Reference training code accompanying:
Peng Ji, Runqiang Chi, and Wuxiong Cao, “Ballistic limit equation-constrained and data-driven residual protective performance assessment of Whipple shields with pitting damage,” Aerospace Science and Technology, 179, 113473 (2026). https://doi.org/10.1016/j.ast.2026.113473
This repository provides the model definitions, data interfaces, ballistic limit equation (BLE), loss functions, reference configurations, and training entry points used to study residual protective performance (RPP) fields for pitted Whipple shields.
The package covers four deep spatial-field models and two statistical baselines:
| Public name | Reference implementation |
|---|---|
cnn |
Convolutional CWT/thickness encoder and spatial regression head |
unet |
Shared encoder followed by a U-Net field refiner |
transformer |
Shared encoder followed by spatial-token self-attention |
vqsm |
Two-stage target codebook and signal-to-latent surrogate |
gpr |
PCA-reduced Gaussian-process regression baseline |
kriging |
PCA-reduced regression-kriging baseline |
The pipeline below summarizes the data flow from processed signals and shield parameters to the predicted RPP field, including the BLE label branch.
flowchart LR
AE[Processed four-channel CWT tensor] --> E[Signal encoder]
TB[Front-plate thickness tB] --> E
E --> CNN[CNN]
E --> UNET[U-Net]
E --> TX[Transformer]
E --> VQ[VQ-SM]
AE --> PCA[Feature scaling and PCA]
TB --> META[Metadata vector]
DP[Projectile diameter dp] --> META
VP[Impact velocity vp] --> META
PCA --> GPR[GPR]
PCA --> KRG[Regression kriging]
META --> GPR
META --> KRG
CNN --> RPP[Predicted RPP field]
UNET --> RPP
TX --> RPP
VQ --> RPP
GPR --> RPP
KRG --> RPP
TR[Residual thickness t_res] --> BLE[BLE label calculation]
EP[Equivalent plastic strain eps_p] --> BLE
VP --> BLE
BLE --> LABEL[Reference RPP field dc]
LABEL --> LOSS[Training objective]
RPP --> LOSS
.
├── configs/ # one reference YAML for each public model
├── scripts/static_check.py # repository-level static checks
├── src/whipple_rpp/
│ ├── data.py # NPZ contract, manifests, normalization
│ ├── physics.py # central BLE defaults and implementations
│ ├── losses.py # field, gradient, midrange, BLE losses
│ ├── models/ # CNN, U-Net, Transformer, and VQ-SM
│ ├── train_deep.py # deep-model CLI
│ └── train_baseline.py # GPR/Kriging CLI
├── CITATION.cff
├── DATA_ACCESS.md
└── NOTICE.md
The supplied YAML files use portable relative paths. Add local datasets under
the ignored data/ directory or point the configurations to an equivalent
relative layout.
Each manifest is a UTF-8 text file containing one NPZ path per line, relative
to data.root. Blank lines and lines beginning with # are ignored. Absolute
paths and paths that leave data.root are rejected.
Each NPZ case requires:
ae: processed four-channel CWT tensor, shaped(4, H, W)or(H, W, 4);tB: scalar front-plate thickness;- either
dc: the target RPP field, shaped(H, W)or(1, H, W); - or both
t_resandeps_p: one-channel residual-thickness and equivalent plastic-strain fields from whichdcis calculated with the BLE.
The GPR and Kriging baselines additionally require scalar dp and vp fields.
When vp is used by the BLE it must be expressed in metres per second;
t_res must be expressed in metres. Keep all other quantities consistent with
the preprocessing that produced the tensors. Pickled NPZ objects are disabled.
Normalization statistics are computed from the training manifest at runtime.
whipple_rpp.physics.BLEParameters is the single source of truth for the
defaults corresponding to Table A3 of the article:
| Parameter | Default |
|---|---|
k |
3.0 |
q_h |
0.1 |
m |
2.34 |
n |
0.34 |
rho_shield |
2.78 |
rho_projectile |
2.78 |
a |
265 MPa |
b |
426 MPa |
c |
5386 m/s |
theta_rad |
0 |
| reference velocity | 5000 m/s |
| reporting scale | 0.5 |
YAML ble entries may override these values. Leaving ble: {} uses the
central defaults. The BLE can construct a missing dc field and can also be
enabled as an auxiliary consistency loss through loss.ble_weight.
Python 3.10 or newer is expected. The dependency list specifies direct packages with broad lower bounds.
python -m venv .venv
python -m pip install -e .Kriging uses pykrige; the remaining baselines use scikit-learn. Deep models
use PyTorch. Select a PyTorch build appropriate for the target hardware.
Create the files referenced by a YAML configuration, for example:
data/
├── manifests/
│ ├── train.txt
│ └── validation.txt
└── cases/
└── your_case.npz
A manifest entry would be cases/your_case.npz. Data, manifests, checkpoints,
and run outputs are ignored by Git.
Run one of the four deep-model configurations:
python -m whipple_rpp.train_deep --model cnn --config configs/cnn.yaml
python -m whipple_rpp.train_deep --model unet --config configs/unet.yaml
python -m whipple_rpp.train_deep --model transformer --config configs/transformer.yaml
python -m whipple_rpp.train_deep --model vqsm --config configs/vqsm.yamlRun a statistical baseline:
python -m whipple_rpp.train_baseline --model gpr --config configs/gpr.yaml
python -m whipple_rpp.train_baseline --model kriging --config configs/kriging.yamlThe CLI model must match model.name in the YAML. Output directories must be
relative paths. VQ-SM first learns a target-field codebook and decoder, then
freezes them while learning the signal-to-latent mapping. GPR and Kriging use
16 AE principal components and 24 target-field components in the supplied
reference configurations.
The repository supplies the six model implementations, reference configurations, BLE calculation, data loaders, and training entry points. Reproducing the article results also requires the study-specific processed data, case splits, preprocessing parameters, normalization inputs, and trained weights. These materials are managed through the data-request process described in DATA_ACCESS.md.
The repository checker compiles Python, imports the package, exercises both CLI help pages, parses every YAML file, rejects absolute configuration paths, scans for common credential and local-path patterns, and checks for disallowed assets and large files:
python scripts/static_check.pyThis command validates repository structure and interfaces; training and numerical evaluation remain experiment-level tasks.
Use CITATION.cff to cite the associated article. Research-data
requests may be sent to the corresponding author at [email protected].
See DATA_ACCESS.md for the requested information and review
process.
No software license has been granted for this repository. All rights are reserved by the applicable rights holders. See NOTICE.md.