The invention relates to a computer-implemented method for predicting or simulating a value distribution of a physical
state variable in a component using a graph-based
machine learning model, wherein the graph-based
machine learning model processes as input data a
data graph in the form of a network of nodes and edges forming a computational grid, wherein the value distribution of the values assigned to the nodes represents the distribution of the physical
state variable in an area or volume region of the component, comprising the steps: - Determining (S2, S12, S22, S32, S42) a computational grid, in particular by transformation, in the form of the
data graph as input data; - Use (S3, S13, S23, S35, S43) of a suitably trained graph-based
machine learning model depending on the input data in order to output as a result of the evaluation a dataed computational grid as the resulting value distribution of the physical
state variable or an aggregated quantity that characterizes the value distribution as a key parameter (S4, S15, S25, S37).