Abstract
Self-driving laboratories (SDLs) increasingly use Bayesian optimisation (BO) to navigate expensive design spaces, yet high-fidelity simulations and experiments remain too costly to query at scale. Multi-fidelity Bayesian optimisation (MFBO) alleviates this by combining abundant low-fidelity evaluations with scarce high-fidelity observations. However, Gaussian process (GP) surrogates can become computational bottlenecks as data volume and dimensionality increase, motivating scalable alternatives. Here, we assess transfer learning based deep neural network (DNN) surrogates that pretrain on low-fidelity data and fine-tune on high-fidelity observations. We construct a chemical process benchmark for glacial acetic acid separation and purification with paired low- and high-fidelity flowsheets. The optimisation considers eight decision variables and minimises the minimum selling price (MSP), while enforcing a product purity threshold via a quadratic penalty. To reflect realistic resource constraints, we define seven training scenarios with approximately constant total simulation time, given the large cost gap between fidelities. Across all scenarios, the transfer learning DNN outperforms a GP regression baseline trained only on high-fidelity data. Test-set R² increases from 0.41 to 0.71 for the DNN, compared with 0.32 to 0.51 for GP, with gains already visible in the most data-scarce setting. These results suggest that transfer learning-enhanced DNN surrogates can offer accurate, scalable multi-fidelity surrogate models for process optimisation, and could be a practical component in BO-based automated discovery settings where low- and high-fidelity data coexist.