Predictive modeling of electrochromic performance in ammonium metatungstate solutions using machine learning algorithms

This research explores the application of machine learning (ML) in the domain of electrochromic (EC) technology, focusing specifically on liquid-state electrochromic devices (ECDs). Unlike traditional solid-state ECDs, liquid devices offer a simpler structure, reducing manufacturing variables and potentially improving prediction accuracy with minimal input data. Two types of ECDs were developed using solutions of ammonium metatungstate-iron(II) chloride and ammonium metatungstate-iron(II) sulfate, resulting in 20 different devices with varying concentration gradients. Transmittance alterations under different current densities were measured to determine modulation range and time response, serving as training data for ML models. Seven regression models were employed to construct EC models and predict optimal device solutions. Subsequent manufacturing and testing of new ECDs validated the predictions, with a comparative analysis of EC characteristics and model fitting performance conducted between the two types of ECDs. For ammonium metatungstate-iron(II) chloride ECDs, under a 5 mA applied current, the maximum optical modulation reached 23.67%, with a coloration efficiency of 17.54 cm2/C (under 700 nm). For ammonium metatungstate-iron(II) sulfate ECDs, under a 5 mA applied current, the maximum optical modulation reached 18.92%, with a coloration efficiency of 17.05 cm2/C (under 700 nm). The coloring time (tc) and bleaching time (tb) for ammonium metatungstate-iron(II) chloride ECDs were ∼14 and 8 s, respectively. The predicted maximum optical modulation for ammonium metatungstate-iron(II) chloride and ammonium metatungstate-iron(II) sulfate ECDs were 23.67% and 18.92%, respectively, with prediction accuracies reaching 97.90% and 96.97%, respectively. Decision tree regression (DTR) and kernel ridge regression (KRR) emerged as the most effective ML methods for these ECDs.