Inverse design of structural colour devices via machine learning

Abstract

The thesis investigates the field of structural colour inverse design, with potential applications in display, anti-counterfeiting, and full-colour nanoprinting. In the conventional structural colour design, the colours are designed by sweeping the structure parameters. Nevertheless, this manner can only produce limited distinct colours, which hardly satisfy the real-world requirements, due to the finite computation resources. An on-demand inverse design method is, hence, crucial in the structural colour inverse design. Thanks to the fast development of data-driven machine learning in the past decades, the deep learning model, after experiencing a training process, is able to fit any function, making it a strong candidate in the structural colour on-demand inverse design. Our research explores the use of deep learning in accurate and diverse structural colour inverse design. A novel loss function defined in the uniform CIELAB colour space is first proposed to improve the deep learning training efficiency, which is different from the conventional usage of nonuniform CIEXYZ colour space. By employing a tandem network, the new loss function presents obvious accuracy superiority in the inverse design of transmissive Ag-SiO₂-Ag Fabry-Perot cavity structural colour, obtaining an average ∆E of 1.2 in the test sets. The work also revealed the limitations of the tandem network in flexibility, while applying the inverse design of reddish colours. This problem was subsequently tackled by introducing a cGAN-based inverse design approach into the parameter-based structural colour inverse design. Based on the distribution control ability, cGAN is able to force the designs to spread the entire solution space, producing an average ∆E of 0.44 and an average solution group of 3.66 in the test set, which exhibits remarkable structural colour inverse design accuracy and diversity. The study also extends the developed cGAN-based method for the VO₂-based dynamic structural colour inverse design, in which an average ∆E of 0.98 is achieved. By using the ability of cGAN to identify multiple solutions, temperature-induced information encryption is implemented. Furthermore, this technique was utilised for the inverse design of all-dielectric metasurface structural colour to evaluate its universality and effectiveness. The reliability of the cGAN-designed metasurface was valid via a series of numerical simulations, proving that the cGAN had an average ∆E of 0.41. Additionally, the inverse design of Lorentzian line shapes in visible ranges was successfully conducted by employing colour as an intermediary to broaden the potential applications of the cGAN-based inverse design method in versatile optical design. The PhD project analyses the challenges of current popular inverse design methods and proposes a novel cGAN-based inverse design approach for structural colour. We expect that the findings of the thesis can deepen the understanding between nanostructure and structural colour and make contributions to future research and applications in photonic inverse design.

More information

Identifiers

ORCID for Peng Dai: orcid.org/0000-0002-5973-9155

ORCID for Kees De Groot: orcid.org/0000-0002-3850-7101

ORCID for Otto Muskens: orcid.org/0000-0003-0693-5504

Catalogue record

Export record