This study predicts the thermoelectric figure of merit (ZT) for defective gamma-graphyne nanoribbons (γ-GYNRs) using binary coding, convolutional neural networks (CNN), long short-term memory networks (LSTM), and multi-scale feature fusion. The approach accurately predicts ZT values with only 500 initial structures (3% of 16,512 candidates), achieving an R2 above 0.91 and a mean absolute error (MAE) of 0.05 to 0.06. The use of artificial feature extraction combined with an attention mechanism reveals that the number and distribution of defects are crucial for achieving high ZT values. γ-GYNRs with moderate and evenly distributed defect count show superior thermoelectric performance. This demonstrates the effectiveness of neural networks in designing low-dimensional materials like γ-GYNRs and offers insights into exploring other materials with excellent thermoelectric properties.
Keywords: Attention mechanism; Convolutional neural networks; Graphyne nanoribbons; Long short-term memory networks; Multi-scale feature fusion; Thermoelectric figure of merit.
© 2025. The Author(s).