A dedicated sensitivity analysis and optimization application for industrial processes

Abstract

The paper describes the architecture and the use case of the developed Modelbox system for sensitivity analysis (SA), uncertainty analysis (UA) and the subsequent optimization of industrial processes. The proposed solution addresses the most common practical and technical problems encountered by researchers and engineers when performing sensitivity analysis. It combines the functions from the numerical toolbox with a simulation management system. Maintaining usability and a good user experience while managing complex investigations of time-consuming industrial process simulations is a very important feature of the system. Several improvements were introduced to optimize the computation time of analysis/modelling tasks, including the automatization of distributed calculations, persistent, transparent caching of simulation data and duration estimations from collected statistics. The system has the ability to perform remote, parallel, asynchronous computations of both analytic algorithms and numerical simulations. The system is dynamically scalable horizontally by using serverless computing endpoints and thus it can be easily adapted to the user's current needs in a flexible way. Modelbox provides web-based access to analysis/modelling tasks from sampling, SA/UA, optimization to metamodelling. It is extended with numerous interactive visualization components for effective results control. In addition, to access data from the completed analysis, the system supports convergence tracking for SA estimates and intermediate optimization results. The process of controlled cooling of rails was considered as a case study. The formulated optimization task was to find a combination of process parameters that ensures a minimum volume fraction of bainite along with required interlamellar spacing and optimal homogeneity of hardness. Different sensitivity analysis methods were used to evaluate the significance of all variables with respect to their influence on the model output.