Project Summary
Overview: Achieving ignition in nuclear fusion requires precise control over plasma temperature, density, and confinement time. This project simulates these conditions and utilizes machine learning to maintain stability against disruptions.
Using a Kaggle-sourced dataset of 100,000 simulated fusion experiments, I developed a two-stage machine learning pipeline to model the complex, non-linear dynamics of nuclear fusion. The system processes 15 key parameters—including magnetic field strength, fuel density, and temperature—to predict two critical outcomes: the likelihood of ignition (Classification) and the resulting power output (Regression).
The approach employs a Decision Tree Classifier to first identify viable ignition cases, which are then passed to a Decision Tree Regressor to forecast energy generation. This methodology allows for the rapid evaluation of fusion configurations, significantly reducing the reliance on expensive and time-consuming physical trial-and-error experiments.
Key Outcomes:
- High-Fidelity Power Prediction: The Decision Tree Regressor achieved a Coefficient of Determination (R²) of 0.999 and a Mean Squared Error (MSE) of 7.885e-6, enabling near-perfect prediction of power output for ignited plasmas.
- Superior Baseline Performance: The model vastly outperformed a Linear Regression baseline (MSE ≈ 823, R² ≈ 0), validating the necessity of non-linear models for capturing fusion dynamics.
- Optimization of Critical Parameters: Successfully analyzed and weighted 15 feature inputs across different reactor configurations (Tokamak, Stellarator, Reversed Field Pinch) to isolate the variables most critical for sustaining ignition.
View Code on Colab