Compressive strength prediction and composition design of structural lightweight concretes using machine learning methods

Introduction. Reducing the density, increasing the strength and other physical-technical characteristics of lightweight concretes are urgent tasks of modern building materials science. To solve them, it is necessary to consider new approaches to the development of compositions of cement systems using effective porous aggregates, binders, chemical and mineral additives, including different nanomodifiers (carbon nanotubes, fullerenes, nanoparticles of SiO2, Al2O3, Fe2O3, etc.). The complexity of design-ing modified cement concretes is largely due to their multicomponent nature and a large number of parameters affecting the key characteristics of material. The qualitative solution of such multicriteria problems is possible with the complex implementation of rational physical and computational experiments using mathematical modeling and computer technology. New opportunities for modeling of structure formation processes and predicting properties of multicomponent building materials are emerging with the development of machine learning methods. The purpose of this study is to develop machine learning algorithms that can ef-ficiently establish quantitative dependences for the compressive strength of modified lightweight concretes on their composition, as well as to identify the optimal variation ranges of prescription parameters based on the obtained multifactor models to achieve the required level of controlled mechanical characteristic. Methods and materials. The processing and analysis of experimental research results were carried out using modern methods of machine learning with a teacher used in the problems of regression recovery, knowledge extraction and forecasting. To implement the developed machine learning algorithms, libraries in the Python programming language, in particular NumPy, Pandas, Scikit-learn, Matplotlib, Seaborn, were used. Results and discussion. It is established that the gradient boosting model is the most accurate type among the obtained machine learning models. It is characterized by the following quality metrics: R2 = 0.9557; MAE = 2.4847; MSE = 12.7704; RMSE=3.5736; MAPE = 11.1813%. According to the analysis of this multifactor model, the optimal dosages of pozzolanic and expanding modifiers amounted to 4.5-6.0% and 6.0-7.5% of the binder weight (Portland cement + modifier), respectively, which ensured achievement of the required level of compressive strength (40-70 MPa) of lightweight concretes at the age of 28 days at material de