Understanding and predicting micro-characteristics of ultra-high performance concrete (UHPC) with green porous lightweight aggregates: Insights from machine learning techniques

Ultra-high performance concrete (UHPC) is an advanced material in construction. Porous lightweight aggregates (PLWA) could reduce the self-shrinkage risk of UHPC by maintaining internal relative humidity. However, understanding and predicting the water migration processes influenced by PLWA are still challenging. Given that machine learning (ML) has shown promise in modeling complex relationships, this study aims to create a reliable ML model to predict and analyze the micro-properties of UHPC with different green PLWAs (pumice and phosphogypsum aggregates). Furthermore, it aims to enhance our understanding of how PLWA influences the microstructure development within the UHPC. The research results demonstrated that convolutional neural network (CNN) algorithm with an R2 2 value exceeding 0.95 in both the test and training data. Meanwhile, the CNN was employed to predict time-dependent water content and hydration degree of UHPC containing various types of PLWA and multiple-aggregates, aiding in the exploration of how different PLWA impact the water migration of UHPC. Based on the ML analysis results, pumice aggregates and multiple-aggregates both contributed to reducing the rate of water migration, subsequently reducing UHPC shrinkage. Finally, some insights on the use of ML techniques for predicting and understanding the micro properties of UHPC were discussed. ML was employed for micro-performance prediction and in-depth analysis, thereby advancing the intelligent evolution of green UHPC products.