Francis-Fan-create/SCaSML

ScaSML solver for high dimensional gradient dependent semilinear PDE

Simulation-Calibrated Scientific Machine Learning (SCaSML)

SCaSML is a novel approach for solving high-dimensional partial differential equations (PDEs) that combines the dimension insensitivity of neural networks with the transparent, unbiased, and physical estimation from simulation solvers.

Abstract

Addressing high-dimensional PDEs has long been challenging due to the 'curse of dimensionality'. Deep learning methods, while promising, often suffer from bias, non-transparency, and insufficient physics embedding. To address these, we propose Simulation Calibrated Scientific Machine Learning (SCaSML), integrating neural network's dimension insensitivity with transparent, unbiased and physical estimation from simulation solvers.

We train Physics-Informed Neural Networks (PINNs) to provide initial solution estimates, whose errors, governed by a new PDE, are corrected via Monte Carlo solvers using the Feynman-Kac and Elworthy-Bismut-Li formulae. We prove SCaSML's rate improvement for two state-of-the-art solvers, quadrature multilevel Picard and full-history multilevel Picard.

Numerical experiments on various high-dimensional PDEs, including the Hamilton-Jacobi-Bellman and Diffusion-Reaction equations, etc., confirm our theoretical results in terms of accuracy. This method advances fields like economics, finance, operations research, and physics by comprehensively considering all involved elements as agents, assets, and resources.

Installation

1. Clone the repository:

git clone https://github.com/francis-fan-create/scasml.git 
cd scasml

2. Create and activate the conda environment:

conda env create -f environment.yml
conda activate myenv

Usage

To set up a new SCaSML solver for specific equations, follow these steps:

1. Configure your equation in equations/equations.py using the Explicit_Solution_Example class as a guide.

2. In the models/ directory, create a new .py file and define your network structure, taking FNN.py as an example. Ensure to set the self.regularizer parameter.

3. In the optimizers/ directory, create a new .py file to customize your training process, following the Adam_LBFGS.py template.

4. To use test methods other than tests/NormalSphere.py, create a new one in the tests/ directory using the NormalSphere.py format.

5. Copy experiment_run.py (Adam_LBFGS training) from results/Explicit_Solution_Example/(certain dimension)/ to results/(your equation)/(certain dimension)/, replacing all "Explicit_Solution_Example" with your equation's name.

6. Review beginning lines to enable or disable wandb online logging.

7. Run the experiment:

python results/(your equation)/(certain dimension)/experiment_run.py

9. View your results in the results/(your equation)/(certain dimension)/ folder and on wandb
10. (Optional) Repeat steps 6-9 for further_training.py after the weights of experiment_run.py are generated to get further trained weights using L_inf training.
11. Replace results/ by results_full_history/ to replace the quadrature MLP solver by the full_history MLP solver.
12. If you still have problems, please submit it to Issues.

Project Structure

• equations/: Contains equation definitions
• models/: Neural network model definitions
• optimizers/: Custom optimization algorithms
• tests/: Test methods for evaluating the solver
• results/: Experiment results and wandb keys
• solvers/: Implementation of SCaSML and other solvers

Contributing

Contributions to SCaSML are welcome! Please follow these steps:

1. Fork the repository
2. Create a new branch (git checkout -b feature/your-feature)
3. Make your changes
4. Commit your changes (git commit -am 'Add some feature')
5. Push to the branch (git push origin feature/your-feature)
6. Create a new Pull Request

Citation

Contact

Maintainer&First Author:

Zexi Fan: 2200010816@stu.pku.edu.cn

Collaborators:

Yan Sun: yansun414@gmail.com

Yiping Lu: yiping.lu@northwestern.edu