Solid-State Detector Characterization with Physics-Based Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for characterizing solid-state detectors are limited in describing material properties on a voxel-by-voxel basis, requiring extensive resource and time, and existing machine learning models struggle with spatial limitations and training difficulties due to weak signals.
Innovation Solution
A physics-based neural network is trained using easily acquired ground truth data, such as electrode signals or free charges, with weighted loss functions to enhance the inference of weak signals, allowing for voxel-by-voxel characterization of semiconductor materials.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional experimental measurements and simulations are used to characterize solid-state detectors, then detailed material properties can be obtained, but the resource and time required are not justifiable and voxel-by-voxel characterization is challenging or impossible
Solution Approach 1:
The patent replaces conventional experimental measurement systems with a machine learning-based physical model. The neural network learns charge transport phenomena from training data and predicts material properties without requiring physical experimental setups for each measurement, dramatically reducing time and resource requirements while maintaining characterization accuracy.
Solution Approach 2:
The patent creates a virtual copy of the physical detector system through a neural network model that replicates charge transport behavior. This digital twin can be queried instantly for material property characterization without requiring actual physical measurements, enabling rapid voxel-by-voxel analysis.
2Productivity
If machine learning models are used to predict charge and voltage distribution, then characterization speed improves, but spatial limitations occur for weaker signals compared to stronger signals
Solution Approach 1:
The patent modifies the loss function parameters during training to apply differential weighting to different signal types. By adjusting the weighting factors in the loss function, the model compensates for the inherently weaker spatial information from hole signals, ensuring accurate spatial prediction across all signal types including the weaker ones.
3Measurement precision
If weighted loss functions are used to enhance weak signal inference, then accuracy of hole transport properties improves, but training complexity increases
Solution Approach 1:
The patent applies local quality by using differential weighting in the loss function where different regions of the parameter space receive different weights based on signal strength. The loss function applies higher weights to weaker hole signals locally, while maintaining appropriate weights for stronger electron signals, achieving improved accuracy without requiring completely complex training procedures.
4Loss of information
If multiple trapping centers are modeled separately, then detailed defect information is captured, but the number of experimental setups and information needed to train the models increases
Solution Approach 1:
The patent merges multiple trapping centers into a single equivalent trapping center model. Instead of requiring separate experimental setups for each trapping center, the model combines their effects into unified parameters that can be trained from a single experimental configuration, reducing the number of required experiments while preserving essential defect information.
Data Source
AI summary
A physics-based network model is trained to learn weights such as trapping, detrapping, and/or transport of holes and/or electrons, as well as voltage distribution on a voxel-by-voxel basis throughout a solid-state detector model. The physics-based network may be used to estimate material property variation throughout the voxels. To reduce the number of experimental setups and information needed to train the models, the models may be trained using more easily acquired ground truth. Just the electrode signals or just the free charge data is used to train the model to characterize the solid-state detector. With this reduced data, the detector may be characterized using equivalency, such as combining multiple trapping centers to an equivalent trapping center. Regularization may be used in the loss calculation, such as where just the electrode signals are used, to deal with the reduced data available as ground truth.