Parametric hierarchical matrix approach for the wideband optical response of large-scale molecular aggregates

Davood Ansari-Oghol-Beig, Masoud Rostami, Ekaterina Chernobrovkina, Semion K. Saikin, Stéphanie Valleau, Hossein Mosallaei, Alán Aspuru‐GuzikJournal of Applied Physics 2013 · DOI · arXiv

Introduction

Modeling the optical response of large-scale molecular structures is a cornerstone of modern biophysics and materials science. In green photosynthetic bacteria, light-harvesting complexes known as chlorosomes contain hundreds of thousands of bacteriochlorophyll (BChl) molecules that work in concert to capture and transport energy. Simulating these systems using full electrodynamics is computationally demanding, often requiring solutions across a wide frequency spectrum to generate absorption spectra or time-domain pulse responses.

This paper introduces a novel parametric hierarchical matrix ($\mathcal{H}$-matrix) method designed to accelerate these wideband simulations. By exploiting spectral redundancies in the system's interaction matrices, the authors achieve significant speedups and memory reductions, enabling the simulation of structures containing over 100,000 molecules—a scale comparable to an entire bacterial antenna complex.

The Computational Challenge

Traditional computational electrodynamics faces three primary hurdles when applied to molecular aggregates:

  1. Memory Scale: Modeling 100,000 molecules using the Discrete Dipole Approximation (DDA) results in a dense matrix of equations. Storing such a matrix would require approximately 160 GB of RAM, exceeding the capacity of most standard workstations.
  2. Frequency Requirements: To capture the full optical profile, researchers must solve these equations at hundreds of different frequency points.
  3. Reconstruction Bottlenecks: While standard $\mathcal{H}$-matrix methods can compress these matrices to save memory, they typically require the matrix to be rebuilt from scratch for every frequency point. In wideband studies, this reconstruction process becomes the primary computational bottleneck.

Methodology: The Parametric Innovation

The authors address these challenges by combining DDA with a specialized parametric $\mathcal{H}$-matrix representation.

Standard $\mathcal{H}$-Matrices

Standard $\mathcal{H}$-matrices partition a large matrix into a hierarchy of blocks. "Near-field" blocks (representing close-range interactions) are stored as dense data, while "far-field" blocks (representing long-range interactions) are stored as low-rank approximations. This reduces the computational complexity from $O(N^2)$ to approximately $O(N \log N)$.

The Parametric Extension

The core innovation of this work is the observation that the spatial interaction patterns—specifically the basis of the low-rank blocks—remain relatively stable across a frequency sweep. Instead of re-compressing the matrix at every step, the authors use Singular Value Decomposition (SVD) on a small set of sample frequencies to extract a common, frequency-independent basis.

They then project the frequency-dependent interaction into a significantly smaller "middle" matrix, denoted as $T$. Interpolating this small $T$ matrix is computationally trivial compared to re-calculating the entire hierarchical structure. For the dense near-field blocks, the authors employ a Vector Fitting method—a rational function fitting technique that accurately captures quantum resonances without requiring an excessively dense frequency grid.

Experimental Setup and Models

The method was validated using structural models of Bacteriochlorophyll (BChl) tubular aggregates, which mimic the architecture of natural chlorosomes:

  • Single Roll Model: 2,160 monomers arranged in concentric cylinders.
  • Extended Roll: A 290 nm long structure consisting of 43,200 molecules.
  • Large-Scale Array: A 2x4 array of rolls totaling 103,680 monomers, representing the scale of a real bacterial light-harvesting complex.

Key Results and Insights

Performance and Efficiency

The parametric method demonstrated dramatic improvements over standard $\mathcal{H}$-matrix approaches:

  • Speedup: For a sweep of 350 frequency points, the parametric method provided a 12x speedup for the 100,000-molecule array.
  • Crossover Point: The parametric approach becomes more efficient than standard methods once a simulation requires more than 20 frequency points.
  • Memory Compression: The 103,680-molecule problem was compressed from a theoretical 160 GB (dense) to just 8 GB of memory.
  • Rapid Generation: Once the parametric model is constructed, generating the system matrix for a new frequency takes between 7 and 80 seconds, whereas standard construction would take over an hour.

Scientific Implications

Beyond computational metrics, the study highlights the necessity of full electrodynamic modeling. The authors demonstrate that for structures of this magnitude, the "electrostatic" approximation (which assumes instantaneous light interaction) fails. Electrodynamic retardation effects—the delay in interaction due to the finite speed of light—are significant, causing a redistribution of peak intensities across the molecular structure that simpler models would miss.

Conclusion and Future Work

The parametric $\mathcal{H}$-matrix approach represents a significant leap in the ability to simulate large-scale biological and synthetic molecular aggregates. While the current implementation is serial, the authors note the method is "embarrassingly parallel," suggesting that future GPU or multi-core implementations could yield even more substantial performance gains. Furthermore, the underlying mathematical framework is general enough to potentially benefit other fields involving wideband wave propagation, such as acoustics or fluid mechanics.

Summary updated

Abstract

Fast and efficient calculations of optical responses using electromagnetic models require computational acceleration and compression techniques. A hierarchical matrix approach is adopted for this purpose. In order to model large-scale molecular structures, these methods should be applied over wide frequency spectra. Here, we introduce a novel parametric hierarchical matrix method that allows one for a rapid construction of a wideband system representation and enables an efficient wideband solution. We apply the developed method to the modeling of the optical response of bacteriochlorophyll tubular aggregates as found in green photosynthetic bacteria. We show that the parametric method can provide one with the frequency and time-domain solutions for structures of the size of 100 000 molecules, which is comparable to the size of the whole antenna complex in a bacterium. The absorption spectrum is calculated and the significance of electrodynamic retardation effects for relatively large structures, i.e., with respect to the wavelength of light, is briefly studied.
BibTeX
@article{ansariogholbeig2013parametric,
  author = {Davood Ansari-Oghol-Beig and Masoud Rostami and Ekaterina Chernobrovkina and Semion K. Saikin and Stéphanie Valleau and Hossein Mosallaei and Alán Aspuru‐Guzik},
  title = {{Parametric hierarchical matrix approach for the wideband optical response of large-scale molecular aggregates}},
  journal = {Journal of Applied Physics},
  year = {2013},
  doi = {10.1063/1.4826189},
}
Figure