The predictions are accurate, as demonstrated by the bulk of data points falling on the diagonal. The prediction’s standard deviation varies based on the target quantity, and the prediction intervals are realistic.
b) confusion matrix for topology This confusion matrix shows a perfect reconstruction of all test points, as all predicted topologies match the original topologies.
To facilitate human interpretation, the latent space is assigned to be two-dimensional. The loss term weights are set to w_1 = 1
, w_2 = 3
, w_3 = 0.1
, and w_4 = 0.01
. The effect of w_3
can be observed by visually inspecting the latent space and coloring the points based on their y value. If the distribution is not normal, w_3
should be increased. Similarly, if there are correlations between z
and y
, w_4
should be increased.
Visualization of the latent space, with the points colored according to their respective cost:
a) density
b) embodied carbon
Visit this Weights & Biases project page for interactive details about the hyperparameters of the trained model.
Engineering assumptions for the case study
Steel material properties:
Timber material properties: