Skip to content

README.md

Latest commit

 

History

History
66 lines (51 loc) · 2.27 KB

README.md

File metadata and controls

66 lines (51 loc) · 2.27 KB
Raw

NeurIPS Open Polymer Prediction 2025

Kaggle Rank Organizer

Kaggle Bronze Medal solution, ranked 217 / 2240 in NeurIPS - Open Polymer Prediction 2025.

The competition focuses on molecular property prediction for polymers from SMILES strings. The final solution uses robust SMILES canonicalization, target-specific descriptor models, graph neural network features, and weighted blending across complementary branches.

Result

Field Value
Competition NeurIPS - Open Polymer Prediction 2025
Organizer NeurIPS / Kaggle
Medal Bronze Medal
Rank 217 / 2240
Awarded September 16, 2025
Targets Tg, FFV, Tc, Density, Rg

Method Overview

  • Canonicalize SMILES and filter invalid polymer R-group notation before training and inference.
  • Build target-specific descriptor tables with RDKit / Mordred features.
  • Train CatBoost and XGBoost regressors per target, using only feature columns shared by train and test descriptors.
  • Build a graph branch from atom and bond features with molecule-level descriptors.
  • Blend GNN, CatBoost, and XGBoost predictions with fixed weights:
0.4 * GNN + 0.3 * CatBoost + 0.3 * XGBoost

Repository Layout

configs/
  default.yaml
scripts/
  infer_tabular.py
  infer_graph.py
  blend.py
src/polymer_prediction/
  constants.py
  smiles.py
  descriptors.py
  tabular.py
  graph.py
  ensemble.py

Quick Start

python -m pip install -e .
python scripts/infer_tabular.py --config configs/default.yaml --model catboost
python scripts/infer_tabular.py --config configs/default.yaml --model xgboost
python scripts/infer_graph.py --config configs/default.yaml --checkpoint /kaggle/input/polymer-gnn/model.pt
python scripts/blend.py --config configs/default.yaml

The default configuration follows Kaggle input paths. For local runs, edit configs/default.yaml.

Notes

Raw competition files, external descriptor datasets, trained checkpoints, and submission artifacts are intentionally excluded. The repository keeps the reproducible solution code and configuration surface.