Deep Learning-Driven Intelligent Prediction for Tailoring Electrical Properties of N2200-Based Donor-Acceptor Conjugated Copolymer OFETs
Organic field-effect transistors (OFETs) based on N2200 donor-acceptor copolymer were fabricated with top-gate bottom-contact architecture. Parametric optimization revealed that a channel length of 150 μm enhances current density while maintaining leakage control. Optimal N2200/poly(methyl methacrylate) (PMMA) concentration ratio of 7/100 mg/mL and annealing at 80 °C for 3 h improved crystallinity and interfacial properties, yielding stable electrical performance. A dataset of 719 experimental data points, surpassing typical TCAD-generated datasets, was used to train convolutional neural network (CNN), back propagation neural network (BPNN), and random forest (RF) models. The CNN achieved R² > 0.9 for all metrics, with R² Vth = 0.95 and R² SS = 0.96. A novel CNN-particle swarm optimization (PSO)-BP hybrid architecture further reduced mean absolute error (MAE) and root mean square error (RMSE) by 15.7% and 14.9% for Vth, 10% and 9% for lg(Ion/Ioff), and 13.5% and 9.5% for SS, respectively. Residual analysis showed that the CNN-PSO-BP model produced the most compact residual distribution, effectively mitigating overfitting and underfitting. This machine learning framework enables autonomous extraction of physical characteristics without predefined formulas, offering a robust pathway for high-throughput device performance tuning.