Lamination parameter informed encoder–decoder framework for stiffness tailoring in composite lattices

Composite lattices, periodic architectures of interconnected fibre-reinforced members, offer exceptional stiffness-to-weight ratios critical for aerospace light-weighting and net-zero emission strategies, yet their directional property tailoring remains underexploited due to design complexity. Composite lattice optimisation confronts a fundamental combinatorial challenge: the design space grows exponentially with ply count, requiring exhaustive searches across ply moduli, fibre orientations, layup sequences, and geometric parameters that rapidly become intractable for multifunctional structural tailoring. We introduce a physics-informed encoder–decoder architecture that circumvents this dimensional curse by projecting the discrete optimisation problem into a continuous latent space of lamination parameters. The encoder, implemented as a feed-forward neural network, maps target homogenised elastic moduli ( E x , E y ) to member-level lamination parameters, while the decoder embeds closed-form analytical expressions from classical laminate theory and homogenisation mechanics. This architecture eliminates data dependency and ensures physical consistency throughout the inverse design process. By minimising reconstruction loss between predicted and target properties, the framework learns physically realisable lamination parameter combinations that simultaneously satisfy stiffness targets and manufacturing constraints for balanced, symmetric laminates. Applied to hexagonal composite lattices, the method reveals that maximum achievable anisotropy depends not only on fibre-matrix modulus contrast but fundamentally scales with lattice effective stiffness, establishing material-dependent upper bounds ranging from 1.03 for E-glass to 1.49 for high-modulus carbon systems. The framework enables rapid exploration of previously inaccessible design regions, reduces inverse design from iterative numerical optimisation to direct analytical mapping, and provides a generalisable template for integrating domain-specific physics with machine learning in structural design.