Open Access Open Access  Restricted Access Subscription Access

ADAPTIVE-WEIGHTED PHYSICS-INFORMED NEURAL NETWORK FOR MULTI-PHASE GUIDED PROJECTILE EXTERIOR BALLISTIC TRAJECTORY SOLVING

Ziyang Zhang, Jinquan Chen, Xugang Wang, Zhongyuan Wang

Abstract


This paper establishes a physics-informed neural network (PINN) framework for solving the multi-phase exterior ballistic initial value problem of guided projectiles, without relying on pre-generated trajectory datasets. Smooth sigmoid functions are employed to unify thrust switching and angle-of-attack regulation across five flight sub-phases into a single continuously differentiable system. To address the optimization imbalance caused by disparate gradient scales among the four coupled equation components, a gradient-norm-based adaptive weighting mechanism is introduced to construct an adaptive-weighted PINN (AW-PINN), which dynamically adjusts component loss weights based on exponentially smoothed gradient norms. Under identical network architecture and training conditions, AW-PINN is compared with standard PINN and residual-adaptive sampling PINN (RAS-PINN), using independently computed RK45 numerical solutions as the evaluation benchmark. Results show that AW-PINN reduces full-trajectory RMSE by approximately 85–92% relative to standard PINN across all state variables, achieving the highest overall accuracy among the three strategies.


DOI
10.12783/ballistics26/37594

Full Text:

PDF

Refbacks

  • There are currently no refbacks.