Closed-loop generative inverse design of lattice structures via reinforcement learning

Lattice structures are widely used in advanced engineering because of their lightweight characteristics and tunable mechanical properties. However, precise generative inverse design remains challenging because of the discrete combinatorial design space and nonlinear physical responses. Existing generative approaches often operate open-loop and focus on reproducing statistical distributions, making compliance with prescribed engineering targets difficult. To address this limitation, a performance-feedback-guided generative inverse-design framework integrates reinforcement learning (RL) with a physics-constrained conditional variational autoencoder (PC-VAE). The PC-VAE provides a continuous latent representation of discrete lattice topologies, while a pretrained surrogate supplies immediate performance feedback. The inverse-design stage is formulated as a one-step goal-conditioned contextual decision problem, in which a task-specific proximal policy optimization (PPO) policy learns a distribution over complete latent vectors from surrogate-based rewards. Policy updates increase the likelihood of candidates satisfying prescribed mechanical conditions while retaining alternative topologies. The framework is evaluated in six scenarios involving point-target, interval-constrained, and simultaneous two-property matching. Compared with the covariance matrix adaptation evolution strategy (CMA-ES), PPO achieves comparable target-hit rates and consistently higher candidate uniqueness and topological diversity. Finite-element (FE) analyses of 48 prespecified validation structures were completed successfully. Median absolute surrogate-to-FE discrepancies were 3.03% for energy absorption and 3.88% for peak stress, with 87.50% of discrepancies within 10% for each property. For a given target, the method generates multiple topologically distinct structures with comparable mechanical responses, capturing the one-to-many relationship between prescribed properties and lattice topology. This study provides a flexible framework for target-directed generation of topologically diverse lattice candidates.