(169y) Li Ion Diffusion in Solid Electrolyte Analyzed Using Deep Generative Models: Dependence of Accuracy of Diffusion Coefficients on MD Data Length.

Molecular Dynamics (MD) are effective in providing insights at the atomic level, However, the computational costs are notably high, particularly ab-initio MD (AIMD). Even as parallelization techniques have expanded the spatial scale of MD, the extension of the temporal scale remains a persisting challenge.

New computational approaches, particularly those incorporating machine learning like MD-GAN [1,2,3], are gaining traction for accelerating computation. MD-GAN, based on Generative Adversarial Networks (GAN), has the potential to generate long-term data from short-term data, thus enabling faster computation through temporal parallelization. To minimize computational costs, it is essential to keep the duration of MD calculations, utilized as training data, as short as possible. Nevertheless, excessively short MD calculations may result in an adequate number of samples, thereby impeding the attainment of the requisite accuracy for estimating physical quantities through training. In the previous study [4], we investigated the accuracy and optimal conditions of MD-GAN when applied to Li ion diffusion systems obtained using AIMD with SIESTA code [5]. In this presentation, we examine the effects on accuracy as the length of input data is shortened, aiming to determine the optimal number of MD calculation steps for MD-GAN training.

1. Endo, Tomobe, and K. Yasuoka, Proceedings of the AAAI Conference on Artificial Intelligence, 32, 2192 (2018).
2. Kawada, R., Endo, K., Yasuoka, K., Kojima, H. and Matubayasi, N., Journal of Chemical Information and Modeling, 63, 76-86(2022).
3. Kawada, R., Endo, K., Yuhara, D. and Yasuoka, K, Soft Matter, 18, 8446-8455(2022).
4. Fukuya, H. Nitta T. Ozawa, and K. Yasuoka , AIChE 2023 Annual Meeting