Data-efficient design of experiments for lipid nanoparticle development through prior-informed machine learning

Experimental design for nanomedicine product development is often limited by high experimental costs and small datasets. Traditional design of experiments (DOE) methods are typically applied to a single development programme with a fixed set of parameters, leading to redundant testing and inefficient exploration of the design space. This work presents a machine learning (ML) enabled DOE framework that leverages prior experimental knowledge to guide efficient sampling across conditions. The framework integrates ML models with adaptive point selection to identify informative regions of the design space while minimizing experimental runs. We propose two strategies for transferring knowledge between conditions: prior-informed learning, and residual learning, which we compare with the baseline approach of direct learning, demonstrating that prior-informed and residual approaches achieve comparable surface reconstruction accuracy with up to 3 times fewer samples than direct learning.