Kromatografian datalähtöinen sijaismallinnus syväoppimisen avulla

Separation of biocompounds by batch chromatography has gained special interest in the biopharmaceutical industry due to its high selectivity and resolution. The design and optimization of this process requires the solution of coupled Partial Differential Equations (PDEs) that often require high computational resources. This project explores the integration of data-driven Machine Learning (ML) aiming to reduce the computational effort and simulation runtimes. A Long Short-Term Memory (LSTM) neural network was trained in PyTorch to predict the elution profiles of binary mixtures under batch chromatography. Training data was generated by simulating the Lumped Kinetic Model (LKM) using the Finite Volumes Method (FVM) to reduce the PDE system to a set of Ordinary Differential Ecuations (ODEs). The JAX library was used to create high fidelity simulations with Just-In-Time (JIT) compilation. The trained LSTM was capable of predicting elution profiles at a significantly lower computational cost and presented a decrease of 99.79% ± 0.050% in the execution time compared to the execution in GPU hardware of the mechanistic models. The inference was analyzed by the Roofline and the GPU Speed-Of-Light (SOL) analyses to assess hardware utilization using NVIDIA Nsight Compute~(2025.3.1), indicating that the LSTM low arithmetic intensity limits the performance during inference. The model developed constitutes in the creation of digital twins for batch chromatography that can respond in real time potentially helping in process optimization and control.