Fig. 3: Deep learning-aided data regression.

a Illustration of the sensor deformation as the attached elastomer bends due to external stimuli. b The relationship between the three parameters (i.e., stimulus magnitude, substrate hardness, and sensor output), where separate colors are used for the hardness of the distinct substrates. c The overall structure of our model shows a stacked hierarchy with initial and final dense layers, and five block operations (details shown in c). Input for block operation: ‘Input 1’ from the feature maps of the previous layer, and ‘Input 2’ from the feature maps of all the preceding layers, except ‘Input 1’. d Conventional implementation of block operation, where each group consists of divided input 1 and whole input 2 for the group dense operation. e Customized implementation of block operation, where input 1 and input 2 are placed in group dense and general dense operation, respectively. Same size outputs from each operation are added for the final output. f Epoch-loss graph of the training and validation sets for the prediction of substrate hardness from the given sensor output and mechanical stimulus. g Epoch-loss graph of the training and validation sets for the prediction of sensor output from the given hardness and mechanical stimulus. h Scatterplot of the true and prediction value of the test set for substrate hardness. i Scatterplot of the true and prediction value of the test set for sensor output.