Neural-physics multi-fidelity model with active learning and uncertainty quantification for GPU-enabled microfluidic concentration gradient generator design

Abstract

The microfluidic concentration gradient generator (μCGG) is an important biomedical device to generate concentration gradients (CGs) of biomolecules at the microscale. Nonetheless, determining their operational parameter values to generate complex, user-specific, biologically desired CGs is not trivial. This paper presents a neural-physics multi-fidelity model (NP-MFM) to predict CGs with equivalent accuracy as high-fidelity CFD simulation at ultra-fast computational speed through a novel uncertainty- and distance-based active learning process. The verified NP-MFM, along with the genetic algorithm, is implemented on a GPU platform to search optimal values of operational parameters that generate CGs closely matching user-prescribed profiles. Results show that the NP-MFM is a feasible multi-fidelity modeling approach for rapid and accurate prediction of CGs (with 0.019s/simulation) and can be used for GPU-enabled μCGG design optimization and automation. Furthermore, design CGs generated by the proposed method match user-prescribed CGs very well with an averaged discrepancy less than 0.34.