# Particle Monte Carlo Feature Topography Model Card v1

## Identity

- Experiment: SA-PROC-FEATURE-PMC-007
- Version: 2026-08-27-v106
- Public execution class: precomputed browser replay only
- Model class: original deterministic stochastic 2D occupancy-grid teaching model
- Compared modes: directional RIE etch and sputter trench deposition

## Intended Use

Use the packet to inspect causal differences among:

- collimated and spread incident directions;
- ion-rich and neutral-rich primary populations;
- high and low reaction or sticking states;
- damped and active reflection or re-emission;
- suppressed and enabled secondary sputter transport;
- nominal aspect-ratio states 1, 2, and 4;
- assumption-only and owned-profile-slot-ready evidence postures.

The model is designed for mechanism literacy, run-card review, exact export, conservation checks, seeded repeatability, and planning of replacement evidence.

## Model Structure

The geometry is a 40 by 34 array of integer cells. A ray starts above a centered feature, takes quarter-cell steps along one of a finite set of declared direction vectors, and stops when it escapes, reacts, sticks, is absorbed without geometry change, or reaches the declared step cap.

The initial RIE state contains a hard-mask opening above a flat substrate. Removing the first exposed non-mask cell evolves a trench.

The initial deposition state contains a 12-cell-deep rectangular cavity. A sticking neutral fills the last vacuum cell before impact. An ion can resputter only an already deposited cell. A spawned secondary fragment can fill one vacuum cell as redeposition or leave the domain.

There is no hidden fit, optimization, external runtime call, imported source result, or website-side solver.

## Randomness And Reproducibility

The apparent randomness is completely deterministic:

- xorshift32 supplies the random stream;
- FNV-1a derives a nonzero 32-bit seed from the physical configuration;
- all direction and species choices use integer-weight tables;
- evidence posture does not participate in seed derivation;
- the complete state is advanced sequentially over 15 batches of 48 primary particles;
- progress state p00 is the untouched initial geometry.

Same seed and same physical configuration must reproduce the same geometry RLE, geometry hash, interaction ledger, and metrics at all 16 states.

## Interaction Logic

Normal alignment uses a bounded L1 direction projection. RIE ion removal favors source-facing incidence. RIE neutral removal receives a modest grazing contribution. PVD neutral sticking favors source-facing incidence. PVD ion resputter is allowed only on primary or secondary deposit cells. Secondary fragment sticking uses its own declared probability.

When a reaction or stick does not occur, the selected reflection state controls whether the entity continues. Reflections are either specular or selected from a finite diffuse half-space set. Bounce count is capped.

Every probability is dimensionless and internal. It is not a material property, sputter yield, sticking coefficient, reaction cross section, or calibration.

## Exact Ledgers

Entity conservation:

entities created = primary launched + secondary created = all termination counts.

Removed material:

removed cells = volatile chemical removals + immediate sputter outflow + secondary created.

Secondary material:

secondary created = secondary redeposited + secondary material outflow.

Grid mass:

final solid cells = initial solid cells + primary deposited cells + secondary redeposited cells - removed cells.

All four identities must close with integer error zero in every row.

## Geometry Metrics

- Etch mean and maximum depth are cell counts below the initial exposed surface.
- Lateral undercut is the maximum void extension beyond the original opening.
- Sidewall roughness is a guarded coefficient of variation of first-solid depths.
- Deposition fill fraction counts initially open cavity cells that become solid.
- Bottom and sidewall coverage count deposited cells in declared cavity bands.
- Accessible void fraction uses four-neighbor flood fill from the top opening.
- Pinch-off means remaining cavity void exists but no bottom-region cavity cell is reachable from the top.
- Bottom, sidewall, and top flux fractions are collision-zone shares, not physical fluxes.

## Evidence Interpretation

The evidence posture changes only labels, readiness score, replacement slots, and decision gates. It cannot change the generated physical state.

The maximum readiness score remains an evidence-planning field. No synthetic row is measured, calibrated, traceable, production-ready, or signoff-ready.

## Source And Software Separation

ViennaPS, TU Wien research pages, and the Klemenschits dissertation provide mechanism and architecture vocabulary. Synopsys provides a commercial scope benchmark. The SemiAgora implementation does not import those tools, source files, examples, equations, coefficients, geometries, or results.

ViennaPS 4.6.2 is GPL-3.0 and is restricted to an isolated local comparison. That comparison cannot supply values to this model and cannot be represented as public replacement evidence.

## Known Limitations

- Two spatial dimensions only.
- Fixed Cartesian cells and finite step rays.
- No level set, mesh refinement, sub-cell interface, curvature, surface diffusion, or continuous normal.
- No chamber, sheath, electric field, plasma chemistry, gas scattering, energy distribution, time-of-flight, temperature, stress, charging, damage, passivation coverage, or endpoint model.
- No material identity beyond integer role codes.
- One removed cell creates at most one secondary fragment.
- No atomistic mass, stoichiometry, density, porosity, microstructure, grain, defect, or composition model.
- No grid-convergence or particle-count-convergence claim in v1; those remain explicit replacement evidence.
- No prediction outside the exact internal comparison.

## Prohibited Uses

This is not a process recipe. No named-tool equivalence is claimed. No live server execution is allowed.

Do not use the packet to select or tune a process recipe, compare vendors, operate equipment, set pressure, flow, source power, bias, temperature, time, endpoint, maintenance, or safety action, predict a physical profile, release a lot, qualify a process, or claim equivalence with ViennaPS, Sentaurus Topography, or any other tool.

## Validation Plan

1. Parse the contract and confirm all axes, constants, and hard boundaries.
2. Independently reconstruct all 192 physical simulations without importing the generator.
3. Compare all 6,144 CSV rows and typed web rows field by field.
4. Recompute every geometry hash and RLE.
5. Require zero entity, removed-material, secondary-material, and grid-mass errors in every row.
6. Require evidence-paired rows to have identical physical fields.
7. Replay the same seed and require byte identity.
8. Run alternate-seed samples and require meaningful geometry variation without violating conservation.
9. Check source reachability and rights metadata without downloading publication or vendor content.
10. Run file, hash, secret-pattern, mojibake, and whitespace hygiene checks before any website edit.