The increasing demand for sustainable packaging solutions drives the adoption of paperboard, a renewable and recyclable material. This study presents a novel approach to optimize the complex, non-linear process of paperboard press forming using Long Short-Term Memory (LSTM) networks. Instead of relying on traditional trial-and-error methods or computationally intensive FEM-methods, our method trains LSTM models tailored to a specific press-forming machine (LUT Packer 2020) and its operational characteristics using experimental data derived from 181 parameter combinations out of 1584 possibilities across two material types, focusing on key process variables like blank holding force (BHF), relative humidity, forming speed, and temperature. We leverage the inherent ability of LSTMs to model sequential data, treating the blank holding force (BHF) as a sequence to capture the causal dynamics of the forming process. A key contribution is the examination of the LSTM’s learned latent space representations using UMAP and distance correlation. This reveals hidden relationships and dependencies between key process parameters are not apparent from the raw experimental data. While the study utilizes a relatively sparse dataset, the results demonstrate the potential of LSTMs to capture underlying process trends and provide a foundation for more robust optimization with larger, more comprehensive datasets.