GFN-v5: D81-Pretrained Fan Topology GFlowNet

This repository contains the GFlowNet checkpoint used for the GFN-D81Pretrain+D5 method in the MCTS+D5 vs GFN+D5 topology-synthesis comparison.

The model is not a Hugging Face transformers model. It is a PyTorch checkpoint for the project-local fan_original.gflownet.GFlowNetAgent implementation in shared_energy.

Main Files

• gflownet_d81_pretrained.best.pt: best checkpoint used for the reported search.
• gflownet_d81_pretrained.final.pt: final epoch checkpoint.
• manifest.json: training configuration and dataset/reconstruction counts.
• training_metrics.jsonl: offline TB pretraining metrics.
• comparison_report.md: MCTS+D5 vs GFN-D81Pretrain+D5 D81 NGSpice result summary.
• mcts_vs_gfn_top5_d81_ngspice_results.xlsx: workbook with target-level results.
• load_gfn_v5.py: minimal checkpoint loading example.

Reported Result

Final physical audit metric:

max over source-selected top-5 topologies of NGSpice D81 best reward.

Under this D81 NGSpice top-5 topology evaluation:

Method	D81 mean	D81 median	Wins	Targets
MCTS+D5	0.473417	0.245649	4	13
GFN-D81Pretrain+D5	0.547092	0.807352	9	13

The final audit used n_cycles=5000, record_cycles=500, samples_per_cycle=40, and timeout_s=180.

Training Summary

• Mode: fan_original_gfn_d81_pretrain
• Training signal: D81 simulator-label retrieval rewards
• Component count: 5
• Generation mode: fan
• Required device types: Sa, Sb, C, L
• Target gammas: [-3, -2.5, -2, -1.5, -1, -0.5, 0.25, 0.5, 0.75, 1.5, 2, 2.5, 3]
• Completed epochs: 120
• Reconstructable examples: 2494

Loading

This checkpoint expects the original repository code to be importable:

git clone <shared_energy_repo>
cd shared_energy
python artifacts_or_downloaded_repo/load_gfn_v5.py --checkpoint gflownet_d81_pretrained.best.pt

The example prints the architecture configuration and verifies that the checkpoint loads into GFlowNetAgent.

Caveats

• GTN surrogate D5 scores were used for search/ranking, but final winner claims should use NGSpice D81 rewards only.
• This checkpoint is intended for research use with the accompanying codebase, not as a standalone inference API.