Fast Simulation Response by a Simulation Based Machine Learning (SMiLe) Approach

Abstract

AbstractProcess simulation is an important tool for designing manufacturing processes before manufacturing a part, optimizing the process, and later identifying the causes of problems and failures during manufacture. Calculation times for complex manufacturing processes can be several hours to days. This prevents rapid response through simulation in the event of a failure during production and integration of process simulation into design optimization loops.Simulation-based Machine Learning (SMiLe) can be an approach to achieve fast response times using simulations: Many simulations of process and material parameter variations and different designs are used to create a database for training machine learning algorithms that predict the key results of the process simulation. The machine learning algorithms can then be integrated into a design optimization loop or provide quick hints for avoiding failures in a production process. The basis for this approach is a simulation that is well calibrated by experimental data for the manufacturing process and the materials involved. If computing resources and software licenses are available, the simulation can be used continuously to improve the simulation database by adding simulation results, thereby increasing the quality of the machine learning algorithm’s prediction. This is a paradigm shift from running simulations when results are needed, to running simulations before results are needed, to getting quick solution hints through simulation when needed.The SMiLe method and workflow are presented and machine learning algorithms are discussed. The method is demonstrated on an industrial manufacturing process, the prediction of cast iron microstructural properties for wind turbine applications.KeywordsProcess simulationMachine learningProcess optimization