Simulating light interaction with complex arbitrary geometry is crucial across the sciences. The Discrete Dipole Approximation (DDA) offers versatility for such problems but faces significant computational challenges, particularly for optically large or high-index systems, limiting its practical scope. Prior circulant preconditioning work, building on frameworks by Chan and Olkin and applied to DDA by Groth et al., demonstrated speedups primarily for quasi-2D geometries while attempts to create stable three-level preconditioners for general 3D structures were unsuccessful. Here we present an efficient and robust DDA implementation featuring a successful three-level circulant preconditioner stabilized through several key enhancements: optimized complex diagonal elements, controlled dimensional expansion and folding of the preconditioner structure, and automated parameter tuning via reinforcement learning. This preconditioning strategy is integrated with a custom GPU iterative solver incorporating stability improvements. Our approach effectively handles arbitrary 3D geometries, including non-homogeneous objects with varying refractive indices and multi-object scenarios with differing material values. The implementation yields substantial computational gains, often exceeding an order of magnitude reduction in iteration count or solution time, enabling convergence for more traditionally difficult problems and reducing demanding simulations from hours to minutes or even seconds on standard hardware. This work significantly extends the range of complex systems amenable to DDA modeling, facilitating advanced electromagnetic simulations relevant to nanophotonics, materials characterization, and atmospheric/biological optics.
扫码关注我们
求助内容:
应助结果提醒方式:
