Answers are grounded in this paper and cite their sources.
Simulation-driven inverse design problems are pervasive in engineering domains where evaluating candidate solutions requires expensive numerical solvers. This work proposes a general artificial intelligence framework for accelerating inverse design in complex engineered systems by coupling a forward surrogate model with an autoencoder-based inverse neural network. The framework addresses the inherent one-to-many nature of inverse mappings through a distribution-aware data generation and filtering pipeline, combined with an information-theoretic evaluation metric based on entropy and Kullback–Leibler divergence. The methodology is validated using a challenging nanophotonic case study involving plasmonic color generation from nanostructured metallic surfaces. A forward deep neural network is trained to approximate the mapping from geometric parameters to chromaticity coordinates, while an inverse deep neural network estimates feasible geometric configurations from target colors using a pretrained encoder–decoder architecture. The models are trained using 49,532 synthetic samples generated through numerical simulations and evolutionary optimization. Experimental results show that the proposed framework achieves an overall pipeline normalized root mean squared error (NRMSE) of 11.63%, an overall pipeline distributional similarity of 84.31%, and a reduction in inference time of 99.27% compared with simulation-based design. The results demonstrate the potential of the proposed artificial intelligence (AI) framework as a practical surrogate for real-time and large-scale inverse design in engineering applications. Limitations related to extrapolation and reliance on synthetic data are discussed, and directions toward physics-informed learning and uncertainty-aware design are outlined.
No takes yet. Share an insight, caveat, or question.
Rivas-Avila et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: