Prediction Method for Drop Impact Dynamics Response of Packaging Based on Hybrid Physics-Guided Neural Network

Addressing the poor physical interpretability and insufficient extrapolation capability of pure data-driven models in predicting drop impact dynamic responses of corrugated-box cushioning packaging, this paper proposes a Hybrid Physics-Guided Neural Network (Hybrid PGNN) that embeds domain-specific mechanical shape priors and adopts a decoupled amplitude–waveform architecture to circumvent the absence of explicit governing PDEs. A dataset comprising 261 measured impact acceleration waveforms at five drop heights was constructed. The network introduces Fourier positional encoding and residual connections to capture high-frequency impact features, and integrates five mechanical constraint terms—derivative constraints at impact peaks, interval monotonicity, and boundary conditions—into a composite physics-prior loss. A decoupled prediction scheme locks the power-law amplitude calibration while the network learns the normalized waveform shape, maintaining physical rationality outside the training domain. On the 110 cm extrapolation test, Hybrid PGNN achieves an R 2 of 0.92 and a peak prediction error of 0.22%, while MLP and Transformer exhibit strong high-frequency fitting within the interpolation domain but weaker extrapolation. Ablation experiments confirm that physical constraints suppress waveform distortion and non-physical oscillations common in out-of-domain prediction. The extracted power-law relation accurately reveals nonlinear softening and energy-dissipation characteristics of corrugated materials. The proposed method provides a physically faithful and extrapolation-reliable simulation tool for anti-drop impact design of packaging structures. The model applies to the tested three-ply B-flute box with a 2.0 kg internal rigid wooden frame under face-drop impact on a rigid floor; other configurations require recalibration.