Research map / Graph learning / Graph self-supervised learning
RingFormer: A Ring-Enhanced Graph Transformer for Organic Solar Cell Property Prediction
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RingFormer, a graph transformer over a hierarchical atom-ring graph with ring-level cross-attention, to predict organic solar cell properties, outperforming 11 baselines on five OSC datasets.
Problem
Predicting organic solar cell (OSC) molecular properties such as power conversion efficiency, where existing GNNs and graph transformers fail to capture ring systems and their interconnections.
Approach
Builds a hierarchical graph with atom-level, ring-level, and inter-level (atom-ring membership) graphs; each layer combines GINE message passing on atoms, a cross-attention over ring-level edges with a virtual molecule node, and inter-level message passing, followed by hierarchical fusion.
Key results
- On CEPDB, RingFormer achieves test MAE 0.189 for PCE vs 0.244 for the best competitor (GraphViT), a 22.77% relative improvement.
- RingFormer achieves best PCE MAE on CEPDB, HOPV, NFA, and PD, and runner-up on PFD (1.776 vs 1.770 for MACCS).
- In multi-task learning on CEPDB, RingFormer is best on all six properties, e.g. Jsc MAE 5.993 vs 7.514 for GPS (20.24% relative improvement).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RingFormer instead of GraphViT | PCE prediction · CEPDB | test MAE: 0.244 -> 0.189 | Helped |
| Use RingFormer instead of GraphViT | PCE prediction · HOPV | test MAE: 1.479 -> 1.477 | Helped |
| Use RingFormer instead of MACCS | PCE prediction · PFD | test MAE: 1.770 -> 1.776 | Hurt |
| Use RingFormer instead of GraphViT | PCE prediction · NFA | test MAE: 2.467 -> 2.259 | Helped |
| Use RingFormer instead of Graphormer | PCE prediction · PD | test MAE: 2.522 -> 2.482 | Helped |
| Add both atom- and ring-level structure rather than using only the atom graph (ablation) | RingFormer · CEPDB | test MAE: 0.550 -> 0.189 | Helped |
| Add both atom- and ring-level structure rather than using only the ring graph (ablation) | RingFormer · CEPDB | test MAE: 0.358 -> 0.189 | Helped |
| Add the ring-level graph (ablation) | RingFormer without ring graph · CEPDB | test MAE: 0.606 -> 0.189 | Helped |
| Add inter-level atom–ring message passing (ablation) | RingFormer without inter-level graph · CEPDB | test MAE: 0.315 -> 0.189 | Helped |
| Remove the virtual molecule node from ring cross-attention (ablation) | RingFormer · CEPDB | test MAE: 0.1886 -> 0.1860 | Helped |
| Use ring cross-attention instead of GINE as ring encoder (ablation) | RingFormer · CEPDB | test MAE: 0.3576 -> 0.1886 | Helped |
| Use rings instead of BRICS motifs as high-level structure (ablation) | RingFormer · CEPDB | test MAE: 0.4556 -> 0.1886 | Helped |
Surprising findings
- Fingerprint-based methods (ECFP, MACCS) outperform most GNN-based and graph transformer baselines on the small experimental datasets (HOPV, PFD, NFA, PD).
Limitations
- Performance on the small experimental datasets degrades with deeper models (more layers), attributed to limited data.
- Gains on PFD are marginal; RingFormer is only runner-up there.
- Evaluated only on OSC property datasets; generalization to other molecular tasks not shown.
Experimental setup
Datasets: CEPDB HOPV PFD NFA PD
Benchmarks: CEPDB HOPV PFD NFA PD
Metrics: MAE
Baselines: MACCS ECFP GINE GINE-VN AttentiveFP O-GNN TopKPool SAGPool Graphormer GPS GraphViT
Builds on: GINE Graphormer GPS O-GNN
Closest papers by meaning
- Complete and Efficient Graph Transformers for Crystal Material Property Prediction (ICLR 2024 · similarity 0.85)
- Structure-Aware E(3)-Invariant Molecular Conformer Aggregation Networks (ICML 2024 · similarity 0.84)
- Learning Over Molecular Conformer Ensembles: Datasets and Benchmarks (ICLR 2024 · similarity 0.84)
- Conformal Crystal Graph Transformer with Robust Encoding of Periodic Invariance (AAAI 2024 · similarity 0.84)
- WGFormer: An SE(3)-Transformer Driven by Wasserstein Gradient Flows for Molecular Ground-State Conformation Prediction (ICML 2025 · similarity 0.84)
- RFL: Simplifying Chemical Structure Recognition with Ring-Free Language (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.