Machine learning prediction and data-driven inverse design of gradient chiral petal honeycombs

This study investigates the structural design, performance prediction and inverse design of gradient chiral petal honeycombs. Firstly, based on the parametric modeling method, four gradient configurations, concave, convex, increasing, and decreasing, are designed to reveal the deformation evolution mechanism under the coupling effect of stiffness gradient and chiral topology, elucidating the regulatory influence of different gradient patterns on shape-morphing curves. Secondly, a grouped parallel, feature-enhanced ensemble multi-output machine learning framework is constructed to achieve high-precision forward prediction ( R 2 > 99 % ) from 63-dimensional geometric parameters to 24-dimensional mechanical properties (including shape-morphing curves, Poisson’s ratio, and macroscopic mechanical indices). A grouped decoupling inverse prediction model is also established to achieve reliable inversion from performance metrics to geometric parameters (average R 2 > 98 % ), overcoming the mapping challenge between high-dimensional nonlinear parameters and multi-dimensional properties. Finally, a two-stage optimization system integrating data-driven prediction and simulation verification is proposed, employing a progressive algorithmic strategy designed for rapid convergence and refined optimization. This system enables the inverse design of four typical target shape-morphing curves: linear, quasi-sinusoidal wave-type, cosine-like double-peak wave-type, and trapezoidal. The average relative error for all cases is controlled within 2%. The research outcomes provide methodologies and tools for the programmable deformation design of gradient honeycomb structures, facilitating a paradigm shift in structural design from empirical trial-and-error to data-driven, closed-loop intelligence.