Research On Material Data Prediction Algorithm Based On Model-Agnostic Meta-Learning

The original intention of the meta-learning method is to solve the problem of insufficient label sample size,and to accelerate the convergence speed of the model by training the initialization parameters of the model in advance.The emergence of subsequent Model-Agnostic Meta-Learning(MAML)has greatly increased the scope of application of meta-learning ideas.Generally,the model solved by the gradient descent algorithm can solve the problem of few-shot learning through it.At present,the application scenarios of meta-learning methods are mostly image and natural language processing.Chemical material data analysis has multi-task training characteristics,which is a typical few-shot learning problem.However,there are still many problems in applying meta-learning methods to material data analysis..In this paper,the structure of MAML algorithm and model is improved to adapt to the analysis of material data.This article takes electrolytic water catalytic material data as the research object,and applies the meta-learning method to material data analysis and prediction.In this application scenario,the composition of the compound divides the data into a multi-task training mode that combines non-metallic elements and transition metals.The same compound can be divided into different tasks by replacing non-metallic elements.Due to the influence of the attributes of the element itself,the correlation between tasks is not uniform and there are redundant and invalid information in the feature letter in the data of a single task.In order to solve such problems,this paper uses the characteristics of Model-Agnostic Meta-Learning multi-task training to deal with the limited amount of data in the material data and the information gap between tasks.On this basis,a special data enhancement method is combined,and the MAML algorithm is improved and optimized.1.First of all,use the information volume and self-encoding of the data to realize the needs of data feature selection and dimensionality reduction,and make full use of the existing information of the data.Due to the fact that there are different types of tasks between tasks in the data,the algorithm of MAML is optimized for this specific data characteristic.When the algorithm updates the parameters,increase the weight of adapting to the relevance of the task,and amplify the influencing factors of the target task in the solution process,so that the model algorithm is more suitable for the data characteristics of the uneven relationship between the training tasks,and the convergence speed of the model and the accuracy of the prediction are improved.rate.2.Secondly,in order to ensure that the new tasks still maintain a strong correlation,a data enhancement method is proposed,which uses the low-information features and invalid features that are filtered out in the data processing process as the subsequent noise factors.Controlling the added type and number of noise factors to generate new tasks not only ensures that there are differences between tasks,but also makes the difference between the two tasks under the same label not too big.At the same time,the first update of the model algorithm is optimized through the idea of LSTM to solve the problem of gradient disappearance when the task sequence increases.In the model angle and data angle,the ability to analyze small samples has been greatly improved.3.Finally,experiments have proved that after the model is optimized by the algorithm,the accuracy of the prediction has been significantly improved when the changes of the parameters are in response to the experimental data that has been determined for a specific category.When the task sequence is too long and the multi-task training,the LSTM optimized MAML algorithm model has a greater improvement effect than the normal training model on the basis of data enhancement.And this proves that the model has strong feasibility in specific applications.Since the data analysis of electrolytic water catalytic materials is representative in the material field,the method implemented in this paper has a certain general applicability.