Estimating the required time for the completion of massively parallel scientific codes poses a significant challenge. This complexity arises from the presence of numerous interdependent algorithms within the code, each potentially responding differently to changes in computational power, vectorization length, memory capacity, network bandwidth, and latency, as well as I/O throughput. Accurately predicting the execution time is crucial for users engaged in large-scale simulations or high-throughput computing tasks such as virtual screening procedures. Using the first version of a pilot for prediction time-per-call, we generalize a practical approach that leverages machine learning techniques to achieve highly precise time-to-solution predictions for a material science code based on Density Functional Theory (DFT). By comparing our results with predictions generated by a parameterized analytical performance model, we demonstrate that deep learning solutions offer superior accuracy, for different machines and different codes, without the need for domain-specific knowledge or explicit algorithmic descriptions within the code implementation.