-checks.sh. Git hooks aren't tracked by version control, so each clone needs a one-time setup:

ln -sf ../../scripts/pre-commit-checks.sh .git/hooks/pre-commit
python3 -m venv .lint-venv && .lint-venv/bin/pip install ruff

clang-tidy also needs build/compile_commands.json, generated by building with cmake -DCMAKE_EXPORT_COMPILE_COMMANDS=ON ... The hook blocks commits on ruff findings; clang-tidy findings are printed but never block.

See Documentation/PANDA_MASTER_DESIGN for the full design philosophy.