Research On Performance Prediction And Reverse Design Of Alloy Materials Based On Deep Learning

New materials are regarded as the prerequisite for technological development and the cornerstone of production practice because of their advanced and diverse characteristics.The discovery of new materials based on traditional methods often requires a lot of time and labor costs,and the properties of the materials developed always cannot meet the requirements of scientific researchers.Driven by the Materials Genome Project,the development of algorithms and data-driven work has achieved great success,and informatics strategies have begun to be widely used in the field of materials science.The data centric informatics method is very effective for solving the problem of material properties that are difficult to measure or calculate with traditional methods due to a large amount of cost or time.Materials researchers have introduced machine learning models to replace traditional methods.The models do not require direct experiments or equation-solving methods and can quickly predict material properties based on known data.The prediction model of material performance must first identify the characteristics of the material,and then establish the mapping between the characteristics and the target attribute through the learning algorithm and can realize the reverse design of the material with specific performance based on the relationship between the material composition and performance.Alloy materials with special optical properties are widely used in various engineering fields from computer displays to solar energy utilization and have always been a research hotspot in the field of materials.In order to accelerate the discovery of new materials with target optical properties,this paper proposes a deep learning-based reverse design algorithm for optical materials,which can give the composition of materials with the required target properties.Our method is based on a combination of deep neural network models and global optimization algorithms(including genetic algorithms and Bayesian optimization).Use element attribute statistical descriptors as features to train neural network models and use transfer learning strategies to solve small data set problems.Experimental results show that the reverse design algorithm has high accuracy and generalization ability.The main research contents of this paper are as follows:A material performance prediction model based on a convolutional neural network is proposed.For the material molecular formula data obtained from the open-source database,the One-Hot coding is used for feature extraction,and the extracted features are input into the convolutional neural network for training to learn the mapping relationship between the composition of the material and its performance.Determine the initial parameters of the model based on the performance comparison of different parameters and update the weights during the training process.Given only the molecular formula of the target material,the convolutional neural network model shows good predictive performance.A method of combining transfer learning with material performance prediction models is proposed to solve the problem of small samples in the data set.First,a large number of similar samples are used for the initial training of the model to obtain the initial parameters of the model,and then small sample data is used to fine-tune the model parameters.After migration learning,the prediction accuracy of the performance prediction model on a small sample data set has been significantly improved.A material reverse design model based on deep learning is proposed to accelerate the discovery of new materials with specific properties.First,genetic algorithms and Bayesian optimization methods are used to simulate and generate material molecular formulas with different constituent elements and then based on the material performance prediction model the possible performance is calculated.A large number of experiments have proved that our reverse design model can design the molecular formula of the material that may have the performance according to the given performance.