Paper List

Journal: ArXiv Preprint
Published: Unknown
BioengineeringComputational Physics

Physics-Guided Surrogate Modeling for Machine Learning–Driven DLD Design Optimization

Department of Mechanical Engineering, Lehigh University | Computational Engineering Department, Lawrence Livermore National Laboratory | Department of Industrial and Production Engineering, Bangladesh University of Engineering and Technology | Precision Medicine Translational Research Center, West China Hospital, Sichuan University

Khayrul Islam, Mehedi Hasan, Yaling Liu
Figure
Figure
Figure
Figure
Figure

The 30-Second View

IN SHORT: This paper addresses the core bottleneck of translating microfluidic DLD devices from research prototypes to clinical applications by replacing weeks-long empirical design cycles with a physics-guided machine learning framework that delivers fabrication-ready specifications in under 60 seconds.

Innovation (TL;DR)

  • Methodology First complete inverse design framework for DLD that transforms measured cellular deformability into optimized device geometry through physics-guided machine learning.
  • Methodology Integration of high-fidelity Lattice-Boltzmann/Immersed-Boundary simulations with XGBoost surrogate models achieving sub-degree predictive accuracy (R²=0.9999, MSE=2×10⁻⁴).
  • Methodology Statistical quantification of deformability-geometry interactions via Type II ANOVA revealing significant interaction effects (F=48.23, p<10⁻³⁴) despite geometric dominance of main effects.

Key conclusions

  • Geometric parameters dominate migration angle variance (F=63.72, p<10⁻³⁷), but cellular deformability exerts statistically significant effects through interactions with device geometry (F=48.23, p<10⁻³⁴).
  • The XGBoost surrogate model achieves exceptional predictive accuracy (R²=0.9999, MSE=2×10⁻⁴), enabling sub-degree migration angle prediction across the design space.
  • Bayesian optimization via tree-structured Parzen estimation identifies optimal DLD architectures in under 60 seconds, reducing design iteration from weeks of experimental prototyping to minutes of automated computation.
Background and Gap： Current DLD design relies on empirical scaling laws and iterative fabrication-testing cycles spanning weeks to months, with existing computational approaches limited to forward prediction and lacking end-to-end pipelines for experimentalists.

Abstract: Microfluidic separation technologies have transformed label-free cell sorting by exploiting intrinsic biophysical properties, yet the translation of these platforms from laboratory prototypes to clinical applications remains constrained by the empirical, trial-and-error nature of device design. Deterministic Lateral Displacement (DLD) represents a paradigmatic example: while demonstrating robust discrimination of cells by size, shape, and deformability across diverse applications including circulating tumor cell isolation and malaria diagnostics, DLD performance exhibits extreme sensitivity to the coupled interplay between cellular mechanical phenotype and micron-scale geometric parameters, necessitating iterative fabrication-testing cycles that span weeks to months. We present the first complete inverse design framework that transforms measured cellular deformability into fabrication-ready DLD specifications through physics-guided machine learning. Our approach integrates high-fidelity lattice-Boltzmann and immersed-boundary simulations with gradient-boosted surrogate models to systematically map cellular mechanical properties to migration behavior across manufacturing-feasible geometric configurations (pillar radius, gap, periodicity). Type II ANOVA quantifies the relative influence of these parameters, revealing that while geometric factors dominate migration angle variance (F=63.72, p<10−37), cellular deformability exerts statistically significant effects through interactions with device geometry (F=48.23, p<10−34). The resulting XGBoost surrogate achieves sub-degree predictive accuracy (R2=0.9999, MSE =2×10−4), enabling Bayesian optimization via tree-structured Parzen estimation to identify optimal array architectures in under 60 seconds—reducing design iteration from weeks of experimental prototyping to minutes of automated computation. By deploying this validated pipeline as an accessible web application that accepts experimentally measured deformation indices and returns optimized device specifications with tolerance analysis, we democratize DLD design for researchers without specialized computational expertise, thereby accelerating the translation of microfluidic technologies from research-grade prototypes to application-specific, clinically deployable devices.