In this study, we present a coupled, dimensional energy-balance model enhanced with machine-learning validation to predict residual-velocity curves and ballistic limits of fiber-reinforced composites. Projectile deceleration is described as a three-term balance involving strength-like, drag-like, and inertial effects, mapped to the nondimensional groups Π₀, Π₁, and Π₂; closed-form and RK4 solutions yield residual velocity and regime boundaries (Π₀ = Π₁, Π₁ = Π₂). Validation against six literature datasets (CFRP and aramid laminates; Vr–V0 curves) shows high accuracy: median R2 = 0.93–0.96 and typical RMSE = 10–30 m·s⁻1, with best case R2 = 0.976 and RMSE = 6.99 m·s⁻1 for thin CFRP. Ballistic-limit predictions accurately capture the nonlinear increase with thickness, with errors less than 1 m·s⁻1 in brittle CFRP and up to 10 m·s⁻1 in Kevlar laminates. A global master curve of wr = Vr/V0 versus ∥Π∥2 collapses all data and shows a consistent trend. Energy-budget analysis quantifies the contributions of the three terms: the strength term Π₀ dominates in about 90% of operational points, while drag-like effects are minimal and inertial effects only appear at thick or high-velocity limits; the dominance fractions and combined contributions support these shifts. The (V₀, h) regime map, derived by setting Π₀ = Π₁ and Π₁ = Π₂, separates design-relevant domains and aligns with observed transitions in Vr–V0 modes and slopes. An independent machine-learning check using Random Forests achieves R2 = 0.992, RMSE = 17.5 m·s⁻1, and MAE = 12.4 m·s⁻1 (fivefold cross-validation: R2 = 0.835 ± 0.145), supporting the mechanistic hierarchy through feature importance. The integrated physics-based model and machine-learning analysis provide traceable parameters (α, β, γ), uncertainty bounds, and practical screening maps for composite and geometric options under high-velocity impact.

Coupled Dimensional Energy Balance and Machine Learning Validation for Ballistic Response Prediction of Fiber Composites

期刊论文
发表日期：

2025年11月27日

查看全文

Bertan Beylergil,
Hasan Ulus,
Mehmet Yildiz

摘要

In this study, we present a coupled, dimensional energy-balance model enhanced with machine-learning validation to predict residual-velocity curves and ballistic limits of fiber-reinforced composites. Projectile deceleration is described as a three-term balance involving strength-like, drag-like, and inertial effects, mapped to the nondimensional groups Π₀, Π₁, and Π₂; closed-form and RK4 solutions yield residual velocity and regime boundaries (Π₀ = Π₁, Π₁ = Π₂). Validation against six literature datasets (CFRP and aramid laminates; V_r–V₀ curves) shows high accuracy: median R² = 0.93–0.96 and typical RMSE = 10–30 m·s⁻¹, with best case R² = 0.976 and RMSE = 6.99 m·s⁻¹ for thin CFRP. Ballistic-limit predictions accurately capture the nonlinear increase with thickness, with errors less than 1 m·s⁻¹ in brittle CFRP and up to 10 m·s⁻¹ in Kevlar laminates. A global master curve of w_r = V_r/V₀ versus ∥Π∥₂ collapses all data and shows a consistent trend. Energy-budget analysis quantifies the contributions of the three terms: the strength term Π₀ dominates in about 90% of operational points, while drag-like effects are minimal and inertial effects only appear at thick or high-velocity limits; the dominance fractions and combined contributions support these shifts. The (V₀, h) regime map, derived by setting Π₀ = Π₁ and Π₁ = Π₂, separates design-relevant domains and aligns with observed transitions in V_r–V₀ modes and slopes. An independent machine-learning check using Random Forests achieves R² = 0.992, RMSE = 17.5 m·s⁻¹, and MAE = 12.4 m·s⁻¹ (fivefold cross-validation: R² = 0.835 ± 0.145), supporting the mechanistic hierarchy through feature importance. The integrated physics-based model and machine-learning analysis provide traceable parameters (α, β, γ), uncertainty bounds, and practical screening maps for composite and geometric options under high-velocity impact.