Answer by immerrr for How to define dictionary structure in ndarray

If you only need named columns and the column set is immutable, you can use record arrays:

In [30]: arr = np.zeros((3,3), dtype=[('position', 'f8', 2),                                      ('boundary', 'i8', (2,2))]); arrOut[30]: array([[([0.0, 0.0], [[0, 0], [0, 0]]), ([0.0, 0.0], [[0, 0], [0, 0]]),        ([0.0, 0.0], [[0, 0], [0, 0]])],       [([0.0, 0.0], [[0, 0], [0, 0]]), ([0.0, 0.0], [[0, 0], [0, 0]]),        ([0.0, 0.0], [[0, 0], [0, 0]])],       [([0.0, 0.0], [[0, 0], [0, 0]]), ([0.0, 0.0], [[0, 0], [0, 0]]),        ([0.0, 0.0], [[0, 0], [0, 0]])]],       dtype=[('position', '<f8', (2,)), ('boundary', '<i8', (2, 2))])In [31]: arr['boundary'][0] = [[1,2], [3,4]]In [32]: arrOut[32]: array([[([0.0, 0.0], [[1, 2], [3, 4]]), ([0.0, 0.0], [[1, 2], [3, 4]]),        ([0.0, 0.0], [[1, 2], [3, 4]])],       [([0.0, 0.0], [[0, 0], [0, 0]]), ([0.0, 0.0], [[0, 0], [0, 0]]),        ([0.0, 0.0], [[0, 0], [0, 0]])],       [([0.0, 0.0], [[0, 0], [0, 0]]), ([0.0, 0.0], [[0, 0], [0, 0]]),        ([0.0, 0.0], [[0, 0], [0, 0]])]],       dtype=[('position', '<f8', (2,)), ('boundary', '<i8', (2, 2))])