# Figure 1. Delivery storage and acquisition assumptions bias inferred half-reaction rates

**Original SemiAgora synthetic numerical study. No measured ALD data or calibrated physical prediction.**

(a) Instantaneous inlet concentration for the previously unseen short-pulse history, comparing a single-pole inlet, a finite-buffer inlet driven by the same command, and a buffer case matched in total delivered amount. Curves connect 193 saved solver-query samples; the panel displays dimensionless time 0–4 without smoothing. The same-command buffer changes both delivered amount and temporal profile. In the equal-dose construction, command amplitude and buffer capacity/initial inventory were scaled together before chemistry; this is a mathematical dose-matched comparison, not a changed valve recipe on identical hardware.

(b) Initial surface-state conversion, defined as one minus the initial surface population, for the same three conditions through dimensionless time 12. The model includes detachment; conversion is not deposited thickness, growth per cycle, complete hydrolysis, or film conformality. Terminal conversion is 0.5542413 for the single pole, 0.7356208 for the same-command buffer, and 0.5545358 for the equal-dose buffer. The equal-dose and single-pole responses nearly overlap; their small difference has not been visually amplified.

(c) Reaction-rate multipliers from the four training-only qualified selections. Row names identify synthetic truth and fitted observation operator. Every fitted delivery model is a single-pole model. Averaged denotes the correct causal acquisition average; point denotes the intentionally misspecified point-sampling ablation. The true multipliers are 1. The buffer-truth fits produce approximately 0.7813 for the first rate and 3 for the second; the dotted line at 3 identifies the imposed upper bound reached by both orange buffer bars. A bound-active stationary solution is not an unconstrained physical rate estimate. The buffer/averaged fit also reaches the lower transit-time bound; that unplotted parameter is retained in the inference CSV and source packet.

(d) The maximum scaled combined RMSE across two held-out pulse histories for each selected fit: 9.83446 × 10⁻¹², 0.00648024, 0.11111629, and 0.11164840, respectively. Horizontal bars start at zero on a common linear axis. The positive near-zero recovery error keeps its actual width; no minimum visible bar length is imposed. Its precise value and the label 'Noiseless recovery control' appear beside the baseline. Channel scales are fixed numerical weights, not noise standard deviations; RMSE is not a percentage. The near-zero matched-model result is a deliberately noiseless synthetic recovery control, not sensor precision. No confidence intervals were estimated and no error bars were invented.

The frozen ALD016 study retained 909 chemical attempts. All 909 raw records passed independent balance and ledger audits; native and separate extent/DOP853 plus refined-reference solutions underwent 69 pairwise checks across 23 control/truth/prediction trios. Training selection was preserved before held-out chemical truth generation. Numerical qualification is distinct from physical calibration. Fixed sensor and wall assumptions, a well-mixed water half-reaction, dimensionless inputs, and absence of measured calibration limit interpretation. Full ALD cycles, spatial transport, film properties, leakage, breakdown, and journal readiness are not established.

## Source and method provenance

Canonical research packet: `07_assets/process_lab/experiments/SA-PROC-ALD-TRANSIENT-002/research/ald-delivery-buffer-comparison-20260918/`.

Panels (a,b): positional columns 0, 1, and 2 of three `analysis/*__1__samples.csv` files. The original header repeats `inlet_concentration`: position 1 is instantaneous inlet and position 9 is the acquisition-averaged sensor channel. Exported CSVs disambiguate these names without changing any numerical value.

Panels (c,d): `analysis/summary.json` and `groups[*]`. Exact input hashes and panel-level mappings accompany the figure. Lines were replotted from existing saved samples; no new solver run, fitting, smoothing, error model, or scientific conclusion was added.

Background motivation: Ahmido et al. (2016), in situ water-vapor delivery metrology, [NIST publication record](https://www.nist.gov/publications/situ-metrology-characterize-water-vapor-delivery-during-atomic-layer-deposition). The source motivates independent inlet measurement. No publication figures, measured traces, tables, or instrument coefficients were copied into this original synthetic model.

## 한국어 설명

**ALD 공급 이력과 센서의 관측 방식이 반응속도 추정에 미치는 영향을 정리한 4패널 Figure입니다.** 기존에 계산하고 독립 검증한 SemiAgora의 합성 수치 연구를 다시 그렸습니다.

(a)는 공급 농도의 시간 변화, (b)는 같은 조건의 표면 초기 상태 전환율, (c)는 학습 조건에서 추정한 두 반응속도 배율, (d)는 보지 않았던 두 펄스 조건에서의 최대 오차입니다. 저장량이 더해지면 공급 총량도 달라지므로 그 효과를 시간 변화만으로 해석하지 않습니다. 총량을 맞춘 두 응답의 작은 차이는 확대 과장하지 않았습니다.

실험 측정 데이터가 아니며 모든 시간·입력은 무차원입니다. 전환율을 막 두께나 성장률로 해석할 수 없습니다. (c)의 점선은 참값 1과 상한 3을 표시하고, (d)는 선형 축의 막대 길이와 실제 수치로 오차를 비교합니다. 거의 0인 기준 사례의 막대를 인위적으로 늘리지 않았으며, 신뢰구간은 산출하지 않았습니다.

## Deliverables and reproduction

- `figure-ald.pdf`: 180 × 150 mm vector figure with embedded text fonts.
- `figure-ald.svg`: editable vector figure retaining SVG text.
- `figure-ald.png`: native 400 dpi figure.
- `figure-ald-preview.png`: 1200 px website preview.
- `public_redraw.py`: redraw with only the four `plot-data-*.csv` exports: `python public_redraw.py --data . --out ./redrawn`.
- `plot.py`: identical portable redraw entry point; only the four exported CSVs are needed.
- `source_manifest.json`, `panel_source_map.json`, `verification.json`: audit trail.

Only bounded plot values are intended for the portfolio download. Full private raw NPZ files and the 397 MB research archive are not included.


## Presentation revision, 2026-09-30

Panels (a,b) retain the same saved traces and axis limits. Panel (c) uses grouped horizontal bars with zero baselines and separate true-value/imposed-bound reference lines. Bar-end labels are rounded for reading; all original floating-point values remain in the unchanged CSV. Panel (d) changes from logarithmic isolated points to zero-based linear bars and explicit values. No scientific value, case, selection, uncertainty model, solver or claim was changed.