Study of nonlinear relationships between dosage mixture design and the compressive strength of UHPC

With advances in building materials research, new materials more suitable for specific applications due to their superior durability and mechanical properties are emerging to meet the requirements of infrastructure construction and maintenance. In this sense, ultra-high-performance concrete (UHPC) is considered one of the most promising materials for concrete construction. However, a traditional UHPC mixture contains large amounts of cement, silica fume, superplasticizer, and other expensive and high carbon footprint components. Hence, several researchers have focused on developing alternative UHPC dosages using locally available materials, including several mineral admixtures, to obtain a more affordable and sustainable UHPC. Therefore, deep knowledge about the relationships between the UHPC dosage and its resulting properties could help to improve these developments. However, these relationships are nonlinear and complex. This paper aims to address this gap by using a random forest model trained with a 931 UHPC mixture collection with 17 input variables and a unique response, namely the UHPC compressive strength. After adjusting a regression model, different tools, such as input variable importance analysis and partial dependence plots, were used to assess the nonlinear relationships between the dosage of UHPC and its resulting compressive strength. In most cases, a proper alignment was observed when the findings provided by these analyses were contrasted with the results provided by other researchers. The results indicated that the inputs with the highest significance on UHPC compressive strength are those related to water content, the packing density, and the quartz powder and superplasticizer dosages, followed by the cement content. Furthermore, mineral admixtures with high amorphous SiO2 content, such as RHA and SF, presented a relevant contribution to the compressive strength, while the other admixtures would seem to have a more modest contribution to it. Finally, it is also worth noting that the machine learning model created might help to create novel UHPC dosages by decreasing the duration and expenses of the experimental campaign by permitting the pre-selection of those components with a superior model response.