Fig. 3: Nonlinear buckling resistance prediction using DNN.
From: Inverse design of truss lattice materials with superior buckling resistance
a Predicted vs. target (simulated) normalized buckling strength values (on test data, never seen by the DNN). The number of times a value occurs is indicated as Counts (90 bins are selected in the 2D histogram). b Probability density of the percentage relative error between predicted and test data. A Gaussian density is assumed. c MAPE and accuracy values on test data plotted against training density. Lines are plotted only to facilitate the interpretation. d Predicted vs. target (simulated) normalized mean buckling strength values, averaged over the loading directions. e Probability density of the percentage relative error between predicted and average data. A Gaussian density is assumed. f MAPE and accuracy values on average data plotted against training density. One run, optimized network’s hyperparameters and training density of 80% is considered, if not otherwise specified. Lines are plotted only to facilitate the interpretation. Error bars correspond to one standard deviation computed over 5 runs.
