A Bayesian machine learning approach for inverse prediction of high-performance concrete ingredients with targeted performance

High-performance concrete (HPC) plays an important role in improving the sustainability and reliability of buildings and infrastructures. Machine learning predictive models have been used for predicting concrete performance from ingredients, however it remains a challenge to achieve inverse prediction of ingredients from targeted performances. This study proposes an in-house coded informatics-based materials analysis framework to enable computational design of HPC with targeted strength performance. The Gaussian processes (GP) emulator is used to construct the surrogate predictive model based-on 453 experimental measurements. The validity of the GP emulator is assessed using the leave-one-out cross-validation (LOO-CV) and also a separate validation dataset. The variance-based global sensitivity analysis, Sobol indices, is applied to understand the impact of physical ingredients on the HPC performances. The results suggest that the trained GP emulator can provide sufficiently accurate and reliable predictions, as well as reflect the real-world physicochemical nature of HPC materials. The inverse material design is achieved by the Bayesian inference method with a Markov chain Monte Carlo stochastic sampling method, the Metropolis-Hastings (MH) algorithm. Combining with the Bayesian inference method, the proposed design framework can infer a list of potential HPC formulae of a targeted performance, each evaluated by the likelihood of resulting in the targeted strength. The data-driven material analysis and design framework proposed in this study provides a novel approach to achieve performance-based design of HPC, with the potential to maximise resource efficiency and reduce economical cost. The methodology presented in this study can also be extended to be applied to a wide range of construction materials, targeting difference service performances including durability.