DataSet for "Modeling Laser Direct Writing Grayscale Exposure Process and Predicting Photoresist 3D Morphology Based on UNet/CGAN Networks”

This dataset is constructed to support the training, validation, and testing of UNet and Conditional Generative Adversarial Network (CGAN) models for predicting the 3D morphology of photoresist in the laser direct writing (LDW) grayscale exposure process. It directly corresponds to the experimental workflow and model development in the manuscript “Modeling Laser Direct Writing Grayscale Exposure Process and Predicting Photoresist 3D Morphology Based on UNet/CGAN Networks”, and provides a standardized input-output mapping for data-driven modeling of LDW grayscale exposure.  1. Data Acquisition Background & Experimental Basis  All data were collected via controlled experiments following a standardized LDW grayscale exposure workflow (detailed in the manuscript’s “Methods” section), with non-exposure parameters (spin-coating, pre-baking, developing) strictly fixed to eliminate covariate interference:  - Substrate & Photoresist Preparation: Clean 3-inch fused quartz wafers were uniformly coated with AZ4620 photoresist (3000 rpm/30 s), then pre-baked on a hot plate (90℃/10 min) to achieve a final thickness of 9.0 ± 0.5 μm.  - Grayscale Mask Generation: Multi-level grayscale masks (2000×2000 pixels, grayscale range: 0–255) covering different periods (e.g., 10 μm, 20 μm) and step numbers were generated via a custom Python script.  - Parametric Exposure: Exposure was performed on a PICOMASTER ATE-200 LDW platform with variable power (250–450 mJ/cm², 50 mW step size) and fixed key parameters (defocus: -1.5 V; step resolution: 100 nm; spot diameter: 230 nm; scanning speed: 200 mm/s).  - Developing & Morphology Measurement: Exposed wafers were developed in AZ300MIF developer (23℃/120 s, nitrogen-dried), and 3D morphology point cloud data were acquired using an Olympus LEXT OLS5100 laser confocal microscope (50× objective, sub-micron measurement accuracy).   2. Data Composition & Scale  The dataset consists of input-output data pairs where “input” represents LDW process conditions and “output” represents the corresponding photoresist 3D morphology:  - Raw Valid Samples: 91 sets of high-quality samples, each including:   - Input: 1×2000×2000 single-channel grayscale mask image (0–255 grayscale values, pixel size matching actual exposure resolution) + 1×5 process parameter vector (power, defocus, step resolution, spot diameter, scanning speed; only power is variable, others fixed as per experimental design).   - Output: 3D morphology point cloud of the exposed photoresist (X/Y/Z coordinates, Z unit: μm) then transformed to grayscale image.  - Data Augmentation: To expand sample diversity and avoid model overfitting, each 2000×2000 mask-image/morphology pair was randomly cropped into 1000 sets of 256×256 sub-images. This resulted in 91,000 final data pairs for model training.  3. Data Preprocessing  Raw data underwent rigorous preprocessing to ensure accuracy and consistency (critical for model convergence):  1. Point Cloud Denoising: Median filtering was applied to raw 3D morphology point clouds to remove high-frequency noise from the laser confocal measurement.  2. Tilt Correction: Three-point leveling was used to eliminate substrate tilt-induced morphology deviations, ensuring the Z-axis height reflects true photoresist thickness variation.  3. Image Registration: Sub-pixel level registration between grayscale masks and morphology images was performed to correct rotational/translational errors, ensuring spatial alignment of input and output.  4. Data Partitioning  The 91,000 data pairs were split into three non-overlapping subsets for model training, validation, and testing (consistent with the manuscript’s model evaluation design):  - Training set: 63,700 pairs (70%) – used for model parameter optimization.  - Validation set: 18,200 pairs (20%) – used for hyperparameter tuning and overfitting monitoring.  - Testing set: 9,100 pairs (10%) – used for unbiased evaluation of model generalization (e.g., UNet’s ~1% mean absolute error and CGAN’s high-frequency information capture, as reported in the manuscript).  5. Data Format & Availability  - Format: Input grayscale images (PNG), process parameters (NPY), and output 3D morphology image (PNG).  - Availability: As stated in the manuscript’s “Data availability” section, the underlying data are not publicly available at this time but can be provided by the corresponding author (Guohan Gao, [e-mail]) upon reasonable request from editors, reviewers, or researchers for reproducibility and further study.  This dataset’s value lies in its standardization (fixed non-exposure parameters) and representativeness (covering diverse grayscale masks and exposure powers), making it a reliable basis for validating data-driven LDW process modeling. It directly enables the key findings of the manuscript (e.g., UNet’s fast convergence and CGAN’s high-frequency prediction) and supports future research on LA-DOE fabrication and 3D microstructure reverse design.