gtn-base-v5

GTN-Base-V5 GNN Surrogate Model

This repository contains the trained GNN surrogate model (gtn_base_v5_toposplit_cleaned) for power converter topology evaluation.

The model predicts the simulated efficiency and output voltage (Vout) for a given converter topology and duty cycle:

f_theta(T, d) -> (efficiency, Vout)

Where:

• T is the power-converter topology graph.
• d is the duty cycle ($d \in [0.1, 0.9]$).

Model Loading & Usage

Load it using the project's shared_energy.seal_models.load_gtn_base_energy_model function:

from huggingface_hub import hf_hub_download
import torch
from shared_energy.seal_models import load_gtn_base_energy_model

repo_id = "DanielJeongsooLee/gtn-base-v5"
checkpoint_path = hf_hub_download(repo_id=repo_id, filename="gtn_surrogate_model.pt")

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model, payload = load_gtn_base_energy_model(checkpoint_path, map_location=device)
model.eval()

Dataset & Training Details

• Dataset: gtn_dataset_5comp_v2_corrected_5000c_cleaned.jsonl (Cleaned 5000-cycle simulation dataset, 704 unique topologies, 57,024 samples)
• Split Mode: topology_random (Topology-disjoint split, ensuring no overlap of topology structures between training, validation, and test splits)
• Training Samples: 45684
• Validation Samples: 5670
• Test Samples: 5670
• Learning Rate: 3e-4
• Total Epochs: 700 (Best validation epoch: 671)

Performance Metrics

Evaluation metrics on unseen topologies in the test split (Topology RSE):

Final Test RSE:

{
  "efficiency": 0.050900354981422424,
  "vout": 0.00417777756229043
}

Final Validation RSE:

{
  "efficiency": 0.05786220729351044,
  "vout": 0.010579720139503479
}

Files in Repository

• gtn_surrogate_model.pt: PyTorch model state dict and config checkpoint.
• manifest.json: Detailed training summary metadata.
• training_history.jsonl: Training and validation loss/RSE history per epoch.