Solid-State Detector Neural Training for Weak Signal Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for characterizing solid-state detectors are limited in providing detailed material properties on a voxel-by-voxel basis, especially for weak signals, and require significant resources and time.
Innovation Solution
A physics-based neural network model is trained with a weighted loss function to enhance weak signals, using different weights for different physical phenomena such as electron and hole transport, trapping, and voltage distribution, to improve accuracy and range of material characterization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional experimental measurements and simulations are used to characterize solid-state detector materials, then material properties can be measured, but the characterization is limited to bulk properties and cannot provide detailed voxel-by-voxel description at micrometer scale
Solution Approach 1:
The patent creates a virtual copy of the solid-state detector material through a machine learning model that replicates material properties at the voxel level. Instead of physically measuring each voxel, the model learns from training data and generates detailed spatial maps of material properties, effectively copying the complex physical measurement process into a computational domain.
Solution Approach 2:
The patent replaces the mechanical/physical measurement system with a computational machine learning system. The physical measurements and simulations are substituted by a neural network that processes input data and outputs detailed material property distributions, eliminating the need for complex dedicated measurement equipment.
2Measurement precision
If conventional measurement methods are used to achieve detailed material characterization, then voxel-by-voxel description can be obtained, but the resource and time required become not justifiable
Solution Approach 1:
The patent performs preliminary training of the machine learning model using training data that captures material properties. Once trained, the model can rapidly predict material properties for new detector materials without requiring time-consuming physical measurements, effectively preparing the knowledge base in advance to enable fast characterization.
Solution Approach 2:
The trained model creates a computational copy that replicates material characterization results instantly. Instead of physically measuring each property, the model retrieves or generates the information from its learned representations, dramatically reducing the time required while maintaining detailed voxel-by-voxel description capability.
3Productivity
If machine learning model is used to predict charge and voltage distribution, then rapid characterization can be achieved, but the model has spatial limitations in prediction for weaker signals
Solution Approach 1:
The patent applies different weighting factors to different signal strengths in the loss function during training. Strong signals receive one weighting while weak signals receive a higher weighting factor, ensuring that the model pays special attention to accurately predicting weak signals in specific spatial regions, thereby removing spatial limitations in prediction accuracy.
Solution Approach 2:
The patent changes the training parameters by introducing weighted loss function with adjustable weighting factors. By modifying these parameters, the model learns to prioritize accurate prediction of weak signals, expanding its reliable prediction range across different signal strengths while maintaining rapid characterization capability.
4Measurement precision
If stronger signals are prioritized in model training, then model accuracy for strong signals improves, but weaker signals are neglected and their spatial range is limited
Solution Approach 1:
The patent implements local quality differentiation in the loss function by applying higher weighting factors to weak signals. This ensures that the model maintains high accuracy for strong signals while simultaneously improving its ability to predict weak signals across extended spatial ranges, making the model adaptable to various signal strengths.
Solution Approach 2:
The patent modifies training parameters by introducing variable weighting factors in the loss function. By adjusting these parameters, the model achieves a balance between accurately predicting strong signals and extending its prediction capability to weak signals, thereby enhancing overall adaptability across different signal conditions.
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. Anode and cathode signals as well as the voltage distribution are relatively strong signals compared to the weaker electron and hole signals. The relatively weaker signals may be limited in range across voxels. In order to expand the range or magnify the effect, the loss function used in training the physics-based neural network may use a weighted combination where the weaker signals are weighted more heavily than stronger signals without substantially reducing the influence of the stronger signals. This improves the inference, resulting in improvement of the accuracy and range of the trained physics-based model.