The introduction of intelligent models, in particular using machine learning methods, opens up prospects for the development of the construction industry. The construction of regression models for predicting the physical and mechanical properties of various types of building materials is a promising and relevant area. The use of such models makes it possible to take into account complex and multifactorial dependencies, while minimizing the influence of the human factor. In the present study, variatropic concrete B30, obtained by centrifugation, acts as the test material. The dataset (351 objects) was assembled during laboratory studies to study the effect of freeze-thaw cycles on the strength characteristics of the material. Using the computer vision method based on the convolutional neural network U-Net, the damage on each of the concrete layers was assessed on different cycles. 4 machine learning models for predicting compressive strength were trained and tested on the collected dataset: Ridge Regression (RR), Random Forest (RF), CatBoost (CB) and Multi-layer Perceptron (MLP). The hyperparameters of the models were optimized using Grid Search + 3-fold cross-validation. As a result of testing the algorithms on a test sample, the best quality metrics were demonstrated by tree architectures: MAE for RF and CB 0.09 and 0.17 MPa, respectively, R2 = 0.99. The results are supplemented by SHAP analysis. The results obtained are a useful tool for optimizing the composition of variatropic concretes used under aggressive conditions.