Figure 1

Convolutional neural network architecture to infer microstructural battery parameters. The CNN is comprised of convolution blocks and fully connected layers, which takes two types of input at different stages. The model takes the color-encoded voltage versus capacity curves as the main input (each color corresponding to a current density), energy density, E, and power density, P, as the second input. Each convolution block has two convolutional layers, followed by a pooling layer. A ReLU activation function is placed after each convolutional layer and hidden dense layer. For each data point, the image with voltage curves are fed into the network. For each curve, E and P are taken into the following fully connected layers, along with the higher-level representation of the input image. The output of this network has two components, the Bruggeman exponent, \(\alpha\), and the area density shape factor, S. See text for details.