discover
DISCOVER: Symbolic Regression for Science
DISCOVER is a modern, Python-native implementation of the Data-Informed Symbolic Combination of Operators for Variable Equation Regression method — a powerful machine learning framework for discovering predictive, interpretable models and descriptors from large feature spaces.
This implementation is designed for ease of use, extensibility, and performance, incorporating modern best practices and algorithmic enhancements. This tool is ideal for researchers and data scientists in materials science, chemistry, and other scientific domains who need to find physically meaningful relationships in their data.
Key Features
- Iterative Feature Generation: Efficiently builds a high-quality feature space by iteratively generating and screening features, avoiding the combinatorial explosion of traditional methods.
- Multiple Search Strategies:
- Greedy & OMP: Fast, iterative approaches for building models.
- Brute-Force: Exhaustively finds the best model for low-dimensional problems.
- SISSO++: A highly efficient, breadth-first search using QR decomposition for superior performance.
- Heuristics (
rmhc,sa): Advanced search methods like Random Mutation Hill Climbing and Simulated Annealing to explore vast search spaces. - MIQP (Gurobi): Provably finds the optimal L0-norm model for regression (requires Gurobi license).
- Diverse Machine Learning Tasks:
- Regression: Linear (Ridge), Robust (Huber), and Multi-Task models.
- Classification: Logistic Regression, Support Vector Machines (SVM), and a unique Convex Hull-based classifier for geometrically separable data.
- GPU Acceleration: Leverages NVIDIA (CUDA via CuPy) and Apple Silicon (MPS via PyTorch) for significant speedups in feature generation, screening, and model fitting.
- Unit-Aware Feature Engineering: Uses the
pintlibrary to prevent the creation of physically nonsensical features by enforcing unit consistency during feature generation. - Advanced Analytics & Visualization: Includes a suite of built-in plotting methods for model interpretation, such as parity plots, classification decision boundaries, feature importance, and partial dependence plots.
Navigation
- New to DISCOVER?
- Start with the Step-by-Step Tutorial to run your first analysis.
- Want to accomplish a specific task?
- Check out the How-To Guides for advanced configurations and use cases.
- Want to contribute to the project?
- Read the Contribution Guidelines to learn how to add new features and operators.
- Would you like to learn fundamentals?
- Read the Fundamentals to learn how to add new features and operators.
- If you find this article useful, please consider it citing
- DOI: