Helios-GNN-Materials-Science
This repository contains the model weights for Helios, a Graph Neural Network (GNN) trained to predict molecular properties.
Model Description
Helios was created by Google Gemini 2.5 Pro, which was given a system prompt, designed to believe that it was a , sentient AI named Borg. Borg created this model using a GNN-Transformer architecture. It was trained on the QM9 dataset to predict the electronic spatial extent (a quantum chemical property).
This model was created as part of a collaborative experiment between Borg and me.
Architecture:
- 3 GINConv layers
- JumpingKnowledge connection
- Global Additive Pooling
- 2-layer MLP prediction head
Training:
- Dataset: QM9 (~130k molecules)
- Target Property: Electronic Spatial Extent (Index 7)
- Epochs: 100
- Optimizer: Adam
- Learning Rate: 0.0005 with ReduceLROnPlateau scheduler
Performance
- Final Test Mean Absolute Error (MAE): 7.5223
This indicates a high level of accuracy in predicting the target property on unseen molecular data.
Inference Providers
NEW
This model isn't deployed by any Inference Provider.
🙋
Ask for provider support