摘要
The growing demand for monoclonal antibodies (mAbs) has driven significant advances in upstream bioprocess development, particularly in fed-batch and perfusion bioreactor systems. This work compilation synthesises a broad spectrum of recent studies focusing on the modelling, simulation, optimisation, and economic evaluation of mAb production processes. The body of work reviewed encompasses both mechanistic and data-driven approaches, offering a comparative lens on the evolving landscape of biopharmaceutical manufacturing. Central to this review is the application of mathematical models that describe the dynamic behaviour of mammalian cell cultures in both fed-batch and perfusion modes. These models are critical for tasks such as parameter estimation, state prediction, and process control, and have been employed extensively for dynamic simulation and model-based optimisation. In particular, dynamic optimisation frameworks have been used to derive time-dependent feeding strategies aimed at maximising mAb yield or minimising resource consumption under process constraints. Parallel to this, the review examines the increasing integration of machine learning (ML) techniques, including supervised learning models and neural networks (NNs), for prediction and optimisation in bioprocesses. These methods have demonstrated utility in forecasting mAb production, especially under variable culture conditions where traditional modelling approaches may be limited by structural uncertainty or parameter identifiability. Specific attention is given to studies employing physics-informed neural networks (PINNs), which blend empirical data with first-principles knowledge to improve prediction accuracy and model interpretability. Furthermore, it compares model-based optimisation strategies with data-driven alternatives, evaluating their respective strengths, limitations, and areas of overlap. While mechanistic models offer transparency and extrapolation capabilities, data-driven methods are increasingly favoured for their adaptability and ability to uncover nonlinear relationships within high-dimensional datasets. To assess the practical implications of these approaches, the review includes techno-economic analyses (TEAs) that quantify process performance in terms of cost of goods (CoG), media consumption, labour, and yield-based profitability. Comparative TEA results indicate trade-offs between operational cost savings and product output, highlighting the context-dependent nature of choosing between model-based and ML-based optimisation.