In semiconductor device simulation, a deep learning Poisson solver can reduce the time spent updating electric potential during particle-based analysis. That matters in transient studies, where the solver must calculate the field repeatedly as the distribution of electrons and holes changes. A demonstrated convolutional neural network (CNN) for a PN junction reproduced the finite-difference method’s I-V curve while solving each time step more than 10 times faster in the reported setup.
The result is a focused example, not evidence that a trained network can replace conventional numerical solvers in every device or operating condition. Its performance depends on the simulated problem, the training data, and how boundary conditions are represented.
Why Poisson’s equation can limit particle simulation
Particle-based semiconductor simulation tracks charge carriers as they drift, diffuse, and scatter under electric fields. In a Monte Carlo simulation, these processes update the positions and distribution of particles over time. The resulting charge distribution is then used to solve Poisson’s equation and calculate an updated potential and electric field for the next step.
Poisson’s equation relates the spatial distribution of electric potential to charge density. For a complex device geometry, a closed-form solution may not be available, so numerical methods are used. The finite difference method (FDM) approximates spatial derivatives on a mesh and iteratively updates the potential at grid points until the solution converges.
FDM is established and straightforward to implement, but its iterative work is repeated throughout a transient simulation. When a study includes thousands of time steps, the Poisson solve can therefore account for substantial computational effort. Any faster alternative must still produce sufficiently accurate potentials, because errors in the field can affect subsequent particle motion and accumulate across steps.
Designing a deep learning Poisson solver for a PN junction
The reported study applied a CNN to a two-dimensional PN junction. The simulated p- and n-regions each measured 200 nm in length and 80 nm in height, with a stated doping density of 1 × 1023 cm-3. With a 10 nm spatial mesh, potential and charge distributions were represented as 9 × 41 matrices. The Monte Carlo simulation used a 10 fs time step.
The network was trained to map charge distribution to potential distribution. Crucially, charge alone was not the only input: the applied voltage bias determines the potential at the device boundaries. The model therefore also received a boundary-condition matrix, with the specified boundary potentials represented at the two ends of the junction. The network output was a potential matrix with the same 9 × 41 dimensions.
This treatment addresses a practical issue in applying machine learning to numerical field solvers. A model that predicts values only in an interior region, or that does not account for boundary values, may not satisfy the conditions that define the physical problem. Here, boundary information was supplied as an input and explicitly emphasized during training.
Training data and boundary-aware loss
FDM solutions provided the examples used to train and validate the network. The dataset contained 580,000 saved simulation steps generated across voltage biases sampled from 0 to 2.8 V in 0.1 V increments. Seventy percent of the data was used for training and 30 percent for validation.
The loss function included one term for the difference between predicted and reference potential distributions and a second term for the boundary-condition matrix. The boundary term received twice the weight of the potential-error term. This weighting was selected empirically in the study. It reflects the design priority: the network must not only approximate the interior potential but also respect the specified boundary values.
The CNN used convolutional layers with multiple kernel sizes to capture patterns at different spatial scales, followed by additional convolutional and fully connected layers. Batch normalization and Leaky ReLU activations were used within the network. Max pooling and dropout, along with the volume of training data, were used to help limit overfitting. Training used the Adam optimizer and stopped after the loss converged over 16 epochs.
Checking accuracy across dynamic simulation
Agreement on individual potential maps is not enough to establish that a faster solver is suitable for a multi-step particle simulation. Small prediction errors could influence later charge distributions, so the study also assessed the CNN in the simulation loop. The trained model replaced the FDM solver, and the PN junction was simulated for 20,000 steps at each tested bias, sampled every 0.05 V from 0 to 2.8 V.
The resulting I-V curve agreed very well with the curve obtained using the finite difference solver. This provides an end-to-end check on the simulated electrical response, in addition to comparisons of individual potential distributions. The authors reported that the CNN’s predicted potentials tracked the FDM results and matched the boundary values.
For the reported hardware and implementation, a single CNN solve took 3 milliseconds on an NVIDIA GeForce GTX 1080 Ti, compared with up to 40 milliseconds for the FDM solve. The study described this as an improvement of more than 10 times per step. These timings apply to the tested setup; they should not be treated as a general performance guarantee for other meshes, devices, hardware, or neural-network designs.
Practical implications for simulation teams
For engineers evaluating machine learning in numerical simulation, the example highlights several considerations:
- Represent the physics that shapes the solution. Boundary conditions were included in both the network input and the loss function rather than treated as incidental details.
- Validate the full workflow. Comparing potential maps is useful, but checking the I-V curve after many simulation steps helps assess whether errors materially affect the simulated device response.
- Define the model’s scope. The CNN learned from FDM-generated data for a specific PN junction setup and range of applied biases. Applying it to different geometries or conditions would require appropriate evaluation and potentially new training data.
- Measure speed in context. The reported per-step timing is meaningful because Poisson’s equation is solved repeatedly, but actual benefits depend on the complete simulation workflow and computing environment.
Conclusion
A CNN can serve as a fast surrogate for the Poisson solve in a particle-based PN junction simulation when its inputs and training objective account for boundary conditions. In the demonstrated case, the model preserved close agreement with the FDM-derived I-V curve over a long simulation and reduced per-step solving time by more than 10 times on the tested hardware. The practical lesson is to assess such methods not only by prediction accuracy on isolated outputs, but also by boundary-condition fidelity, cumulative behavior, and performance within the complete simulation loop.