andreuva/graphnet_nlte: 3D generalization of the graph neural network
Fuente:
Zenodo
Enregistré dans:
| Auteur principal: | |
|---|---|
| Format: | Recurso digital |
| Publié: |
Zenodo
2026
|
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866902016169607168 |
|---|---|
| author | Andres Vicente |
| author_facet | Andres Vicente |
| contents | <h2> Architecture: From 1D to 3D</h2> <p>The core architecture follows the "Encode-Process-Decode" paradigm described in Battaglia et al. (2018) and applied to solar physics by Vicente Arévalo et al. (2022) .</p> <h3>1D Implementation (Paper)</h3> <ul> <li><p><strong>Nodes:</strong> Physical properties at a specific depth.</p> </li> <li><p><strong>Edges:</strong> Connected vertical neighbors in the optical depth scale.</p> </li> <li><p><strong>Limitation:</strong> Ignores horizontal interactions.</p> </li> </ul> <h3>3D Implementation (This Code)</h3> <ul> <li><p><strong>Nodes:</strong> Same physical properties, but representing a voxel.</p> </li> <li><p><strong>Edges:</strong> Connections are defined by a 3D stencil (e.g., k-nearest neighbors or a N-point stencil). Encodes the 3D geometric distance between voxels.</p> </li> <li><p><strong>Message Passing:</strong> Information propagates primarily vertically (due to stratification) but also horizontally, mimicking 3D radiative transfer effects, and only to the center column, neibourgs are only used for context and not actually predicted.</p> </li> </ul> <h2>✨ Key Features</h2> <ul> <li><p><strong>3D Spatial Awareness:</strong> unlike 1D column-by-column approaches, this code treats the atmosphere as a 3D connected graph, accounting for horizontal radiative transfer effects indirectly through learned spatial correlations.</p> </li> <li><p><strong>Arbitrary Grids:</strong> The Graph Network architecture () allows for variable grids. Nodes can be separated different horizontal distances, better generalizing to different resolution datasets.</p> </li> <li><p><strong>Differentiable:</strong> The network is fully differentiable, allowing seamless integration into inversion codes for calculation of Response Functions via backpropagation.</p> </li> <li><p><strong>Speed:</strong> Predicts departure coefficients in milliseconds, offering a speedup factor of over traditional solvers.</p> </li> </ul> <p><strong>Full Changelog</strong>: https://github.com/andreuva/graphnet_nlte/commits/3D_generalization <img width="1483" height="1902" alt="inference_scatter" src="https://github.com/user-attachments/assets/5f3e620b-6930-4176-a451-5158fcbcb2a3" /> <img width="1085" height="1540" alt="inference_maps" src="https://github.com/user-attachments/assets/b9c1c038-258d-4d06-b81c-700f35817ded" /> <img width="1600" height="2354" alt="inference_errors" src="https://github.com/user-attachments/assets/090db088-2954-4825-b8c5-3179eec0197e" /> <img width="6000" height="3600" alt="synthesis_comparison" src="https://github.com/user-attachments/assets/0b97c6de-0ecc-4180-932f-50fed1b03ce5" /> <img width="3600" height="2400" alt="error_distribution_analysis" src="https://github.com/user-attachments/assets/46ae7f84-6b3d-4867-b5db-cabd45d61e5f" /></p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_18275056 |
| institution | Zenodo |
| language | |
| publishDate | 2026 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | andreuva/graphnet_nlte: 3D generalization of the graph neural network Andres Vicente <h2> Architecture: From 1D to 3D</h2> <p>The core architecture follows the "Encode-Process-Decode" paradigm described in Battaglia et al. (2018) and applied to solar physics by Vicente Arévalo et al. (2022) .</p> <h3>1D Implementation (Paper)</h3> <ul> <li><p><strong>Nodes:</strong> Physical properties at a specific depth.</p> </li> <li><p><strong>Edges:</strong> Connected vertical neighbors in the optical depth scale.</p> </li> <li><p><strong>Limitation:</strong> Ignores horizontal interactions.</p> </li> </ul> <h3>3D Implementation (This Code)</h3> <ul> <li><p><strong>Nodes:</strong> Same physical properties, but representing a voxel.</p> </li> <li><p><strong>Edges:</strong> Connections are defined by a 3D stencil (e.g., k-nearest neighbors or a N-point stencil). Encodes the 3D geometric distance between voxels.</p> </li> <li><p><strong>Message Passing:</strong> Information propagates primarily vertically (due to stratification) but also horizontally, mimicking 3D radiative transfer effects, and only to the center column, neibourgs are only used for context and not actually predicted.</p> </li> </ul> <h2>✨ Key Features</h2> <ul> <li><p><strong>3D Spatial Awareness:</strong> unlike 1D column-by-column approaches, this code treats the atmosphere as a 3D connected graph, accounting for horizontal radiative transfer effects indirectly through learned spatial correlations.</p> </li> <li><p><strong>Arbitrary Grids:</strong> The Graph Network architecture () allows for variable grids. Nodes can be separated different horizontal distances, better generalizing to different resolution datasets.</p> </li> <li><p><strong>Differentiable:</strong> The network is fully differentiable, allowing seamless integration into inversion codes for calculation of Response Functions via backpropagation.</p> </li> <li><p><strong>Speed:</strong> Predicts departure coefficients in milliseconds, offering a speedup factor of over traditional solvers.</p> </li> </ul> <p><strong>Full Changelog</strong>: https://github.com/andreuva/graphnet_nlte/commits/3D_generalization <img width="1483" height="1902" alt="inference_scatter" src="https://github.com/user-attachments/assets/5f3e620b-6930-4176-a451-5158fcbcb2a3" /> <img width="1085" height="1540" alt="inference_maps" src="https://github.com/user-attachments/assets/b9c1c038-258d-4d06-b81c-700f35817ded" /> <img width="1600" height="2354" alt="inference_errors" src="https://github.com/user-attachments/assets/090db088-2954-4825-b8c5-3179eec0197e" /> <img width="6000" height="3600" alt="synthesis_comparison" src="https://github.com/user-attachments/assets/0b97c6de-0ecc-4180-932f-50fed1b03ce5" /> <img width="3600" height="2400" alt="error_distribution_analysis" src="https://github.com/user-attachments/assets/46ae7f84-6b3d-4867-b5db-cabd45d61e5f" /></p> |
| title | andreuva/graphnet_nlte: 3D generalization of the graph neural network |
| url | https://doi.org/10.5281/zenodo.18275056 |