Machine Learning for Simulation

Machine learning and classical simulation traditionally follow different modelling paradigms. PN 6 investigated how both could be combined to develop data-integrated methods capable of addressing simulation problems that neither approach could solve as effectively on its own. A particular focus was on adapting machine-learning methods to the specific characteristics of simulation data and incorporating physical constraints, symmetries and prior knowledge.

The research also explored the reciprocal relationship between simulation and machine learning: simulations could provide additional training data for learning methods, while machine learning could support modelling, prediction and the analysis of complex simulation results. Visualization and visual analytics played an important role in making machine-learning methods more understandable and enabling users to explore and interact with them.

The research addressed three closely connected questions:

• Machine Learning for Simulation: How could machine learning be adapted to simulation problems while incorporating physical constraints and prior knowledge?
• Simulation for Machine Learning: How could simulation data complement experimental and real-world data to improve machine-learning methods?
• Visualization and Machine Learning: How could visualization help users understand and control machine-learning methods – and how could machine learning, in turn, support data-integrated visualization?