Development of machine learning models for predicting properties of carbon materials and design of process conditions for production of materials with desired multiple properties

Cokes are an essential material in the iron and steel industry, and widely used as a fuel and for reducing iron ore. The product properties of cokes are influenced by raw materials and process conditions, which in turn affect its performance. Because raw material components vary depending on the origin and production period, product properties can fluctuate even under the same process conditions, and thus, it is necessary to design appropriate process conditions for each raw material. The objective in this study is to construct machine learning models that predict product properties from raw materials and process conditions using data sets measured in needle and pitch coke production processes, and to design appropriate process conditions that achieve desired values of properties by inverse analysis of the constructed models. To select appropriate process conditions and their time delays for a data set, we propose a method called NSGA-II-VDS, combining elitist non-dominated sorting genetic algorithm (NSGA-II) and genetic algorithm-based process variable and dynamics selection (GAVDS), for multiple product properties, and confirmed that the proposed NSGA-II-VDS could construct models with higher prediction accuracy than the conventional GAVDS. Furthermore, we constructed models using a data set where time-series data was interpolated based on the maximum batch time for each process condition for a batch process data set. By inputting time-series data generated virtually into the constructed models for raw material information for a target product, we successfully designed process conditions predicted to achieve the desired product property values by selecting time-series data with promising predicted product quality values.