Sparse Narrow-Band Topology Optimization for Large-Scale Thermal-Fluid Applications

TL;DR

This paper introduces a scalable fluid-based topology optimization method for thermal-fluid applications that efficiently handles millions of design variables, enabling detailed heat-exchanger designs on standard hardware.

Contribution

It presents a novel narrow-band optimization approach that reduces computational cost and reliably produces binary designs without filtering, suitable for large-scale thermal-fluid problems.

Findings

Optimized a heat exchanger with 5×10^7 variables on a desktop workstation.

Achieved efficient large-scale topology optimization using a local narrow-band approach.

Demonstrated the method's effectiveness on multiple thermal-fluid design examples.

Abstract

We propose a fluid-based topology optimization methodology for convective heat-transfer problems that can manage an extensive number of design variables, enabling the fine geometric features required for the next generation of heat-exchangers design. Building on the classical Borrvall-Petersson formulation for the Stokes flow, we introduce an optimization algorithm that focuses computational effort on the fluid-solid interface, where it is most needed. To address the high cost of repeated forward and adjoint analyses and to avoid leakage through nominally solid regions, we exclude fictitious solid voxels from the analysis by imposing the no-slip boundary conditions in the vicinity of the fluid-solid interface. In contrast to the prior approaches, the fictitious solids are also excluded from the global optimization problem via reducing it to a sequence of local narrow-band subproblems with a variable design space. The contribution of our method is that large-scale optimization can be solved efficiently by continuous simplex method while reliably obtaining binary designs without additional filtering or projection. We demonstrate efficiency of the method on multiple examples, including the optimization of a two-fluid heat exchanger at $Pe=10^4$ on a $370^3$ grid comprising $5\times10^7$ design variables using only a single desktop workstation.