Manufacturing processes are characterised by parameter fluctuations and noise inherent to the industrial setting. These uncertainties are convoluted with variations like changes of material parameters or varying geometric properties of the manufactured products. Consequently, these interactions must be considered during the design of the forming tool’s active surface to ensure the quality of the manufactured products. Due to the iterative nature of the tool design process and the stochastic character of the parameter variations, i.e. process noise, the finite element method is not a suitable approach to efficiently design an active surface which is robust against process noise. Yet, data-driven models like artificial neural networks grant a holistic modelling method. However, these models demand a substantial amount of training data to ensure accurate predictions across diverse geometric specifications and material types. By using transfer learning, the scalability of these models can be ensured by pre-training the main effects and interactions on a baseline domain dataset obtained from finite element simulations and subsequently fine-tuning on a specialised dataset representing a target domain which is distorted from the baseline domain by a parametric shift. Active learning helps to iteratively find the subset of additional data points that need to be selected to learn the algorithm efficiently. In the presented approach, active transfer learning is used to minimise the amount of data to adapt between parameter domains representing variations within the deep drawing process, thereby improving the reusability of already existing datasets and optimising the design of specialised datasets.