v0.23.0

📅 October 05, 2025 👤 Alex Jokela 🔖 bc7c67d
🎫 Tickets: MBA-42

Breakthrough in gyroscopic stability prediction using transfer learning - a machine learning approach that combines physics-based calculations with data-driven corrections. Achieves 77.7% improvement over the traditional Miller formula while maintaining the ability to generalize to any bullet, even those never seen during training.

Changes

✨ Features

  • Transfer learning stability model (77.7% improvement over Miller)
  • Weighted ensemble (RF + GB + XGB) with uncertainty estimation
  • Production API: predict_minimum_twist()
  • Enhanced stability calculator with confidence metrics
  • Inertia ratio integration (272 bullets with measured tensors)

📝 Improvements

  • Graceful degradation to Miller for unseen bullets
  • Confidence-based blending (high/medium/low)
  • Physics-informed features (not just raw geometry)
  • Works for 164 calibers (trained on 14)

📝 Model_Performance

  • MAE: 1.45" (vs Miller: 6.52")
  • MAPE: 14.4% (vs Miller: 58.2%)
  • Correction Factor R²: 0.927
  • Uncertainty-aware: 33% high, 33% medium, 33% low confidence

📝 Research

  • Compared 5 model architectures (RF, GB, XGB, Stacked, Weighted)
  • Investigated AB Analytics CDM extraction (not used - quality/legal issues)
  • Explored synthetic inertia prediction (failed - >20% error)
  • Tested PINN approach (failed - R²=0.007)