Fig. 4
From: Machine learning enables polymer cloud-point engineering via inverse design
Inverse Design. (a) Framework of the selection criteria, where the data set is used to train GBR and NN ensemble, PSO predicts polymer design (x*) a desired CP (y*), and the design is verified for accuracy by the NN ensemble where CP agreement is a downselection criteria. (b) Illustration of the validity of the filtering procedure. We observe that given limited training data, not all extrapolated points are valid. However, when an ensemble of neural networks trained with distinct initializations agree on a certain input, then we have a much greater confidence in the validation of their predictions. (c) Final PSO-based inverse-design performance, with experimental values (orange triangles) showing an RMSE of 3.9 °C. (d) Forward model (NN ensemble) performance of the polymers synthesized from design (orange triangles)
