A machine-learning surrogate framework for predicting energy absorption and initial peak force of combined lattice metamaterials across full design space

Additive manufacturing enables the design of lattice metamaterials by combining different unit-cell architectures, but exhaustive evaluation of large design spaces is computationally expensive. This study develops a machine-learning surrogate framework for lattices comprising six unit-cell types arranged at four positions. Stratified index sampling (SIS) selected configurations for finite-element (FE) analysis, and the FE model was validated through quasi-static compression experiments. Support vector regression (SVR), neural networks (NN), and ridge regression (RR) were evaluated using one-hot and scaled-relative-density (SRD) encodings with five-fold cross-validation. One-hot encoding consistently outperformed SRD encoding. Following initial model selection, the preferred models were compared with an additive ordinary least-squares (OLS) baseline. OLS was selected for energy absorption (EA), achieving a held-out R 2 of 0.991 and RMSE of 1.938 J, while SVR was retained for initial peak force (IPF), achieving R 2 of 0.955 and RMSE of 0.294 kN. EA exhibited predominantly additive behavior, whereas IPF benefited more from nonlinear modeling. Overall, the framework enables efficient exploration and application-specific selection across the full lattice design space while reducing the need for exhaustive simulations and substantially lowering computational evaluation costs.