Data-Driven Accelerated Discovery of LNMCO Cathodes Materials via 37-Dimensional Parameter Space Mining With a Cascaded Neural Network.

Abstract

The vast, unexplored synthesis space of LNMCO cathode materials contains potential solutions to the long-standing trade-off between energy density and stability. To navigate this high-dimensional space, a predictive tool capable of accurately mapping the complex relationships between synthesis parameters and electrochemical performance is essential. Here, a cascaded neural network (CaNN) architecture was designed to simulate the material synthesis workflow. This model synchronously maps 37 process dimensions within the composition-processing-structure-property-performance (CPSPP) paradigm. By utilizing a cascaded structure, the output of upstream prediction tasks serves as the input for downstream tasks, enabling the capture of hierarchical dependencies that govern material properties. This design choice proved highly effective, achieving superior predictive accuracy with an overall coefficient of determination (R²) of 0.85. Furthermore, SHAP analysis was integrated to open the model's "black box," demonstrating a mechanism-informed approach where predictions align closely with the underlying physical laws of structural inheritance. Experimental validation of nine candidates spanning diverse compositions and synthesis routes confirms the predictive accuracy of the strategy (prediction errors < 10%), establishing a "CaNN modeling → Latin Hypercube Sampling → Wa Screening → Experiments validation" framework for the accelerated discovery of high-performance cathode materials.