Spatial decorrelation in two IFU cubes of the structured mock dataset for directly imaged companions. Left panels: brown dwarf signals inserted in a structured manner, as varying Gaussian decaying ellipses within a delimited aperture, showing variations in position, size, and shape. The white areas indicate removal of real companions. Right panels: plots illustrating datasets after flattening and transformation, with the horizontal axis representing radial velocity and the vertical axis representing stacked spatial dimensions. Visible lines denote spax-els with inserted planetary signals, showing variations based on the inserted planets’ properties. This method prevents the ML algorithm from learning redundant spatial artefacts; it emphasises learning from cross-correlation patterns in the spaxel dimension.