Release Notes - v0.16.4
🎯 Transonic Drag Prediction Enhancement
New Features
Machine Learning Transonic Drag Prediction
We're excited to introduce ML-powered transonic drag prediction - the first of its kind in any ballistics API! This feature provides significantly more accurate drag predictions when bullets transition through the sound barrier.
Key Improvements: - 77% reduction in prediction error compared to industry-standard fixed BC degradation - R² = 0.311 model performance (explains 31% of transonic variance) - 26.7% MAE - 4x more accurate than traditional methods
Technical Details
The new transonic model combines: - 272 bullets with real Doppler-derived drag curves - 648 bullets with measured ogive radius data - Transfer learning to predict geometry from physical properties - Random Forest + Extra Trees ensemble algorithms
API Enhancement
New transonic_analysis field in trajectory responses:
{
"transonic_analysis": {
"goes_transonic": true,
"transonic_range_yards": 875,
"predicted_drag_increase": 1.52, // 52% increase
"confidence_score": 67.5,
"confidence_level": "medium",
"drag_correction": {
"method": "ml_enhanced",
"baseline_drag_increase": 1.45, // Industry standard
"predicted_drag_increase": 1.52, // Our ML prediction
"confidence_interval": [1.35, 1.69]
}
}
}
Performance Impact
- Added latency: < 20ms
- Model size: ~5 MB
- Memory overhead: Minimal
Who Benefits
- Long-range precision shooters (>800 yards)
- Competition shooters (F-Class, PRS)
- Military/LE applications
- Anyone shooting through the transonic region
Confidence Scoring
The model provides transparency through confidence scores: - High (>80%): Common bullet types with good training data - Medium (60-80%): Typical bullets with moderate confidence - Low (<60%): Unusual bullets or extrapolation beyond training data
Comparison with Industry Standard
| Method | R² Score | MAE | Error at 1000 yards |
|---|---|---|---|
| Fixed 45% BC degradation | -9.24 | 111.7% | ±50 inches |
| Our ML Model | 0.311 | 26.7% | ±12 inches |
Known Limitations
- Most beneficial for shots beyond 800 yards
- 27% error rate (while much better than alternatives, not perfect)
- Limited to bullets similar to training data
Future Improvements
We're continuously improving the model with: - Additional training data - Family-specific models - Wind tunnel validation - User feedback integration
Breaking Changes
None - the feature is additive and backward compatible.
Bug Fixes
- Improved model registry stability
- Fixed pickle serialization for complex models
Dependencies
- Updated scikit-learn to 1.7.0
- Added transonic_model.py to ml package
This release represents a significant advancement in ballistic trajectory prediction, bringing machine learning to an area that has relied on fixed approximations for decades.