A T-matrix database to promote information-driven research in nanophotonics

TL;DR

The Daphona T-matrix portal is introduced, a web-based platform for interactive searching, filtering, and exporting standardized data containing structure-property relations for a wide range of scatterers, as expressed by their T-matrices, and it is demonstrated how the available data enables addressing scientific questions in the broader context of information-driven research.

Abstract

Information-driven methods from machine learning and artificial intelligence for exploring the optical response of metasurfaces and, more generally, photonic systems rely on well-annotated datasets for training. For metasurfaces made from a periodic or aperiodic arrangement of scatterers, the primary information encoding their response is the optical properties of these individual scatterers. In the linear regime, that response is entirely contained in the transition or T-matrix of the individual scatterer. However, despite the widespread use of these T-matrices in exploring advanced photonic materials within the larger community, there is no common infrastructure for distributing them with consistent metadata and a standard representation. That would be important to avoid the repetitive, resource-intensive computation of these T-matrices by researchers worldwide and to enable data-driven research. To overcome this limitation, we introduce the Daphona T-matrix portal at https://tmatrix.scc.kit.edu/, a web-based platform for interactive searching, filtering, and exporting standardized data containing structure-property relations for a wide range of scatterers, as expressed by their T-matrices. Besides introducing this infrastructure, we demonstrate how the available data enables addressing scientific questions in the broader context of information-driven research. The multiple illustrative examples in our contribution cover both surrogate forward models and inverse design models, and operate either directly on the T-matrix or alternatively on optical observables of metasurfaces made from these scatterers.

Figures (7)

• Figure 1: Overview of the main features of the Daphona T-matrix Portal, including upload, validation, storage, search and filtering, visualization, and API access.
• Figure 2: Main exploration interface of the Daphona T-matrix Portal, showing the searchable repository view with structured dataset cards and physics-aware filtering controls.
• Figure 3: Upload, exploration, and export capabilities offered by the Daphona T-matrix Portal through both the graphical user interface and the OpenAPI endpoints.
• Figure 4: Test set evaluation of the surrogate model predicting T-matrices. a) Schematic of the surrogate network: The model maps geometry class and parameters to the real and imaginary parts of the $\mathrm{T}$-matrix of individual scatterers. (b) Real and imaginary parts of the $\mathrm{T}$-matrix elements for the reference ($\mathrm{T}_{\mathrm{true}}$) and the prediction ($\mathrm{T}_{\mathrm{pred}}$), together with the residual $\mathrm{T}_{\mathrm{pred}}-\mathrm{T}_{\mathrm{true}}$ for a representative test geometry: a cylinder with radius 110 nm and height 190 nm. (c) The empirical cumulative distribution function (ECDF) of the reflectance values computed from the predicted T-matrices per geometry class of those samples contained in the test data. (d) Predicted (markers) and reference (solid lines) reflectance spectra $R(\lambda)$ for the same sample at three angles of incidence and two polarizations.
• Figure 5: Test set evaluation of the reflectance surrogate model. (a) Schematic of the surrogate network: the model maps geometry class and parameters to reflectance spectra over wavelength for multiple incidence angles and polarizations. (b) ECDF of the mean-absolute error (MAE) of the predicted reflectance spectra, shown per geometry class. (c) Predicted (markers) and reference (solid lines) reflectance spectra $R(\lambda)$ for a representative test geometry at angles of incidence $0^{\circ}$, $30^{\circ}$, and $45^{\circ}$ for TE and TM polarizations. Predictions are evaluated on a denser wavelength grid than the actual training grid.