Deep learning from physicochemical information of concrete with an artificial language for property prediction and reaction discovery

•An artificial language is created to represent various variables of complex concrete mixtures.•The physicochemical information of concrete ingredients is considered using this language.•An automatic approach is proposed to configure deep learning models for concrete mixtures.•The approach can assess the importance of variables and discover chemical reactions.•The approach shows high accuracy and high generalizability for predicting complex mixtures.