Neural Network Modeling Semiconductor Charge Transport
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for characterizing semiconductor materials are limited in detailing material properties on a voxel-by-voxel basis, especially at the micrometer scale, due to resource and time constraints, and are unable to accurately model charge transport properties across the bulk of the material.
Innovation Solution
A physics-based neural network is trained using input charges and output signals to model semiconductor materials, with weights optimized through a gradient descent algorithm, representing transport, trapping, de-trapping, recombination, and diffusion of electrons and holes as tensor fields, allowing for voxel-by-voxel property description.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional experimental measurements and simulations are used to characterize semiconductor materials, then material properties can be measured with specialized equipment, but detailed voxel-by-voxel characterization at micrometer scale is impossible due to resource and time constraints
Solution Approach 1:
The patent replaces conventional mechanical/experimental measurement systems with a neural network-based computational system. The neural network is trained on experimental data and then used to predict material properties at voxel-by-voxel resolution, substituting physical measurements with computational predictions that are both detailed and efficient.
Solution Approach 2:
The patent creates a computational copy of the semiconductor material's physical properties through the neural network model. Instead of physically measuring each voxel, the network learns the underlying patterns from limited measurements and generates virtual measurements for all voxels, effectively copying the material's behavior without physical intervention.
2Measurement precision
If conventional neural network methods are used for charge transport prediction, then electronic coupling elements can be predicted, but the methods are limited in finding material properties along the bulk of the material and require optimizing many weights with no direct relation to physical properties
Solution Approach 1:
The patent changes the parameters being optimized in the neural network from abstract weight values to physical material properties such as mobility, trapping lifetime, and de-trapping lifetime. This transformation makes the optimization process more efficient and the results directly interpretable in physical terms.
Solution Approach 2:
The patent extracts the essential physical properties from the complex neural network training process. Instead of optimizing all possible weights, the method identifies and optimizes only the key physical parameters that govern charge transport, removing unnecessary complexity while maintaining predictive accuracy.
3Measurement precision
If analytical solutions for nonlinear transport equations are solved using conventional numerical methods, then charge transport can be modeled, but material properties are assumed to be constant across the bulk and cannot be found along the bulk of the material
Solution Approach 1:
The patent segments the semiconductor material into discrete voxels, allowing material properties to vary spatially across the bulk. The neural network processes each voxel independently, enabling detailed spatial characterization without requiring complex bulk-level numerical solutions to nonlinear transport equations.
Data Source
AI summary
A method of training a neural network modeling physical phenomena of semiconductor material includes receiving plurality of training pairs corresponding to a semiconductor material. Each training pair comprises an input charge to a distinct voxel of the semiconductor material and one or more output signals generated by the distinct voxel in response to the input charge. A neural network is trained using the training pairs. The neural network models the semiconductor material and each voxel is represented in the neural network by a tensor field defined by (i) a location of the voxel within the semiconductor material and (ii) one or more physics-based phenomena within the voxel at the location.