Pytorch operation fails for surfaces with only one contact.

[benk-mira]

Describe the bug

When I use the pytorch backend in release 2025.2.1a2, gempy fails if a surface is only constrained by one contact point. Here is the traceback:

To Reproduce

import gempy as gp
import numpy as np

geo_model = gp.create_geomodel(
    project_name='Model1',
    extent=[0, 780, -200, 200, -582, 0],
    resolution=(50, 50, 50),
    # refinement=4, # We will use octrees
    structural_frame=gp.data.StructuralFrame.initialize_default_structure()
)

gp.add_surface_points(
    geo_model=geo_model,
    x=[225],
    y=[0],
    z=[-95],
    elements_names=['surface1']
)
gp.add_orientations(
    geo_model=geo_model,
    x=[350],
    y=[0],
    z=[-120],
    elements_names=['surface1'],
    pole_vector=[np.array([0, 0, 1])]
)
element = gp.data.StructuralElement(
    name='surface2',
    color='#FFA833',  # color=next(geo_model.structural_frame.color_generator),
    surface_points=gp.data.SurfacePointsTable.from_arrays(
        x=np.array([460]),
        y=np.array([0]),
        z=np.array([-280]),
        names='surface2'
    ),
    orientations=gp.data.OrientationsTable.initialize_empty()
)

geo_model.structural_frame.structural_groups[0].append_element(element)

gp.compute_model(geo_model, engine_config=gp.data.GemPyEngineConfig(backend=gp.data.AvailableBackends.PYTORCH, use_gpu=False))

Expected behavior

I expect that as long as another layer has two surface points, as in the example I provided, the interpolation should be well defined and result in no errors. This is documented on the website:

Desktop (please complete the following information):

I'm on windows using:
gempy:

gempy-engine:

Additional context

There is no issue if I set the backend to 'numpy'.

[javoha]

Hi @benk-mira,

thanks for the question. This is interesting. Let's break it down: First of all this behavior does not contradict what is written in the documentation: "Two surface points for at least one surface in a structural group/series". In your case you have one surface point for each of two units within the same group.

But that is nitpicking and this model actually runs in my environment using gempy '2025.2.0' and torch '2.3.1+cu121'. So I guess my advice would be to use the current main release version instead of the pre-release you are on unless you have a specific reason for that.

Apart from that I will keep this issue open and we will look into it before the next main release.

Cheers,
Jan

[benk-mira]

Ah ok, so you don't consider the orientation a surface point. I can confirm the error exists if I add another point to 'layer1' in my example. The reason I'm looking at a pre-release is that it fixes a bug where the model cells were being mislabelled:

gempy-project/gempy_engine#20