Work

Interpretable inverse design of optical multilayer thin films based on global neural adjoint

Year: 2026

Type: article

Abstract: Existing artificial intelligence methods for the inverse design of optical multilayer thin films (OMTs) often function as “black boxes,” offering limited insight into parameter–performance relationships. We propose the global neural adjoint (GNA) framework, which achieves a strong balance across six criteria: accuracy, efficiency, diversity, scalability, flexibility, and interpretability. At its core is the OMT-mixer, a forward architecture that accurately and efficiently predicts optical properties across diverse material configurations, a predefined set of layer numbers, and thicknesses. By incorporating material and constraint loss functions into the neural adjoint loss, the GNA enables scalable and flexible inverse design, while gradient-based sensitivity profiles quantify the accumulated sensitivity of the design loss to the thickness and material index of each layer during optimization. The OMT-mixer achieves accuracy comparable to transformer-based surrogates while reducing the training time per epoch by approximately 37–42%, and by approximately 61% relative to OptoGPT. Evaluated against a genetic algorithm, GNA without material loss, and OptoGPT across six test spectra and six band-pass-filter targets, the GNA achieves mean accuracy comparable to the strongest reference methods while providing significantly better peak accuracy and substantially greater diversity. Under a fabrication constraint fixing the first layer to 100-nm-thick $\hbox {SiO}_2$ , the GNA retains significantly better peak accuracy and greater diversity than OptoGPT with top-k sampling, albeit at a higher computational cost. Gradients in the relaxed material space and the effects of rounding and filtering are quantitatively analyzed, while the reliability of the sensitivity profiles is further validated through gradient-aggregation, perturbation, and ablation analyses. These results establish the GNA as a robust, interpretable, and versatile framework for advanced OMT inverse design.

Source: Scientific Reports

DOI: 10.1038/s41598-026-75220-w

Authors: Sungjun Kim, Jungho Kim

Institutions: Electronics and Telecommunications Research Institute, Kyung Hee University

Language: English

Topic: Machine Learning in Materials Science

Subfield: Materials Chemistry

Field: Materials Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $2,690

Funders: National Research Foundation of Korea, Ministry of Science and ICT, South Korea