Physics-knowledge-integrated neural network for quantum transport prediction of field-effect transistors

Quantum-mechanics-based transport calculations play an important role in the design of nanoelectronic field-effect transistors (FETs), but their computational intensity is the primary challenge. To address this, we developed a physics-knowledge-integrated neural network framework that directly predicts quantum transport current-voltage curves for gate-all-around FETs. This model significantly accelerates FET design optimization while maintaining the density-functional theory--level accuracy. By integrating key physical parameters of the channel materials, our model achieves a mean absolute error as low as 0.39. In particular, 98% of the current prediction residuals are within one order of magnitude, even with a limited training dataset. The differentiable nature of the neural network allows for automatic differentiation analysis, providing interpretable insights into our model. This ensures that the model has successfully learned quantum transport physics, thereby enhancing its reliability. Using our model, we efficiently identified symmetric p-type gate-all-around FETs with performance comparable to n-type ones, which are crucial for developing homogeneous complementary metal-oxide semiconductor circuits. This work provides an effective and reliable method for rapidly screening and optimizing FET designs, paving the way for accelerating the design process of next-generation electronic devices.