|
|
Project: Flexible Graph Construction for Neural Weather Models
MLLAM (Machine Learning for Limited Area Models) is an open-source collaboration of meteorological institutes and researchers building data-driven regional weather forecasting - GraphCast-style GNN models that run the encode → process → decode cycle over mesh graphs.
My project makes mesh-graph construction topology-agnostic across the MLLAM stack, so researchers can swap rectangular, triangular (Delaunay), or fully custom meshes into the same forecasting pipeline - instead of being locked into one hardcoded grid.
🔗 Official project page: summerofcode.withgoogle.com → Flexible Graph Construction
🧑🏫 Mentors: Leif Denby · Hauke Schulz · Joel Oskarsson
| Contribution | Repo | Status |
|---|---|---|
#81 - Two-step mesh_layout architecture separating mesh coordinate creation from connectivity - the foundation for all alternative mesh topologies |
weather-model-graphs |
✅ Merged |
| #92 - Triangular (Delaunay) multi-range meshes as the first alternative topology, built on the new layout architecture | weather-model-graphs |
🔄 In review |
#123 - to_torch_tensors_on_disk: standardized on-disk graph serialization so any WMG graph loads directly into neural-lam |
weather-model-graphs |
✅ Merged |
#596 - create_graph_with_wmg bridge CLI, replacing 600+ lines of duplicated graph-construction code in neural-lam with the shared WMG library |
neural-lam |
🔄 In review |
| #79 - Prebuilt / bring-your-own mesh layouts: design for injecting externally-generated meshes (e.g. from existing NWP model grids) | weather-model-graphs |
🧩 Design agreed with maintainers |
| #144 - CI benchmark regression detection for graph-construction performance | weather-model-graphs |
🧩 Proposal accepted |
Layer 1 ✅ Mesh layout architecture - rectangular + triangular (Delaunay) topologies
Layer 2 🔄 Bridge: one shared graph pipeline + on-disk format between WMG and neural-lam
Layer 3 🔜 Migrate neural-lam's internal graph representation to PyG HeteroData
Layer 4+ 🌟 Graph quality metrics, density-adaptive meshes, spherical coordinates
Why it matters: ML-based weather forecasting is becoming operational at national weather services. Flexible mesh construction lets scientists match the graph to the physics - finer resolution where the weather is complex - rather than the other way around.
| Achievement | Year | |
|---|---|---|
| 🌦️ | Google Summer of Code Contributor - MLLAM · Flexible Graph Construction for neural weather models (350h project) | 2026 |
| 📚 | Amazon ML Summer School - Shortlisted (Amazon India) | 2026 |
| 🦈 | GitHub Pull Shark & Pair Extraordinaire - earned through merged PRs and collaborative work | - |
| 🧩 | Competitive programming on LeetCode | - |
I contribute to open-source Machine Learning & Scientific Computing projects - currently as a Google Summer of Code 2026 contributor with MLLAM, working on graph construction for neural weather prediction models used in real forecasting research.
I believe in building in the open and giving back to the community.
💡 I'm always open to interesting conversations, collaborations, and opportunities!
