NEGF simulation acceleration is an important goal in nanoscale semiconductor design, where engineers need to understand quantum transport in devices with increasingly small dimensions. The nonequilibrium Green’s function (NEGF) method can model transport at this scale, but its repeated calculations of carrier density and electrostatic potential make self-consistent simulations computationally demanding. A convolutional neural network (CNN) approach offers a way to reduce that burden for double-gate MOSFETs by predicting key spatial distributions within the simulation loop.
Why nanoscale device simulation is computationally demanding
As MOSFET dimensions shrink, gate control and carrier behavior depend on detailed distributions of potential and charge within the device. Multigate architectures, including double-gate structures, use electrostatic control to influence carriers in a thin channel. Evaluating how a device behaves requires more than a simple current-voltage calculation: the spatial variation of the potential and carrier density matters.
NEGF is a quantum-mechanical transport method suited to analyzing nanoscale devices. In a typical self-consistent calculation, the simulation determines carrier density from the device potential, then uses that charge distribution in Poisson’s equation to update the potential. The process repeats until the potential converges. NEGF calculations involve repeated matrix operations and numerical integrations, so calculating charge at every iteration can become a significant part of the computational workload.
Using a CNN for NEGF simulation acceleration
The approach studied for double-gate MOSFETs uses a convolutional autoencoder to estimate two outputs from an input potential distribution: the carrier charge density and the local quantum capacitance. These quantities are represented as two-dimensional distributions across the device, with grid positions treated similarly to pixels in an image.
This spatial representation is central to the method. Convolutional networks are designed to process data with local spatial relationships, making them a natural fit for learning patterns across the channel and transport directions. The network is trained on results from conventional NEGF calculations, learning how a potential distribution corresponds to charge and quantum capacitance distributions.
Drain voltage is included as an additional input to the network. This matters because the voltage affects the source and drain carrier populations used in the NEGF calculation. During training, the potential and output distributions are normalized for the network; predicted outputs are then returned to their physical ranges before they are used in the device calculation.
Integrating predictions into the self-consistent loop
The CNN does not replace the electrostatic solution. Instead, it substitutes for the expensive NEGF calculation of charge density and quantum capacitance within the iterative process. The predicted quantities feed into the linearized Poisson equation, which provides a potential correction. The potential is updated, and the loop continues until the convergence condition is met.
- Provide the device potential and drain voltage as inputs to the trained CNN model.
- Predict charge density and local quantum capacitance for the device grid.
- Update the potential by solving the linearized Poisson equation.
- Repeat the process until the potential converges, with optional NEGF calculations to refine the result.
This workflow preserves the role of Poisson’s equation in establishing electrostatic consistency while reducing how often the full NEGF transport calculation is required. In the reported study, using the CNN in this self-consistent process produced accurate potentials in significantly less computational time than conventional NEGF calculations.
Accuracy, generalization, and practical considerations
Tests compared CNN-based self-consistent results with converged NEGF results for double-gate MOSFETs. The CNN reproduced potential distributions and current-voltage characteristics across the gate lengths examined. A model trained on several gate lengths was also applied to other lengths within the studied range, demonstrating that the learned relationship was not limited to a single device geometry. The study reports useful performance for gate lengths around 5 to 15 nanometers under its modeled conditions.
Predicted results were not identical to the NEGF reference in every case. The researchers found that additional NEGF self-consistent iterations could refine a CNN-based solution, providing a practical hybrid option: use the faster learned model to approach convergence, then use the conventional solver to improve the result. The work also notes a trade-off in training: smaller batches can improve current predictions but may make convergence less stable, while larger batches can support more stable updates at the cost of some accuracy.
These findings should be understood within the scope of the modeled device and training data. The study focused on silicon double-gate MOSFETs with specified transport and confinement directions and device assumptions. The CNN’s ability to generalize across tested gate lengths does not, by itself, establish equivalent performance for different materials, device architectures, or conditions outside the training domain.
Implications for semiconductor modeling workflows
For device researchers and simulation teams, the key opportunity is to reduce the cost of repeated self-consistent calculations without discarding the physical framework used to model transport and electrostatics. A trained CNN can serve as a rapid predictor inside a simulator, while selected conventional NEGF calculations remain available for dataset generation, validation, and solution refinement.
The practical value depends on how often simulations are run and whether the device cases of interest are represented by suitable training data. Preparing that data and training the model requires effort, but once trained, the model can be reused for subsequent calculations within its demonstrated range. This makes CNN-assisted NEGF a potentially useful complement to established quantum transport simulation—not a universal substitute for it.
Conclusion
Convolutional autoencoders can support NEGF simulation acceleration by learning to predict carrier density and local quantum capacitance from device potential distributions. Embedded in the self-consistent Poisson loop, these predictions can shorten calculation time while retaining an option for NEGF-based refinement. For engineers evaluating nanoscale MOSFETs, the approach offers a promising way to make repeated simulations more efficient, provided its accuracy and applicability are checked against the device conditions and training range.