Data-driven and mechanistic modelling of PLGA nanoparticle formation in hydrodynamic flow focusing microfluidics.

Microfluidic devices offer precise control over polymeric nanoparticle (NP) synthesis. However, rational design of process optimization to obtain a target NP size remains challenging because the relationship between operating parameters, mixing dynamics, and particle formation is nonlinear. Here, we integrate mechanistic analysis, statistical design of experiments (DoE), and machine learning (ML) to develop predictive and inverse design frameworks for the synthesis of poly(lactic-co-glycolic acid) (PLGA) NPs. The effects of flow rate ratio (FRR), PLGA concentration, and microchannel width were systematically investigated using hydrodynamic flow focusing (HFF) microfluidic devices. Mixing time analysis demonstrated that NP size is primarily governed by solvent-to-antisolvent mixing, with equivalent mixing times producing comparable particle sizes despite different combinations of FRR and channel width. A Box-Behnken DoE yielded a highly predictive quadratic model (R2 = 0.9979), identifying microchannel width as the dominant design variable, followed by PLGA concentration and FRR. Eight machine learning algorithms were subsequently developed and evaluated with gradient boosting achieving the highest predictive accuracy (test R2 = 0.9900, RMSE = 2.65 nm). Finally, inverse design experiments were performed for target NP sizes of 80, 120, and 160 nm using DoE and a selected subset of ML models. Process conditions predicted by the developed models were implemented experimentally. The gradient boosting model produced NP sizes in closest agreement with the preassigned targets, outperforming DoE and other ML models. This combined mechanistic and data-driven framework enables reliable prediction and inverse design of PLGA NP synthesis and provides a practical strategy for intelligent and on-demand microfluidic-based synthesis of polymeric nanomedicines.