One-Sided Material Property Prediction With Virtual Imputation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models face increased difficulty in predicting one-sided material properties, such as negative bandgap, negative thermal expansion, and stepped electrical resistivity, due to their deviation from the norm in the same material class, complicating training and prediction processes.
Innovation Solution
A machine learning system and method that imputes a fixed value for one-sided material properties, trains a model using this fixed value, and predicts properties for both trained and untrained materials, employing techniques like Gaussian Process regression and supervised models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are trained to predict one-sided material properties, then the ability to predict unusual material properties is improved, but the training difficulty and model complexity increase significantly
Solution Approach 1:
The patent segments the training data into two distinct subsets: one for materials with properties within the normal range and another for materials with one-sided properties. This segmentation allows the model to learn different patterns separately, reducing training complexity while improving prediction accuracy for unusual properties.
Solution Approach 2:
The patent applies parameter changes by transforming the target property values through a mapping function that converts one-sided property values into a standardized range. This transformation simplifies the learning task for the model while preserving the ability to predict unusual material properties accurately.
2Adaptability or versatility
If machine learning models are trained with diverse material properties including one-sided properties, then the versatility of the model is improved, but the training data requirements and computational resources increase
Solution Approach 1:
The patent performs preliminary action by pre-processing the training data to identify and separate one-sided properties before model training. This preliminary segmentation reduces the computational burden during training and allows the model to be trained on smaller, more focused datasets while maintaining versatility.
Solution Approach 2:
The patent introduces an intermediary mapping function that transforms one-sided property values into a standardized format. This intermediary transformation allows the model to handle diverse material properties uniformly without requiring separate processing pathways, reducing overall computational requirements.
Data Source
AI summary
A system for predicting a one-sided property value for one or more material candidates includes a processor and a memory communicably coupled to the processor. The memory includes a stored acquisition module and a machine learning (ML) module. The acquisition module is configured to select a training data set with a given material property. The training data set includes a first subset of materials having the material property within a predefined range and a second subset of materials having the material property outside the predefined range. Also, the ML module is configured to impute a fixed value for the material property outside the predefined range and train a ML model to predict the material property using imputed fixed value.