In the past decade, optoelectronic devices have seen rapid performance gains across platforms ranging from photovoltaics and light emitters to photodetectors and sensors. Nevertheless, despite remarkable advances in materials and device engineering, their efficiencies and functionalities still lag behind both fundamental limits and application-driven targets. Leveraging light–matter interactions through nanophotonic strategies can not only push devices toward these ceilings but also unlock application-specific properties. In this context, plasmonic approaches that provide intense subwavelength field confinement have shown great potential to uplift emerging technologies and accelerate their impact across solar energy conversion, solid-state lighting, displays, and sensing, while also enabling compact polarization control and chiral-selective functionality [1].

To this end, we theoretically model perovskite-based optoelectronic systems that explicitly leverage localized surface plasmon resonances (LSPRs), using large-scale 3D finite-difference time-domain (FDTD) simulations. We engineered LSPR-mediated near- and far-field effects in all-perovskite tandem solar cells to enhance near-infrared absorption, while allowing thinner active layers and controlled parasitic losses. Our calculations show up to a remarkable 2% absolute increase in power conversion efficiency with randomly distributed, optimized plasmonic nanoparticles (NPs) inside the perovskite layer, considering NP metal, size and concentration [2]. In perovskite light-emitting diodes (PeLEDs), we calculate LSPR-induced local density of optical states (LDOS) enhancement to boost radiative recombination and tune emission directionality, employing optimized metallic NPs in a disordered distribution across the perovskite thin film. We introduce the first 3D model that explicitly captures photon recycling (PR), enabling green PeLED designs with reduced radiative thresholds by jointly exploiting LSPRs and PR. Our calculations predict a fourfold enhancement in external quantum efficiency at a low injection current density of 0.02 mA cm⁻² [3].

Across these studies, optimization was primarily driven by physical understanding of the system, complemented by grid searches and metaheuristic algorithms to navigate parameter space. Yet, as nanostructure geometries and target optical responses become more complex, these strategies can face combinatorial growth, local optima, and prohibitive simulation costs. This motivates a transition to AI-driven inverse design capable of targeting on-demand optical properties far more efficiently. In this regard, we develop a generative framework employing denoising diffusion probabilistic models (DDPMs). Our framework designs chiral nanophotonic metasurfaces, conditioned on target optical responses. Our model is trained on a large dataset of FDTD simulations and conditioned on the desired transmission/reflection (T/R) spectra and circular dichroism (CD) targets, while enforcing fabrication priors such as minimum lithographic feature size. A physics-aware selection loop is designed to filter and refine candidates, reducing the number of full-wave simulations required compared with classical optimization algorithms. Furthermore, we are able to generate free-form geometries beyond predefined shape families. The generated structures achieve CD values exceeding 0.15 in transmission and reflection across the 350–1200 nm spectral range. These results broaden the design possibilities for nanostructures with tailored optical functionality in optoelectronic materials and devices  [4].