Machine Learning-Driven Surrogate Modeling and Operating-Point Selection for a Microfluidic Diffusion-Membrane...

Microfluidic diffusion systems provide a powerful in vitro platform for evaluating transdermal drug delivery (TDD), yet their predictive capability is often constrained by limited experimental datasets and nonlinear transport behavior across membrane-device configurations. This study integrates machine learning (ML) with microfluidic experimentation to develop accurate and generalizable surrogate models for cumulative drug permeation under different hydrodynamic and membrane conditions. This work presents an ML-augmented microfluidic TDD framework for predicting cumulative drug permeation from small experimental datasets. Caffeine cream permeation was examined across twelve device-membrane configurations (sMDC, mMDC, and LiveBox2 paired with PET, CA, rat skin, and alginate) at three perfusion flow rates. For each configuration, SVR, MLP, RFR, GBR, XGB, and KNN models were trained and cross-validated using only 33 experimental measurements. SVR showed the strongest overall performance among the evaluated models, achieving test R² values typically above 0.97 and RMSE values of <1-3 µg/cm², accurately capturing the nonlinear time-flow-cumulative mass behavior of TDD profiles. A domain-bounded Gaussian-noise augmentation strategy was used to increase local sampling density while keeping augmented values within the experimentally observed time and cumulative-mass ranges. Polynomial equations were obtained from the predictions of SVR to capture the interaction between inputs and outputs. The trained SVR surrogates were then used for automated steady-state identification and surrogate-based operating-point selection, revealing the dependence of the selected flow rate on membrane permeability and device geometry. Alginate consistently delivered the highest steady-state cumulative mass across all systems (up to ~448 µg/cm²), establishing it as the most efficient TDD membrane among those evaluated. Finally, compact third-degree polynomial equations were derived from the SVR predictions, enabling explicit analytical prediction and rapid design-space exploration. Overall, these ML-derived models and analytical equations provide a fast, low-cost tool for predictive design, enabling rapid microfluidic system evaluation and operating-condition selection, and significantly accelerating the development and screening of next-generation TDD platforms.