Model based learning for efficient modelling of heat transfer dynamics

Predictive models are crucially relevant for the design and control of manufacturing processes, particularly when process quality is highly sensitive to unknown physical parameters. This paper describes a method to model industrial processes using neural networks and physical prior knowledge. The approach is applied to heat transfer problems, particularly relevant for additive manufacturing processes, and the results are compared to alternative methodologies. Obtained results show that the integration of partial differential equations in the neural network model leads to reduced amounts of required training data and increases the model stability. Such outcomes represent promising characteristics for novel model predictive control strategies.
Procedia CIRP, 102
ISSN:2212-8271
18th CIRP Conference on Modeling of Machining Operations (CMMO)