Research

Materials Imaging & Integration

AI Materials Informatics

Goal

  • To utilize machine learning (ML) models that can optimize the cathode synthesis for enhanced discharge capacity, and predict the composition and cycling states of the lithium-ion batteries (LIBs)

  • To demonstrate the ability of a convolutional neural network (CNN) model to analyze computationally generated data and make predictions on experimental materials science data

Approach

  • Combine inverse design and ML model to predict optimal synthesis parameters for LIB cathode materials

  • Train a CNN model on phase field-generated images to segment experimental data of materials microstructures

  • Develop a CNN model trained on SEM images to classify the composition and cycling states of LIB cathode materials

Impact

  • Accelerate the discovery and optimization of high-energy-density LIB cathodes, reducing reliance on the traditional trial-and-error methods

  • Accelerate the process of materials design by reducing the dependency on labor-intensive manual segmentation

  • Improve the efficiency of battery production and development by minimizing human-dependent analysis and increasing the speed of materials characterization

Code Repository

Machine learning-assisted synthesis of LIB cathode materials. (a) Schematic illustration of three steps of the design-to-device pipeline, (b-d) SEM images based on the prediction of imputed datasets, (e) charge-discharge curves, and (f) comparison between the predicted and experimental target discharge capacity [link]

Segmentation of experimental datasets via CNN trained on phase field simulations. (a) Phase field output and the addition of the (b) luminance (c) pixel noise, (d) XCT test image, (e) associated overlay from the best CNN, (f) higher magnification image of the red box section and (g) associated overlay, (h) saliency map, and (i) 3D visualization of the Aluminum dendritic structure [link]

Composition and state prediction of LIB cathode via CNN trained on SEM images. (a) Example images of true cases and their grad-CAM overlays from the best-trained network, (b) probability of each class for the false case and grad-CAM overlays of top-most highest classes, and (c) results of composition and cycling state prediction of SEM images from domain experts [link]

Materials and Molecular Modeling, Imaging, Informatics, and Integration (M3I3) Project. (a) Schematic diagram of the processing–structure-property relationship and materials hierarchy for the M3I3 project, aiming to achieve the seamless integration of the multiscale “structure–property” and “processing–property” relationships. An instance of the M3I3 project research involves predicting the discharge capacity as a function of the composition of lithium-ion battery materials, as shown in Figures (b) and (c) for real and predicted discharge capacities, respectively [link]