Neural Network Material Property Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Experimental datasets of material properties in materials science are limited due to high costs and limited material diversity, while simulation data, although abundant, lacks accuracy for predicting material properties.
Innovation Solution
A neural network system that learns transferable embeddings from large-scale simulation databases to predict material properties, using graph neural networks and experimental data to improve accuracy and generalize to new materials even with limited training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If simulation data is used to predict material properties, then the quantity of data is increased, but the accuracy of prediction deteriorates
Solution Approach 1:
The patent merges simulation data and experimental data into a unified training dataset. The neural network is trained on both simulation data (providing large quantity and structural information) and experimental data (providing accuracy), allowing the model to leverage the strengths of both data sources simultaneously to achieve both high quantity and high accuracy in predictions
Solution Approach 2:
The neural network acts as an intermediary that processes simulation data through learned material descriptors and embeddings, transforming the less accurate simulation data into accurate property predictions by mediating through the experimental training data and learned structural representations
2Measurement precision
If experimental data is used to predict material properties, then the accuracy of prediction is improved, but the quantity of data is reduced
Solution Approach 1:
The neural network model serves multiple functions: it learns material descriptors from simulation data, predicts properties for materials in the training distribution, and generalizes to new unseen materials. This multi-functionality allows the system to achieve high accuracy on limited experimental data while also providing broad applicability across diverse material spaces
Solution Approach 2:
The system performs preliminary learning of material descriptors and structural features from large-scale simulation data before making actual property predictions. This preliminary action of learning transferable embeddings from simulation data prepares the model to achieve high accuracy when applied to experimental data and new materials
3Productivity
If machine learning is applied to materials science, then the productivity of materials discovery is increased, but the reliability of predictions deteriorates due to limited experimental datasets
Solution Approach 1:
The patent changes the parameters of the learning process by using learning rates, batch sizes, and training epochs that are optimized for the specific combination of simulation and experimental data. The neural network parameters are tuned to achieve reliable predictions by balancing the influence of large-scale simulation data with limited experimental data through careful parameter optimization
Solution Approach 2:
The system uses feedback from experimental data to refine and validate the predictions made from simulation data. The experimental measurements provide feedback that confirms or corrects the model's predictions, increasing reliability while maintaining high productivity through the primary use of simulation data for training
Data Source
AI summary
Methods, computer systems, and apparatus, including computer programs encoded on computer storage media, for predicting one or more properties of a material. One of the methods includes maintaining data specifying a set of known materials each having a respective known physical structure; receiving data specifying a new material; identifying a plurality of known materials in the set of known materials that are similar to the new material; determining a predicted embedding of the new material from at least respective embeddings corresponding to each of the similar known materials; and processing the predicted embedding of the new material using an experimental prediction neural network to predict one or more properties of the new material.