Say we have a point cloud of a surface (np.array of shape=(N, 3)) like the below image, on which we want to define a trajectory.

So far using scipy.interpolate.griddata I can project the 2D defined trajectory and get the following:

but then calculating surface normal for each of these points is not quite easy.

Is there a way using PyTorch/TensorFlow to do something similar while using the automatic differentiation feature of them to calculate such surface normal easier for each point?

P.S. Although off-topic and better for a separate question: I am also interested if there is a way to define the trajectory on a local coordinate system on this surface and map it to the world 3D coordinate system.