The growing demand for sustainable composite materials has promoted interest in repairable architectures capable of extending service life and reducing end-of-life waste. Vitrimers, a class of covalent adaptable networks (CANs), provide intrinsic repair and reprocessability; however, their use as full composite matrices is still limited by cost and processing complexity. A promising alternative is represented by multi- matrix composites, where vitrimers are selectively integrated only in regions requiring repair functionality. This work investigates the delamination behavior and repair capability of an epoxy–vitrimer multi-matrix glass-fiber composite. Through-thickness bi-matrix laminates were manufactured using three curing strategies—Sequential Vacuum, Sequential Press, and Co-Curing—to identify the route providing the best interfacial performance. Results show that Co-Curing significantly enhances Mode I interlaminar fracture toughness, reaching values close to fully vitrimer-based composites. This improvement is associated with stronger interfacial bonding and enhanced energy-dissipating mechanisms. Healing efficiency was evaluated by re- testing specimens after hot-press repair cycles, analyzing the effects of temperature, time, and pressure on fracture toughness retention. The dominant healing mechanism was identified as mechanical interlocking, driven by vitrimer penetration into epoxy surface roughness. Although toughness recovery remains partial, the proposed architecture enables damage-mitigating repair while reducing manufacturing and material costs.
Innovation in the coupled design of unit-cell geometry and cell wall-level architecture can be crucial for multifunctional metamaterials, enabling simultaneous programming of stiffness, strength, and energy absorption beyond conventional straight-beam lattices. Here, we propose an anti-curvature strategy that retains the typical advantages of bending-dominated architectures, such as high energy absorption, while also improving effective stiffness and strength that are normally associated with stretching-dominated lattices. Building on this premise, we integrate physics-informed curvature-guided geometric design of ligament profiles with machine-intelligence for efficient predictive modelling. Nonlinear finite element simulations are performed to quantify the influence of curvature direction under different modes of loading, including compression, tension, and shear. The numerical results reveal significant curvature-dependent enhancements without any notable increase in weight: up to 51.56% in stiffness, 60.42% in onset-of-failure stress, and 57.80% in energy absorption under mode-dependent normal loading. Additionally, anti-curvature designs under shear can improve effective stiffness, onset-of-failure stress and energy absorption by up to 66.72%, 81.26% and 91.67%, respectively. The coupled design space of physics-informed cell-wall curvature and unit cell geometry, augmented by a machine learning-assisted efficient predictive modelling, provides a scalable route to engineer lightweight lattices with superior specific strength, tunable stiffness, and enhanced energy absorption for multifunctional applications
This paper proposes a feature-attention-enhanced long short-term memory (FA-LSTM) framework for data-driven prediction of the complete nonlinear in-plane buckling response of parabolic deep arches made of homogeneous, laminated-composite, and functionally graded materials (FGMs). A unified design parameter vector integrating equivalent sectional stiffnesses, geometric parameters, and non-dimensional elastic support coefficients is formulated to represent the three material systems within a single feature space. A multi-head feature self-attention module is employed to fuse the static design parameters with the time-varying load–displacement state at each load step, enabling the network to adaptively shift its reliance from geometry- and material-dominated features in the pre-buckling regime to displacement- and load-dominated features as instability approaches. A joint loss function combining stepwise-increment and absolute-state supervision, together with a dual-head decoder and a curriculum-based training strategy, is adopted to suppress error accumulation during autoregressive prediction. Trained and tested on 2513 nonlinear finite element (FE) simulations and evaluated over five independent runs, the proposed model achieves a full-sequence coefficient of determination (R2) of 0.9920 ± 0.0041 and a buckling-load R2 of 0.9919 ± 0.0074, while reducing the computational cost by approximately two orders of magnitude compared with nonlinear FE analysis. The framework provides an efficient surrogate tool for the buckling analysis and design of parabolic deep arches.
