Abstract
Electrochromic devices have demonstrated considerable potential in a range of applications, including smart windows and automotive rearview mirrors. However, traditional cycle life testing methods are time-consuming and require significant resources to process a substantial amount of generated data, which presents a significant challenge and remains an urgent issue to be addressed. To address this challenge, we proposed the use of Long Short-Term Memory (LSTM) networks to construct a prediction model of the cycle life of electrochromic devices and introduced an interpretable analysis method to further analyze the model's predictive capabilities. The original dataset used for modeling was derived from preliminary experiments conducted under 1000 cycles of six devices prepared with varying mixing ratios of heavy water (D(2)O). Furthermore, validation experiments confirmed the feasibility of the D(2)O mixing strategy, with 83% of the devices exhibiting a high initial transmittance modulation amplitude (ΔT = 43.95%), a rapid response time (t(c) = 7 s and t(b) = 8 s), and excellent cyclic stability (ΔT = 44.92% after 1000 cycles). This study is the first to use machine learning techniques to predict the cycle life of electrochromic devices while proposing performance enhancement and experimental time savings for inorganic all-liquid electrochromic devices.