Metadata-Version: 2.4
Name: tooppy
Version: 2.8.2
Summary: A topological optimization library for python 3
Author: Shengzhi Wu
Author-email: e1124755@u.nus.edu
Classifier: Programming Language :: Python :: 3
Classifier: License :: OSI Approved :: MIT License
Classifier: Operating System :: OS Independent
Classifier: Topic :: Scientific/Engineering :: Mathematics
Classifier: Topic :: Scientific/Engineering :: Physics
Description-Content-Type: text/markdown
License-File: LICENSE
Requires-Dist: numpy
Requires-Dist: scipy
Requires-Dist: pyvista
Requires-Dist: sympy
Requires-Dist: taichi
Dynamic: author
Dynamic: author-email
Dynamic: classifier
Dynamic: description
Dynamic: description-content-type
Dynamic: license-file
Dynamic: requires-dist
Dynamic: summary

![logo](README.assets/logo.png)

# tooppy

tooppy is a powerful [**to**pological **op**timization](https://en.wikipedia.org/wiki/Topology_optimization) library for **py**thon 3.

What is topological optimization? Consider designing a part using a material, aluminum alloy for instance, where we want its weight not to exceed a given value, while maximizing its strength. The optimal solution is often similar to a truss structure, containing a series of holes. If we have no prior knowledge about the structure, such as starting with a simple rectangular block, the optimization process will inevitably involve changes in the [topology](https://en.wikipedia.org/wiki/Topology). This type of optimization is called topology optimization. Topology optimization is useful in aerospace engineering, where structures often need to be strong while minimizing weight.

If you're a visual artist rather than an engineer, you can still enjoy this library, as the results of topology optimization often yield beautiful, organic structures. You can find many similar forms in nature, such as animal skeletons. I believe you can draw interesting inspiration from the rich variety of the results.

In principle, tooppy is capable of handling topology optimization problems in any dimension, not just the usual 2D and 3D cases.

## Install

```bash
pip install tooppy
```

## Features

### 2D

In this example, we seek a structure with both ends fixed (vertically fixed, horizontally slidable), and a vertical force applied at the center as the load. See `tests\test_2d.py`.

<img src="README.assets/result_2d_0.png" alt="result_2d" style="zoom:50%;" />

### 3D

This example is the 3D version of the one above, with the structure fixed at the four corners (slidable) and a load applied at the center. See `tests\test_3d.py`.

<img src="README.assets/result_3d_0.png" alt="result_3d_0" style="zoom:67%;" />

You can conver the 3D array to a mesh and save it as an STL file. See `tests\conver_to_mesh.py`.

<img src="README.assets/result_3d_mesh_0.png" alt="result_3d_mesh_0" style="zoom:67%;" />

tooppy provids methods for calculating **volume**, **centroid** and **inertia tensor** of meshes. See `tests/get_geometric_information.py` for details.

### Multiple Loads

tooppy supports solving for structures that are robust under multiple load conditions. The following example designs a structure with 4 fixed corners, where the center may be subjected to loads in the X, Y, and Z directions, requiring robustness under all three load conditions. Note that, since there may be horizontal loads, the previously used slidable supports can no longer be applied. All supports here are fully fixed. See `tests\test_3d_multiload.py`.

<img src="README.assets/result_3d_1.png" alt="result_3d_1" style="zoom:50%;" />

Requiring the structure to be robust under various load conditions often leads to relatively moderate designs.

### Mask

Another powerful feature of tooppy is the mask. Users can specify certain regions where material placement is prohibited. This is particularly useful when the part is in a constrained environment or needs to fit with other parts in space. The regions are specified using a boolean array.

This example has the same configuration as the first 3D example, except that material cannot be placed within the specified four cylindrical regions. See `tests\test_3d_mask.py`.

<img src="README.assets/result_3d_2.png" alt="result_3d_2" style="zoom:67%;" />

Elements disabled by mask will not be involved in the calculation. This can significantly improve the calculation speed, as the time consumed by the calculation is approximately proportional to the square of the degrees of freedom (dofs) in the scene. So one can calculate the results of different resolutions in stages. See `tests\use_different_resolutions_in_stages.py`, where a low-resolution result is first calculated, and then it is used as a mask to calculate a high-resolution result. In this way a high-resolution result can be obtained in a shorter time.

## More Interesting Examples

### Compression / Shear / Bending Resistant Structures

We aim to find 2D structures that are highly resistant to compression, shear or bending.  We completely fix the lower end of a rectangular region, then apply pressure (vertical), shear force (horizontal), or bending force (rotational) to the upper end. See `tests\test_2d_compression_resistant_structure.py`, `tests\test_2d_shear_resistant_structure.py` and `tests\test_2d_bending_resistant_structure.py`.  The optimized structure is shown in the figure below.

![result_2d_csb_resistant](README.assets/result_2d_csb_resistant.png)

We notice that in the first case, it tries to use arch-like or catenary structures to increase stiffness under pressure, while also attempting to use tree-like structures to save material. You can see similar structures in Antoni Gaudí's famous masonry building, the [Sagrada Família](https://en.wikipedia.org/wiki/Sagrada_Fam%C3%ADlia). See the picture.

<img src="README.assets/Sagrada_Família.jpg" alt="Sagrada_Família" />

## Performance

The solver used to solve the sparse linear system can be set through the option `linalg_solver` of `tooppy.solve`. There are two choices:

|  `linalg_solver`   |                         Explanation                          |
| :----------------: | :----------------------------------------------------------: |
|     `'scipy'`      | Based on `from scipy.sparse.linalg import spsolve`. Has broader support and exhibits more consistent performance across different hardware and Python versions. The performance of this solver is relatively superior for two-dimensional problems |
|     `'taichi'`     | Based on `taichi` and the conjugate gradient method. Supports GPU acceleration. For large-scale problems (3d, with 60k+ degrees of freedom), the solving speed is significantly higher than that of `scipy`. When choosing this option, you can also use another option `taichi_arch` to set on which hardware to solve. You can choose `'cpu'` or `'gpu'` (default) |
| `'auto'` (default) |  Automatically determined based on the size of the problem   |

The performance comparison is shown in the following figure (the problem in `test_2d.py` and `test_3d.py`, the different resolutions constructed by scaling each axis proportionally, with the CPU being AMD Ryzen 9 9900X3D and the GPU being NVIDIA GeForce RTX 5090).

![performances](README.assets/performances_2d.png)

![performances](README.assets/performances_3d.png)

Although the Taichi solver performs significantly better than the SciPy solver for large-scale 3D problems, numerical experiments have revealed that for 2D problems, the efficiency of the Taichi solver is lower than that of the SciPy solver. The reason for this is unknown.

## The Numbering of Vertices and Elements

To solve a problem, you need to construct one or more of the functions `get_fixed`, `get_load`, and `get_mask`, which are used to set constraints, loads, and masks. In these functions, you need to return an array to inform the program which degrees of freedom are fixed, which vertices are subjected to external forces in which directions, and which elements must be left empty without allowing material placement.

The finite element method (FEM) solution domain is composed of a series of square/cubic/hypercubic elements with edge lengths of 1, and the number of elements is the product of the entries in the `resolution`. The vertices are the endpoints of the elements and outnumber the elements by one along each axis. For example, if `resolution = [2, 3, 5]`, there are $2 \times 3 \times 5 = 30$ elements and $3 \times 4 \times 6 = 72$ vertices.

Constrains and loads are defined on the vertices and must follow the vertex numbering. The mask and solution results are defined on the elements and follow the element numbering.

Generally, you don't need to pay too much attention to numbering. Both `get_fixed` and `get_load` accept the parameters `resolution`, `ndof`, and `coordinates`, while `get_mask` accepts `resolution`, `number_of_cells`, and `coordinates`. In most cases, you can easily construct the constraints, loads, and masks you want using this information. A typical approach is to use `coordinates` to construct a boolean array to select which items of the NumPy array to manipulate, as demonstrated in `tests\test_3d_mask.py`.

When you need to find the indices of vertices on the edges or faces of the solution domain (rectangle, cuboid, or hypercuboid), you can easily achieve this using `tooppy.get_indices_on_boundary_elements(resolution, axis_selection)`. For example, if you want to select the leftmost column in a $5 \times 5$ grid, you can write `tooppy.get_indices_on_boundary_elements([5, 5], [[True, False], None])`. `[[True, False], None]` means selecting the start point but not the end point on the first axis, and selecting all on the second axis. If you want to select the two faces of a cuboid along the positive and negative Y directions, your `axis_selection` should be written as `[None, [True, True], None]`. If you want the four edges in the Z direction of a cuboid, you would write `[[True, True], [True, True], None]`.

Note that since the `resolution` passed into `get_fixed`, `get_load`, and `get_mask` represents the size of the element array, when you try to obtain the vertex indices instead of the element indices—specifically when writing the content of `get_fixed` and `get_load`—you must manually add one to each entry of `resolution` before passing it into `get_indices_on_boundary_elements`.

## Element Stiffness Matrix Cache

The element stiffness matrix is automatically cached in`element_stiffness_matrices/` on the first time you solve a problem by default. In subsequent calculations, this matrix will be automatically loaded to save time. When you use new materials (with new [Young's modulus](https://en.wikipedia.org/wiki/Young%27s_modulus) `E` or [Poisson's ratio](https://en.wikipedia.org/wiki/Poisson%27s_ratio) `nu`), or calculate problems in different dimensions, the matrix needs to be recalculated.

You can disable this behavior by using the `skip_calculating_element_stiffness_matrix_if_exists=False` option. You can also use the `element_stiffness_matrix_file_dir` option to specify the location where the matrix is stored or to be stored.

## Important Updates

### 2.7.0

Elements disabled by mask are not included in the calculation. This significantly improves the performance of scenarios with masks, and supports the use of different resolutions at different stages to enhance computational efficiency.

### 2.8.0

Introduced GPU acceleration, enabling a speedup of 900% for large-scale problems (3d, with 90k+ degrees of freedom).

## TODO

1. Save memory usage
1. Generate incidents based on geometry elements (lines, circles, etc.)

## Reference

This project referenced the following projects during development:

1. [ToPy](https://github.com/williamhunter/topy)
2. [topopt2D](https://github.com/worbit/topopt2D/tree/master)

The picture of Sagrada Família is provided by Zhangqin Zheng.
