Long short-term memory (LSTM) networks for precision prediction of Schottky barrier photodiode behavior at different ıllumination levels.

This study has focused on modeling and predicting the electrical properties and parameters of CdZnO interlayered Al/p-Si Schottky Diodes (SDs) using the Long Short-Term Memory (LSTM) algorithm. The primary aim of this study was to develop a robust predictive model that accurately captures how dopant concentration and illumination levels influence the electrical behavior of SDs. Using the temporal gating and memory capabilities of the LSTM architecture, we proposed a time- and cost-efficient alternative deep-learning model to extensive experimental procedures, ensuring that the diode characterization process could be accelerated without compromising accuracy. The dataset comprises a combination of three Al/CdZnO/p-Si SDs containing different Cd dopant ratios (10%, 20%, and 30%) and five different levels of illumination (50, 100, 150, 200, and 250 mW/cm). Predictions for electrical parameters, including ideality factor (n), barrier height (F), and series resistance (R), were conducted using the traditional I-V method, Cheung's analysis, and Norde's method. To evaluate the LSTM model predictions, one diode at a specific illumination level was selected as the test set. At the same time, the remaining dataset was divided into 80% for training and 20% for validation. The optimization algorithm was selected as Adaptive Moment Estimation (Adam), and the root mean squared error (RMSE) served as the loss function. Hyperparameters, including the number of epochs (150) and batch size (64), were determined empirically to balance computational efficiency and model performance. Results indicate that predictions on Diode 1 demonstrate strong performance at 50, 100, and 150 mW/cm illuminations, exhibiting RMSE values of 11.5, 7.2, and 11 mA, respectively, and R² values exceeding 0.98. LSTM shows on Diode 2 consistently lower errors, achieving a minimum RMSE of 6.22 mA at 100 W (R²=0.993). Diode 3 predictions elevated RMSE and mean absolute error (MAE) at both 50 and 250 mW/cm. Across Traditional I-V, Cheung's, and Norde's analyses, the LSTM model yields close agreement with experimental measurements, particularly for barrier height and ideality factor. In conclusion, the LSTM model offers a reliable, cost-effective, and time-efficient alternative to exhaustive Schottky diodes experimental measurements. By accurately capturing the nonlinear interplay of doping level and illumination in SDs, this method provides a practical way to expedite device characterization. These findings highlight the potential of data-driven deep learning approaches in semiconductor research and open avenues for broader applications of LSTM architectures in predicting electronic and optoelectronic device parameters.