Answers are grounded in this paper and cite their sources.
The prediction of key properties of cathode materials using machine learning can significantly accelerate the development of nextgeneration secondary batteries with enhanced performance, safety, and reduced costs. In this work, we present an application of the GeoCGNN graph neural network for predicting the redox potential of cathode materials for Na-ion batteries. We analyze three different approaches for the prediction of redox potential, and evaluate which one is the most accurate: (i) the single crystal direct prediction, which considers only the discharged material; (ii) the double crystal direct prediction, which uses both discharged and charged materials; and (iii) the indirect prediction, where the redox potential is derived from the formation energy predictions of discharged and charged materials. Training, testing and validation are performed using data from the Materials Project database. We apply the trained GeoCGNN graph neural network to study sodium cathode material based on layered transition metal (M) oxide NaMO2. Our model predicts compositions that yield high redox potential and low formation energy. We validate our approach by comparing predictions with calculations based on density functional theory. We found that the double crystal approach is better by the mean absolute error, while for the NaMO2 crystals, the indirect approach outperforms. Moreover, we found that the prediction performance of each approach depends on the amount of Ni or Ti. Finally, we define a revised predictor that aligns the outputs on non-relaxed crystals with those obtained from the first-principles relaxed structures, ensuring consistency with the most accurate prediction.
No takes yet. Share an insight, caveat, or question.
Catillo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: