In the context of the industrial shift towards carbon neutrality and electrification, heat pumps have emerged as feasible solutions for decarbonizing the heat supply previously associated only with fired or resistive heating technologies. The integration of heat pumps into industrial processes reduces the cooling and heating demand, while it capitalizes on the waste heat, which eventually enhances the overall energy efficiency. However, a heat pump device typically interacts with other competing energy systems, such as fired boilers and electric heaters. This renders the synthesis, design and optimization more complex. For this, the characterization of the grand composite curve of the industrial process, representing the amount and temperature of the available heat, is necessary to select the best levels of temperatures and refrigeration fluids that minimize the total cost of the system. Up till now, a mixed integer nonlinear programming (MINLP) approach is used to optimize the integration of a heat pump superstructure into any type of grand composite curve, bearing in mind economic and thermodynamic constraints. To reduce the problem size for large, intractable MINLP problems, and to streamline the selection of temperature levels and refrigerant fluids, supervised machine learning models are developed. By composing a database with randomly generated grand composite curves and corresponding optimal temperature levels and refrigerants, artificial neural networks, as nonlinear mapping functions for temperature prediction and refrigerant selection, are trained and are able to identify distinct patterns within the curve that influence the selection of specific heat pump structures. The models are employed in four case studies in the dairy, brewery, sugar, paper and pulp industry to evaluate the application potential and to assess the impact of the shape of the grand composite curve on the predictions. Processes with substantial waste heat availability and moderate temperatures are the most suitable for heat pump integration. In particular, plateau levels, representing streams with high heat of vaporization or thermal capacities, are identified as optimal locations for heat pumps. Self-sufficient pockets and sloping lines render less accurate model predictions. Despite the often suboptimal temperature level predictions, they are still usable for well-performing systems achieving significant energy consumption reductions of up to 50% in the case studies.