aw-write-experiment — Experiment Section Writer

Purpose

Write the Experiment section of an academic paper (Phase 4 of IMRAD) by generating independent paragraph files that are later merged into a complete chapter by aw-execute.

This skill is a section-writing subagent called by the Wave Executor during Phase 2. It receives a specific task (e.g., "Write 4.1 Dataset Description") and outputs a single paragraph .tex file.

When to Trigger

• aw-execute wave executor calls this skill with a specific task
• User runs /aw-write-experiment directly
• Orchestrator delegates during GSDAW pipeline execution

Inputs

Input	Source	Description
Task description	Wave executor (objective field)	Which paragraph to write (e.g., "4.1 Dataset Description")
research-brief.json	.planning/research-brief.json	Author intent, novelty claims
methodology.md	.planning/methodology.md	Full experiment design — datasets, baselines, metrics, ablation
literature.md	.planning/literature.md	Related work context (for baseline positioning)
Elsevier template	templates/elsevier/	LaTeX format reference

Outputs

Paragraph files written to sections/experiment/ with naming convention {task-id}.tex:

File	Task ID	Content
sections/experiment/4-1-datasets.tex	4.1	Dataset description with dataset table
sections/experiment/4-2-baselines.tex	4.2	Baseline method configurations
sections/experiment/4-3-metrics.tex	4.3	Evaluation metric definitions
sections/experiment/4-4-ablation.tex	4.4	Ablation study setup

Dataset Table Format

\begin{table}[htbp]
\centering
\caption{Experimental Datasets}
\label{tab:datasets}
\begin{tabular}{llcccc}
    \toprule
    Test Set & Source & Material & Defects & Signal Count & SNR Range \\
    \midrule
    Sim-train & FEM simulation & Al + CFRP & FBH, SDH, delamination & 10,000 pairs & -10 to +20 dB \\
    Sim-val & FEM simulation & Al + CFRP & FBH, SDH, delamination & 1,000 pairs & -10 to +20 dB \\
    Sim-test & Held-out FEM & Al + CFRP & FBH, SDH, delamination & 500 pairs & -10 to +20 dB \\
    Al-exp & Experimental & Al 2024-T3 & FBH (3, 5, 8 mm depth) & 50 A-scans & -5 to +5 dB \\
    CFRP-exp & Experimental & CFRP laminate & Delamination, fiber breakage & 50 A-scans & -8 to +2 dB \\
    \bottomrule
\end{tabular}
\end{table}

Baseline Configuration Table

\begin{table}[htbp]
\centering
\caption{Baseline Methods}
\label{tab:baselines}
\begin{tabular}{lll}
    \toprule
    Method & Configuration & Implementation \\
    \midrule
    Wiener filter & Adaptive, 5x5 local neighborhood & SciPy signal.wiener \\
    DWT denoising & Daubechies-4, 4-level decomp., soft thresholding & PyWavelets wtmm.denoise \\
    BM3D & Block matching 3D, 8x8x8 blocks, sigma\_est=auto & bm3d package \\
    Sparse coding & Overcomplete DCT (256 atoms), OMP reconstruction & SPAMS toolbox \\
    Bandpass filter & Butterworth 4th order, passband 1--10 MHz & SciPy signal.butterworth \\
    \bottomrule
\end{tabular}
\end{table}

Evaluation Metrics Definitions

\subsection{Evaluation Metrics}

We evaluate denoising performance using five metrics.

\textbf{SNR Improvement} measures the gain in signal-to-noise ratio:
\[
\Delta\text{SNR} = 20 \log_{10}(\text{RMS}_{denoised}) - 20 \log_{10}(\text{RMS}_{noisy})
\]
where RMS denotes the root mean square of the signal amplitude.

\textbf{Mean Squared Error} quantifies the squared difference to the clean reference:
\[
\text{MSE} = \frac{1}{N}\|y_{clean} - y_{denoised}\|^2
\]

\textbf{Lin's Concordance Correlation Coefficient} (CCC) measures waveform morphology preservation independent of scale:
\[
\text{CCC} = \frac{2\text{Cov}(y_{true}, y_{pred})}{\text{Var}(y_{true}) + \text{Var}(y_{pred}) + (\bar{y}_{true} - \bar{y}_{pred})^2}
\]

\textbf{Waveform Similarity Index} (WSI) measures zero-lag normalized cross-correlation between denoised and clean waveforms.

\textbf{F1-score for Defect Detection} evaluates the probability of detecting defect echoes in denoised A-scans via peak detection, at SNR = -5 dB.

Workflow

aw-write-experiment invoked with task "4.1 Dataset Description"
    │
    ▼
Read: .planning/methodology.md (Experiment Design section)
    │
    ▼
Read: .planning/research-brief.json (novelty)
    │
    ▼
Read: templates/elsevier/ (LaTeX format)
    │
    ▼
Write: sections/experiment/4-1-datasets.tex
    │
    ▼
Return completion with word count and preview

Step-by-Step Execution

Step 1: Read Inputs

Read .planning/methodology.md Section 3 (Experiment Design):

• 3.1 Datasets (training, validation, test sets)
• 3.2 Baseline Methods (configurations)
• 3.3 Evaluation Metrics (definitions)
• 3.4 Ablation Studies (variants)
• 3.5 Statistical Analysis (seed, significance testing)

Also read .planning/research-brief.json for novelty claims relevant to experimental validation.

Step 2: Extract Content per Task

4-1-datasets.tex — Dataset Description:

• Training set: 10,000 paired noisy-clean A-scan signals from FEM
• Validation: 1,000 held-out pairs
• Test sets: Sim-test (500), Al-exp (50), CFRP-exp (50)
• Materials: Aluminum 2024-T3 (60%), CFRP laminate (40%)
• Defect types: FBH, SDH, delamination, fiber breakage
• FEM parameters: Aluminum $v_L=6320$ m/s, CFRP $v_L=3000$ m/s
• Signal specs: 2048 samples at 100 MHz, SNR -10 to +20 dB

4-2-baselines.tex — Baseline Methods:

• Wiener filter (adaptive, 5x5 neighborhood)
• DWT denoising (Daubechies-4, 4-level, soft thresholding)
• BM3D (block matching 3D, 8x8x8 blocks)
• Sparse coding (overcomplete DCT, 256 atoms, OMP)
• Bandpass filter (Butterworth 4th order, 1-10 MHz)
• Include justification for why these are fair comparisons

4-3-metrics.tex — Evaluation Metrics:

• SNR improvement: $\Delta\text{SNR}$ target 8-12 dB
• MSE: target < 1.5 x 10^-3
• CCC: target > 0.95
• WSI: target > 0.90
• F1-score: target > 0.85 at -5 dB
• POD: Probability of Detection, slope > 2.5 dB^-1

4-4-ablation.tex — Ablation Studies:

• A1: Loss function (MSE-only vs. MSE+CCC vs. Full)
• A2: Architecture depth (3-level vs. 4-level vs. 5-level)
• A3: Skip connections (no skip vs. concat vs. concat+attention)
• A4: Training SNR distribution (fixed +10 dB vs. uniform -10 to +20 dB)
• A5: Input window size (256 vs. 512 vs. 1024)
• A6: Attention gating (with vs. without)

Step 3: Write Paragraph File

Write the .tex file with:

1. \paragraph{Section Title} heading with label
2. Running text with technical detail
3. Tables using booktabs format
4. Metric definitions with inline math
5. Cross-references via \ref{tab:}

Step 4: Verify Output

• At least 150 words per paragraph
• Academic register
• Elsevier citation format
• Tables use booktabs
• No hardcoded numbers in \ref{}
• No TODO/FIXME placeholders

Step 5: Report Completion

Paragraph 4.1 (Dataset Description) written.
Word count: 312
File: sections/experiment/4-1-datasets.tex
Preview: "The training dataset consists of 10,000 paired noisy-clean A-scan signals..."

Elsevier LaTeX Conventions

Same as aw-write-methodology:

• \documentclass[review]{elsarticle}
• \usepackage{booktabs} for tables
• \cite{key} for numbered citations
• \ref{tab:}, \ref{fig:}, \ref{eq:} for cross-references

Error Handling

Missing methodology.md

错误：未找到方法论设计文件 (.planning/methodology.md)。

请先运行 /aw-methodology 生成实验设计，
或确认方法论已通过 Discuss #2 审批。

Incomplete Experiment Design

If a section is missing content:

• Write available content with gap note
• Report in completion message

File Locations

manuscripts/[paper-name]/
├── .planning/
│   ├── research-brief.json
│   ├── methodology.md         ← Primary input
│   └── literature.md
├── templates/elsevier/
└── sections/
    └── experiment/
        ├── 4-1-datasets.tex       ← Output
        ├── 4-2-baselines.tex      ← Output
        ├── 4-3-metrics.tex        ← Output
        └── 4-4-ablation.tex       ← Output

Integration Points

Connection	Agent/File	Direction
Called by	aw-execute (Wave Executor)	Input: task
Feeds into	aw-execute (Phase Merger)	Output: paragraphs
Reads	.planning/methodology.md	Input
Later review	aw-review	After wave

Quality Gate Checklist

• At least 150 words in paragraph
• Academic register
• Elsevier citation format
• Dataset table uses booktabs
• Baseline table uses booktabs
• Metric definitions include formulas
• No hardcoded numbers in cross-references
• No TODO/FIXME placeholders
• File saved to correct path sections/experiment/{filename}.tex