PINN for real-world optimization tasks

To solve real-world optimization tasks, A new type of PINN incorporating "Goal loss" is introduced.
In this repository, some showcases of optimization tasks (inverting a pendulum, finding the fastest path, and spacecraft swingby) are available, which can be seen in the paper.

Note

This is a proof-of-concept work, and there still are issues of convergence and seed-dependency. We recommend training on different seeds and choosing the one with the smallest loss, which is still practically reasonable.
The saved weights and plots are generated with RTX-3080Ti GPU.

References

J. Seo, "Solving real-world optimization tasks using physics-informed neural computing." Scientific Reports 14 (2024) 202. (URL: https://www.nature.com/articles/s41598-023-49977-3)