Quantitative analysis of LiF in solid electrolyte interphases using AMDG-1D CNN combined with far-infrared spectroscopy

Quantifying LiF in the solid electrolyte interphase (SEI) is critical for developing novel interface engineering and understanding battery degradation. However, conventional techniques (XPS, FTIR) face severe limitations in quantifying LiF due to shallow detection depth or low sensitivity. Far-infrared (FIR) spectroscopy offers a potential advantage via the characteristic lattice absorption of LiF, yet its application is hindered by broad, strongly overlapping peaks and the lack of reliable concentration labels in real SEI samples. To overcome these challenges, inspired by the phenomenon that metastable SEI components can be transformed into stable products in a high-humidity atmosphere, this work constructs a simplified simulated SEI mixture and systematically collects its FIR spectra. A novel one-dimensional convolutional neural network, AMDG-1D CNN, is developed for LiF quantification, integrating adaptive preprocessing, multi-scale feature extraction, dual attention refinement, and global multi-statistics pooling. The model achieves a test R² of 0.904 and RMSE of 3.706 wt%, outperforming conventional methods. It also shows strong robustness under various interferences. When applied to real SEI samples, predicted LiF concentrations show small deviations from ion chromatography measurements, validating the transferability from simulated to practical systems. This FIR-deep learning framework provides a reference strategy for LiF quantification in complex electrochemical interfaces.