Multitask Neural Network for Material Property Prediction with Scarce Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning approaches require substantial data and domain knowledge to build accurate models, leading to high costs and delays in predicting material properties, especially when data is scarce, limiting their transferability and prediction power.
Innovation Solution
The use of multitask learning and transfer learning methods, such as compositionally restricted attention-based networks and residual neural networks, allows for the generation of accurate models with significantly less data, enabling simultaneous training of multiple properties and leveraging commonalities between related data points, thereby reducing the need for extensive data and domain expertise.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning approaches are used with scarce data, then model accuracy is limited, but requiring more data increases cost and time
Solution Approach 1:
The patent merges multiple related tasks into a single unified model that predicts multiple material properties simultaneously. By combining property prediction tasks (e.g., mechanical, thermal, electrical properties) into one model framework, the system leverages shared patterns across properties to improve accuracy while reducing the total data required compared to training separate models for each property.
Solution Approach 2:
The patent creates a universal machine learning model that can predict multiple different material properties using the same underlying architecture and training approach. This multi-functional model generalizes patterns from available data across different property types, enabling accurate predictions for scarce data scenarios without requiring property-specific models for each material characteristic.
2Measurement precision
If physical modeling is used to build models with scarce data, then domain knowledge is required, but this increases complexity and time consumption
Solution Approach 1:
The patent replaces physical modeling approaches with data-driven machine learning models. Instead of using physics-based equations and domain knowledge to model material properties, the system uses neural networks and other ML algorithms that automatically learn patterns from experimental data. This substitution eliminates the need for complex physical modeling while achieving comparable or superior accuracy, especially when data is available.
Solution Approach 2:
The patent changes the fundamental parameters of model development by moving from physics-based parameterization to data-based parameter learning. Rather than requiring explicit physical equations and domain expertise to formulate models, the system uses statistical patterns and numerical relationships extracted from data, fundamentally changing how models are constructed and reducing development complexity.
3Productivity
If traditional machine learning models are trained with limited data, then processing power requirements increase, but reducing data decreases model accuracy
Solution Approach 1:
The patent combines multiple property prediction tasks into a single unified model that processes data once and produces multiple predictions simultaneously. This merging approach improves training efficiency by avoiding redundant processing for each property separately, while the shared learning mechanisms maintain accuracy through cross-property pattern recognition that compensates for limited data.
Data Source
AI summary
A method of forming a model for predicting one or more properties is provided. The method includes determining a plurality of datasets of properties. The method also includes training a common encoder and one or more individual decoders utilizing the plurality of datasets of properties. Each individual decoder of the individual decoder(s) is distinct from each other and is used to model different properties. The method also includes determining a transfer learning dataset for one or more properties, training a new decoder using the transfer learning dataset and the common encoder, generating a predicted property using the common encoder and the new decoder, and preparing an item using the predicted property.