Temperature-aware learning of current–voltage behavior in Schottky diodes: From classical machine learning to neural basis-expansion models under a leave-one-temperature-out validation protocol

We present a leakage-averse and reproducible benchmark for temperature-dependent current–voltage (I–V) prediction in Al/CdZnO/p-Si Schottky diodes. Using a unified Leave-One-Temperature-Out (LOTO) protocol, we hold out one temperature per fold to assess prediction at held-out temperatures within the studied Al/CdZnO/p-Si device system. We compare classical regressors, including support vector regression (SVR), multilayer perceptron (MLP), and random forest (RF), with LSTM-Many2One, CNN-LSTM, Attention-LSTM, and N-BEATS. We also include three benchmark methods to contextualize this comparison: nearest-temperature transfer, two-dimensional polynomial-ridge response-surface regression, and a thermionic-emission model incorporating series and shunt resistances. The seven data-driven models use sliding windows of 30 consecutive voltage points with the target temperature appended, whereas the three benchmark methods use method-specific direct or physically parameterized representations. Across the 11 temperature folds, N-BEATS yielded the lowest descriptive mean errors, with MSE=0.263±0.28, RMSE=0.461±0.236, and R 2 =0.989±0.008. MLP followed closely, while LSTM-Many2One remained competitive but more variable; CNN-LSTM performed poorly in this data regime. The benchmark methods yielded mean RMSE values from 0.755±0.167 to 1.091±0.417, with the thermionic-emission R s –R sh model producing the lowest benchmark-method error. N-BEATS yielded the lowest descriptive mean relative errors for the ideality factor and barrier height. Reconstructed interface-state density profiles also showed qualitative agreement with the experimental trends across the representative models.