> ## Content Index > Fetch the complete content index at: https://research-pop.com/llms.txt > Use this file to discover other available public pages before exploring further. # USTC and Suzhou Laboratory use dual-loop active learning to design white circularly polarized light - URL: https://research-pop.com/ustc-and-suzhou-laboratory-use-dual-loop-active-learning-to-design-white-circularly-polarized-light/ - Published: 2026-09-21T21:02:20.000Z - Updated: 2026-09-21T21:02:20.000Z - Author: ResearchPOP - Tags: Career, Sustainbility ![](https://storage.ghost.io/c/7c/5e/7c5e0911-739b-4a35-94eb-3e810bf78ba7/content/images/2026/09/Screenshot-2026-09-21-at-22.46.42.png) Source: [https://www.nature.com/articles/s41467-026-77409-z](https://www.nature.com/articles/s41467-026-77409-z?ref=research-pop.com) ## At a glance A light source can appear white without rendering colours well. Producing strong circular polarization across the same broad spectrum adds another design challenge because changing the emitters or layered structure can improve one property while shifting another. Researchers at the University of Science and Technology of China and Suzhou Laboratory connected these tasks through two active-learning loops. One searches for compositions with a high colour rendering index, or CRI, and chromaticity close to the target white point. The other searches for processing conditions that strengthen circular polarization across the visible spectrum. Results measured in the laboratory guide the recommendations made in subsequent rounds. The optimized four-component film reaches a CRI of 90 at CIE coordinates of (0.33, 0.33), with a representative luminescence dissymmetry magnitude of about 1.8 and a spectral fill factor of about 0.8\. A prototype pumped by an ultraviolet light-emitting diode produces white light with a CRI of 92\. The team also fabricated separate devices giving warm, neutral and cool white outputs. ## Background White-light quality has several dimensions. CIE chromaticity coordinates describe the colour of the light itself, while CRI indicates how well that light renders the colours of objects relative to a reference. A balanced white point and good colour rendering depend on the full emission spectrum rather than the colour of a single emitter. Circularly polarized luminescence adds information about the rotating electric field of emitted light. The luminescence dissymmetry factor, gₗᵤₘ, describes the difference between left- and right-circularly polarized emission. Its magnitude is used to compare the strength of that imbalance. The authors also use a spectral fill factor, FF, to describe performance across a wavelength interval, allowing the optimization to consider broad spectral coverage rather than only a strong response at one wavelength. Combining several emitters makes the search more complex. Energy transfer between components can change their relative contributions, so the spectrum of a blend may differ from a simple sum of its ingredients. Molecular alignment and the optical response of an overlying polymer layer also influence polarization. The material composition and processing conditions must both be considered in designing the final output. ## Research question The study examines whether a learning workflow can use a limited experimental dataset to recommend both the ingredients and processing conditions needed for a specified combination of colour and polarization properties. The team begins with three polyfluorene derivatives, PFO, F8BT and PFO-DBT, and later introduces rubrene as a fourth component. The emitters form an oriented film beneath a stretched poly(vinyl alcohol), or PVA, layer. A relative twist between the two layers creates the optical architecture used to generate circularly polarized output. ## Inside the study The first learning loop focuses on composition. ChemDFM, a chemistry-focused large language model, helps identify candidate emitters. Experimental composition and emission data are then used to train a Wasserstein autoencoder, which represents the nonlinear relationship between mixtures and optical performance in a continuous space. Candidate compositions are generated by sampling this space, and a TabPFN regression model ranks them using a score that balances CRI with proximity to CIE coordinates of (0.33, 0.33). ![](https://storage.ghost.io/c/7c/5e/7c5e0911-739b-4a35-94eb-3e810bf78ba7/content/images/2026/09/Screenshot-2026-09-21-at-22.46.52.png) The recommendations are returned to the laboratory. Selected mixtures are prepared and measured, and the new results are added to the dataset before another round begins. For the initial three-component system, the researchers started with 147 measured films. Each of two learning rounds generated 1,500 virtual candidates and selected four for experimental testing, adding eight films to the dataset. The selected ternary film reached a CRI of 87 at CIE coordinates of (0.36, 0.35). The second loop carries the selected composition into a processing search. The PVA layer acts as a phase retarder, while its thickness and stretching degree affect the polarization of light emerging from the oriented emissive layer. With the twist angle fixed at ±45°, Bayesian optimization explores the space defined by PVA thickness and stretching degree. Its selection strategy considers both predicted improvement and uncertainty, allowing the model to explore less familiar regions before refining promising conditions. ![](https://storage.ghost.io/c/7c/5e/7c5e0911-739b-4a35-94eb-3e810bf78ba7/content/images/2026/09/Screenshot-2026-09-21-at-22.47.07.png) For the ternary system, the researchers selected an experimentally accessible PVA thickness of 30 μm and a stretching degree of 25%. Here, stretching degree describes the increase in length relative to the original length. Under 365 nm excitation, the resulting stack gave a representative |gₗᵤₘ| of about 1.7 and an FF of approximately 0.70\. Clockwise and counterclockwise arrangements produced polarization signals with opposite signs. The two loops remain distinct but interact through experimental feasibility and feedback. Polarization results from the processing loop help determine whether a composition is retained and inform later recommendations from ChemDFM. They do not replace the colour-related objective used to rank compositions in the first loop. The connection is made through the films that are prepared, tested and returned to the workflow as new evidence. ![](https://storage.ghost.io/c/7c/5e/7c5e0911-739b-4a35-94eb-3e810bf78ba7/content/images/2026/09/Screenshot-2026-09-21-at-22.47.16.png) The team then introduced rubrene to adjust the spectral balance and expanded the search to a four-component system. Two further composition rounds added eight experimentally tested films. The selected film reached CRI 90 at the target white point of (0.33, 0.33). Processing optimization again led to practical conditions of 30 μm PVA thickness, 25% stretching and a ±45° twist. Under these conditions, the representative |gₗᵤₘ| was about 1.8 and the FF was about 0.8. These measurements describe complementary aspects of the light. CRI and chromaticity characterize colour quality, while dissymmetry and FF describe polarization strength and its distribution across the spectrum. The authors also define a combined comparison metric, M = |gₗᵤₘ| × (CRI/100), which reaches about 1.62 for the optimized film. M is a performance descriptor introduced in the study, rather than a measure of energy-conversion efficiency. ![](https://storage.ghost.io/c/7c/5e/7c5e0911-739b-4a35-94eb-3e810bf78ba7/content/images/2026/09/Screenshot-2026-09-21-at-22.47.25.png) The researchers next placed the optimized fluorescent film and PVA layer above a commercial ultraviolet LED. The LED optically excites the film, which converts the ultraviolet pump light into visible emission. Four emission features near 435, 520, 550 and 625 nm combine to form a white spectrum spanning approximately 400–700 nm. At 3.5 V, the prototype has a reported CRI of 92, CIE coordinates of (0.32, 0.33) and a luminance of about 2,400 cd m⁻². The clockwise and counterclockwise stacks produce closely overlapping emission spectra, while their dissymmetry curves carry opposite signs. The plotted representative values are approximately +1.7 and −1.6\. These measurements connect white-light output with polarization handedness in the assembled colour-conversion device. ![](https://storage.ghost.io/c/7c/5e/7c5e0911-739b-4a35-94eb-3e810bf78ba7/content/images/2026/09/Screenshot-2026-09-21-at-22.47.37.png) The workflow was also used to target different correlated colour temperatures, or CCTs, which describe whether white light appears warmer or cooler. Recommendations derived from the composition–structure–property relationship guided the fabrication of three separate devices rather than continuously adjusting a single device during operation. The warm-white device reached 3,428 K, with a CRI of 81 and a luminance of about 2,600 cd m⁻². The neutral-white device reached 4,153 K, with a CRI of 81 and approximately 2,400 cd m⁻². The cool-white device reached 6,309 K, with a CRI of 86 and approximately 2,200 cd m⁻². Their representative |gₗᵤₘ| values were 1.5, 1.6 and 1.5, respectively, and each device had an FF of about 0.7\. The results show how the design framework can accommodate different colour targets while retaining strong circular-polarization responses. ## Takeaways and outlook The study connects a sequence of decisions that are closely linked in optical-material design. Composition determines the balance between emission bands, while processing controls the layered optical structure. Measurements establish whether each proposed combination delivers the intended colour and polarization properties, and the resulting data guide the next recommendation. The work progresses from a three-component starting point to a four-component film with improved white-light characteristics, followed by an LED-pumped prototype and separate devices designed for different colour-temperature targets. The two learning loops support this progression by selecting experiments and refining their recommendations as new evidence becomes available. The authors present the work as a proof of concept for combining active learning and large language models in the design of circularly polarized luminescent materials. They identify lighting and full-colour three-dimensional displays as potential application directions. More broadly, the study shows how an inverse-design workflow can connect requested optical properties with specific formulations and processing choices that are tested in the laboratory. ## About the researchers Peng Yang (University of Science and Technology of China and Suzhou Laboratory), Xiaoyue He (Suzhou Laboratory and University of Science and Technology of China) and Hongli Zhang (University of Science and Technology of China) are equal-contributing first authors. Mingjun Xiao (Suzhou Laboratory and University of Science and Technology of China), Xin Chen (Suzhou Laboratory) and Gang Zou (University of Science and Technology of China and Suzhou Laboratory) are the corresponding authors. The other authors are Zeyu Feng, Li Wen and Yin Xu from the University of Science and Technology of China, Liyang Wen and Bo Chen from Suzhou Laboratory. ## Original research Peng Yang, Xiaoyue He, Hongli Zhang, Zeyu Feng, Liyang Wen, Li Wen, Yin Xu, Bo Chen, Mingjun Xiao, Xin Chen and Gang Zou. “Design of white circularly polarized luminescence with high gₗᵤₘ across entire visible regions by dual-loop active learning.” *Nature Communications* (2026). DOI: [10.1038/s41467-026-77409-z](https://doi.org/10.1038/s41467-026-77409-z?ref=research-pop.com). Published online 21 September 2026. --- ## Research POP Notes This article reflects the independent interpretation of the Research POP team and does not represent the views of the authors, their institutions or the journal. If you identify any inaccuracies or have concerns regarding the content, figures or attribution, please contact us at [team.researchpop@gmail.com](mailto:team.researchpop@gmail.com). We will review the matter promptly and make corrections or remove the relevant material where appropriate.