Stellarator Data Transformation via CycleGAN

This project aimed at transforming non-good stellarator configuration data into “good” ones that satisfy our design criteria. The overall approach employs a CycleGAN model to adjust the distributions of key parameters, thereby generating more useful stellarator configurations (referred to as Gstels and, when meeting an additional plot criterion, as XGstels).

Overview

• Problem Statement:
  Stellarator configuration data is divided into many subsets (from six different folders). Configurations that meet a set of criteria (listed in the following table) are labeled as Gstels. Within these, configurations that can also be plotted using r = 0.1 are marked as XGstels. The goal is to transform many non-good data points into ones that satisfy both checks.

  

• Approach:
  A CycleGAN model is used to learn the mapping from non-good to good data, particularly focusing on transforming parameter distributions such as the etabar distribution. The model exploits the adversarial interplay between a generator and a discriminator to gradually reshape the data distribution.

---

CycleGAN Model Training

Transformation of Distributions

• Initial State:
  The original etabar distribution for non-good configurations resembles a normal distribution.

• Target Distribution:
  The desired distribution (from Gstels) is bimodal with two peaks. Early in training, the generator produces outputs that still follow a normal distribution.

• Evolution Through Adversarial Training:
  1. Generator Version 1 (V1):
     Initially produces data following a normal distribution.
  2. Discriminator V1:
     Learns from the target bimodal distribution, identifying discrepancies (e.g., the presence of two peaks).
  3. Generator V2:
     Adjusts to produce distributions with two peaks but still may require improvements in peak height ratios.
  4. Discriminator V2:
     Further refines the criteria by learning that the left peak should be taller than the right.
  5. Generator Version 3 (V3):
     Finally evolves to produce the expected distribution where the left peak is more pronounced than the right.

---

Training Process

• Infinite Loop of Epochs:
  An infinite while loop is defined during training, where in each epoch the model is updated.

• Periodic Evaluation:
  Every 1000 epochs, the current model is used to transform 30,000 non-good data points. The proportion of data passing both checks is computed.

• Convergence Criteria:
  The training loop terminates when there is no increase in the pass rate for 10 consecutive evaluations.

• Final Testing:
  An additional one million non-good stellarator data points are processed, resulting in approximately 26% of the data being successfully transformed into XGstels.

---

Evaluation & Observations

Performance Analysis

• Evaluation Metric:
  The evaluation is based on hard binary criteria (e.g., whether an axis length is greater than zero), which makes the metric non-continuous and may contribute to fluctuations in the evaluation curve.

• Plot Analysis:
  The change in the proportion of data passing both checks during training exhibits a curve that is not as smooth as expected. Three possible reasons for this behavior are:

  1. Binary Evaluation Criteria:
     The use of a hard 1/0 pass/fail system results in a non-smooth metric.
  2. Model Complexity:
     The relative simplicity of the current model may limit its performance.
  3. Training Dynamics:
     Imbalances where either the generator becomes too strong or the discriminator lags behind could also affect performance.

  
  Figure: Training curve indicating the pass rate over different epochs.

Data Distribution Shifts

• Pre- and Post-Transformation Comparison:
  A comparison between the data that passed both checks before and after transformation reveals that the model effectively shifts the distribution toward the target profile.

  
  Figure: Distribution of data before (blue) and after (orange) transformation.

• Comparison with Real XGstels:
  Distributions of transformed data that passed both checks were nearly identical to those of the actual XGstels, confirming the efficacy of the transformation process.

Folder-specific Performance

• Uneven Model Performance Across Folders:
  The data is sourced from six folders. Testing revealed that the model performs differently:

  • Better Performance:
    In the last three folders.
  • Suboptimal Performance:
    In the first three folders.

  

This suggests that the quality or characteristics of the data across folders may vary, influencing the transformation effectiveness.

---

How to Apply the Code

1. Environment Setup

Ensure that you have installed the following Python libraries:

• argparse
• torch
• pandas
• numpy
• matplotlib
• scikit-learn

Make sure the project's dependencies (model and qsc) are available in your project directory.

2. Data Preparation

• Place your input CSV file (e.g., new_data.csv) containing configuration data into the data/ folder.
• Ensure that the scaler parameter files (scaler_X_mean.npy, scaler_X_std.npy, scaler_Y_mean.npy, scaler_Y_std.npy) exist.
• Verify that the trained generator model file G_A2B_best_model.pth is located in the project root or update the --model_path parameter accordingly.

3. Running the Code

The code performs data generation and (optionally) additional filtering. It consists of the following phases:

Filtering Phase (Optional):

1. Basic Filtering:
   Uses Qsc evaluation to filter out configurations that do not meet pre-defined criteria.

2. Fourier-Based Filtering:
   Further processes the data by converting them into Fourier coefficients and applies additional checks.

Example Command:

python your_script.py --input_csv data/new_data.csv \
                      --generated_csv data/generated_output.csv \
                      --model_path G_A2B_best_model.pth \
                      --filter True \
                      --basic_out_csv data/generated_output_basic.csv \
                      --fourier_out_csv data/generated_output_fourier.csv