Performance Optimization of Strength Prediction Models Based on Cementitious Materials Literature Data

In light of the growing significance of high-quality datasets for the advancement of artificial intelligence, there has been a notable shift in focus towards utilizing literature data for the construction of high-quality R&D datasets for performance prediction models and composition design. This has emerged as a pivotal objective within the domain of materials research and development. This article employs the SupperalloyDigger algorithm in conjunction with the cement property dictionary to mine the literature data set comprising 57,037 related documents on cementitious materials from Elsevier Press. The cosine similarity algorithm is employed for the effective classification of the data items within the dataset. The classified data is then subjected to a process of deep cleaning in conjunction with expert knowledge of cementitious materials, resulting in the creation of a literature dataset with a data capacity of 22,810 and 53 data dimensions. A variety of algorithms, including linear regression, support vector machine, multilayer perceptron, random forest, artificial neural network, and recurrent neural network, were employed to construct prediction models for 28-day compressive strength in cementitious materials. Additionally, weight optimization algorithms were utilized to enhance the model performance. The weight optimization method (EK-RFWO), which has been optimized using expert knowledge, has been demonstrated to enhance the performance of six models in both the training and testing sets. The mean square error (MSE) of the training set decreased by 0.18–0.37, and the MSE of the testing set decreased by over 0.40 in comparison to the initial model. These results demonstrate that the EK-RFWO method can markedly enhance the prediction accuracy of data models constructed based on literature data, and improve the performance of the model.