naman-00/Energy_Landscape_Wafer_Variation_Benchmark

---

Energy Landscape Wafer Variation Benchmark (ELWVB)

Introduction and Overview

The Energy Landscape Wafer Variation Benchmark (ELWVB) is a synthesized dataset designed to serve as a comprehensive benchmark for machine learning models applied to semiconductor manufacturing yield prediction. It tackles the crucial challenge of modeling the spatial variability of critical process parameters—specifically Critical Dimension (CD) and Film Thickness—and their ultimate impact on wafer failure probability and manufacturing yield.

The dataset simulates a realistic scenario where process variations across the wafer surface lead to non-uniform defect probabilities. It provides a rich set of features including normalized spatial coordinates, physical parameter values, a calculated "Energy Landscape" value, and two critical target variables: the ground truth probability of defect and the simulated binary failure outcome.

Core Dataset Components

ELWVB provides a comprehensive view of the simulation with both tabular data and key visualizations:

• wafer_site_data.csv (Tabular Data): This file contains 40,000 data points, representing individual sites on a $200 \times 200$ grid. Each site is characterized by 13 detailed features, enabling site-specific analysis and model training.
• true_p_defect_map.png (Ground Truth Visualization): A map illustrating the actual, underlying probability of defect across the wafer.
• Modeled Output Maps (e.g., mle_p_defect_map.png, mc_p95_map.png): Visualizations of the defect probabilities predicted by established modeling techniques (Maximum Likelihood Estimation and Monte Carlo simulations), providing a baseline for comparison.
• simulated_failure_map.png (Hotspot Visualization): A visual representation of the final, binary failure outcomes, clearly indicating the high-risk "hotspot" regions.
• results_summary.json (Benchmark Metrics): A summary file providing true parameters used in the simulation and baseline performance metrics (e.g., AUC) achieved by simple benchmark models like Logistic Regression.

Research and Application Focus

The ELWVB dataset is ideal for researchers and practitioners in data science and semiconductor engineering who are focused on:

1. Spatial Machine Learning: Developing advanced models (e.g., those using convolutional or graph neural networks) that leverage the spatial arrangement of the wafer sites to improve prediction accuracy over non-spatial models.
2. Yield Modeling and Optimization: Training robust classification models to predict the binary failure status, which is essential for yield management and quality control.
3. Probabilistic Forecasting: Benchmarking regression models for predicting the continuous p_defect (probability of defect), providing crucial insights into manufacturing margins and risk assessment.
4. Feature Importance Studies: Investigating the relative impact of physical parameters (cd, thickness) and spatial location on the final defect rate.

By providing both the ground truth defect probability and actual failure outcomes, ELWVB supports both continuous probability modeling and binary classification tasks, offering a challenging and highly relevant scenario for industrial AI.

Dataset Structure

The dataset primarily consists of one CSV file (wafer_site_data.csv), a JSON file with model summary, and four image files for visualization.

File Name	Format	Description
wafer_site_data.csv	CSV	Contains the site-specific parameters, true probability of defect, and the binary failure outcome for 40,000 sites.
results_summary.json	JSON	Summary of model parameters and performance metrics (e.g., AUC).
true_p_defect_map.png	PNG	Visualization of the true probability of defect across the wafer.
mle_p_defect_map.png	PNG	Visualization of the MLE (Maximum Likelihood Estimation) predicted probability of defect.
mc_p95_map.png	PNG	Visualization of the 95th percentile probability from Monte Carlo simulation.
simulated_failure_map.png	PNG	Visualization of the simulated binary failure outcomes (hotspots).

Dataset Fields (wafer_site_data.csv)

The main data file, wafer_site_data.csv, includes 13 columns for each wafer site:

Column Name	Type	Description
x, y	float	Normalized Cartesian coordinates on the wafer.
r	float	Normalized radial distance from the wafer center.
cd	float	Critical Dimension variation.
thickness	float	Film thickness variation.
hotspot	float	A value indicating a defect-prone region.
hotspot_mask	int	Binary mask for the hotspot region.
E_true	float	True Energy Landscape value (related to the defect probability).
p_defect	float	Target: The true probability of defect at the site.
failure	int	Target: Binary outcome (0 = no failure, 1 = failure).
p_mle	float	Probability of defect predicted using Maximum Likelihood Estimation.
p50_mc	float	50th percentile (median) probability from Monte Carlo simulations.
p95_mc	float	95th percentile probability from Monte Carlo simulations.

Dataset Information

• Total Number of Sites: 40,000
• Grid Size: $200 \times 200$ grid
• Modeling Task: Classification (predicting failure from wafer parameters).
• Benchmark Metrics (as per results_summary.json):
  • AUC (Area Under the Curve) for MLE Model: 0.5841
  • AUC (Area Under the Curve) for Logistic Regression Model: 0.5842

Potential Tasks

This dataset is suitable for:

1. Wafer Yield Prediction: Training models to predict the binary failure outcome based on process parameters (cd, thickness, x, y, etc.).
2. Defect Probability Modeling: Regression tasks to model the true probability of defect (p_defect) or the modeled probabilities (p_mle, p50_mc, p95_mc).
3. Spatial Analysis: Studying the effect of spatial variation (i.e., x, y, r, cd, thickness) on defect probability.