Bayesian optimization of deterministic lateral displacement devices for circulating tumor cell separation

Circulating tumor cells (CTCs), shed from primary tumors into the peripheral bloodstream, are important biomarkers for early cancer diagnosis and disease monitoring. Deterministic lateral displacement (DLD) microfluidic devices enable label-free CTC isolation through continuous, size-based separation. However, DLD performance is highly sensitive to coupled geometric and operating parameters, making optimal design challenging. Conventional trial-and-error approaches are computationally expensive and provide limited insight into parameter interactions, necessitating data-driven optimization. Here, a Bayesian optimization (BO) framework integrating high-fidelity numerical simulations with Gaussian process (GP) surrogate modeling is developed for efficient DLD design. Using pillar gap size, post-array tilt parameter and flow rate as design variables, BO iteratively constructs a surrogate of the nonlinear target-particle lateral-displacement response and guides adaptive sampling. The GP model achieved high predictive accuracy, with an [Formula: see text] of 0.9873 and a mean absolute error of [Formula: see text] on an independent validation dataset. Compared with Evolution Strategy (ES), BO reduced the evaluations required to identify the best-performing design within the prescribed design space by 67.2% and the time-to-solution by 80.0%, while outperforming Genetic Algorithm (GA), Differential Evolution (DE), Particle Swarm Optimization (PSO) and conventional parametric sweep approaches. The BO-identified design produced a target-particle lateral displacement of [Formula: see text]. Post-optimization particle-size and outlet-region analyses identified an effective critical diameter of approximately [Formula: see text] and demonstrated distinct collection of the investigated NSCLC CTC target surrogate and smaller blood-cell-relevant background-size particles across three release positions. The GP surrogate also enables response-surface mapping, global sensitivity analysis and uncertainty quantification. Results identify the post-array tilt parameter as the dominant factor governing target-particle lateral displacement, while flow rate strongly interacts with geometric variables. Uncertainty analysis confirms stable performance under perturbations. The proposed framework provides an efficient, interpretable, and reliability-aware approach for DLD microfluidic device design.