Airport Runway Blockade Encoder-Only Transformer: Model Weights and Configuration
This repository contains the trained encoder-only Transformer used for scene-conditioned takeoff-and-landing window (TLW) candidate generation on crater-damaged runways. The network proposes ten normalized center/heading poses from the complete runway-crater configuration. A deterministic geometric evaluator and bounded local search must then compute feasible window lengths, refine the poses, and select the longest verified result.
Model details
| Item | Value |
|---|
| Architecture | Encoder-only Transformer without positional encoding |
| Inputs | One runway descriptor and up to 20 crater descriptors |
| Outputs | 10 normalized candidate poses (x, y, heading) |
| Hidden dimension | 128 |
| Encoder layers / attention heads | 2 / 4 |
| Feed-forward dimension | 512 |
| Trainable parameters | 458,078 |
| Framework | PyTorch |
| Training seed | 2026 |
The runway feature vector is (length / 3600, width / 60, aspect_ratio / 80). Each crater is represented as (x / runway_length, y / runway_width, radius / runway_width). Crater sequences are padded to 20 entries and accompanied by a binary mask. Predicted x and y are in [0, 1]; the predicted heading is in [-1, 1] and is rescaled by atan(runway_width / runway_length) in degrees.
Training
The model was trained on 10,000 synthetic military-runway scenes with ES-derived supervision. A fixed split produced 8,500 training scenes and 1,500 validation scenes. Training used AdamW for 40 epochs with a batch size of 256, learning rate 1e-3, and weight decay 1e-5. model_best.pt is the checkpoint with the lowest validation loss.
The checkpoint is a PyTorch dictionary containing:
model: model state dictionaryconfig: architecture and training configurationepoch: selected epochval_loss: selected validation loss
Independent-test results
Results below use a separate 2,000-scene synthetic test set. Ratios compare returned TLW length with the stored pseudo-expert length.
| Method | Mean ratio | Median ratio | 5th percentile | Zero-length results |
|---|
| Raw Transformer Top-10 proposals | 0.7758 | 0.8090 | 0.4444 | 0 |
| Transformer Top-10 + 3-round/16-point refinement | 0.9590 | 0.9961 | 0.7850 | 0 |
The pseudo-expert is a numerical search result, not a proof of the global optimum, so an individual ratio may exceed 1.0.
Checkpoint format
model_best.pt is a PyTorch checkpoint containing the trained state dictionary, architecture/training configuration, selected epoch, and validation loss. The model architecture is documented above; implementation and optimization source code are intentionally not distributed in this repository.
Repository contents
model_best.pt: selected trained model checkpointconfig.json: model architecture and training configurationREADME.md: architecture, training, evaluation, and limitation notes
Intended use and limitations
This artifact is intended for research, reproducibility, and independent evaluation of learning-guided geometric optimization. It was trained and evaluated on synthetic crater-damage scenes within the distributions and normalization limits described above. Performance outside those conditions is not guaranteed.
Neural outputs are proposals, not feasibility certificates. Every candidate must be checked by the deterministic runway-boundary and crater-clearance evaluator before use. This model must not be used by itself for operational aviation, safety-critical, military, or emergency decision-making.
No training dataset is included in this model repository. Dataset publication and its terms should be handled separately.
Citation
The associated manuscript citation will be added after publication. Until then, please cite this model repository and record the exact Hugging Face revision used.
