With the demand for renewable energy and efficient devices rapidly increasing, a need arises to find and optimize novel (nano)materials. With sheer limitless possibilities for material combinations and synthetic procedures, obtaining novel, highly functional materials has been a tedious trial and error process. Recently, machine learning has emerged as a powerful tool to help optimize syntheses; however, most approaches require a substantial amount of input data, limiting their pertinence. Here, three well‐known machine‐learning models are merged with Bayesian optimization into one to optimize the synthesis of CsPbBr3 nanoplatelets with limited data demand. The algorithm can accurately predict the photoluminescence emission maxima of nanoplatelet dispersions using only the three precursor ratios as input parameters. This allows us to fabricate previously unobtainable seven and eight monolayer‐thick nanoplatelets. Moreover, the algorithm dramatically improves the homogeneity of 2–6‐monolayer‐thick nanoplatelet dispersions, as evidenced by narrower and more symmetric photoluminescence spectra. Decisively, only 200 total syntheses are required to achieve this vast improvement, highlighting how rapidly material properties can be optimized. The algorithm is highly versatile and can incorporate additional synthetic parameters. Accordingly, it is readily applicable to other less‐explored nanocrystal syntheses and can help rapidly identify and improve exciting compositions’ quality.
20 August 2021 in ACS Biomaterials Science & Engineering
Precise and reproducible control over nanocrystal synthesis is essential for tailoring optical properties, yet remains a long‐standing challenge in halide perovskites. A broadly adoptable machine learning–guided framework, the Synthesizer, is introduced that combines Gaussian Process regression and Bayesian optimization with chemistry‐aware molecular encodings and systematic feature engineering. Rather than new algorithms, the advance lies in translating interpretable machine learning tools into a practical, benchtop platform for nanocrystal optimization under ambient conditions. Using CsPbBr 3 as a model system, nm‐level precision in photoluminescence peak tuning (430 nm to 520 nm) is achieved, along with benchmark narrow linewidths down to 70 meV via lateral confinement control, and robust photoluminescence quantum yield optimization linked to surface trap density. Mapping the two‐dimensional parameter space (Cs/PbBr 2 and antisolvent/PbBr 2 ratios) across multiple antisolvents enables predictive optimization and identifies the antisolvent/PbBr 2 ratio as a previously underappreciated mechanistic parameter, offering a quantitative basis for antisolvent‐accelerated nanocrystal growth. Transfer tests across distinct chemical spaces, including alcohols and cyclopentanone, confirm generalizability to unseen molecules, while application to CsPbI 3 demonstrates extension to new material systems. These results establish an adoption‐ready platform for data‐efficient, uncertainty‐aware synthesis design, providing reproducible pathways to accelerate materials discovery beyond halide perovskites.
In this paper, we propose a novel flexible optimization pipeline for determining the optimal adsorption sites, named AUGUR (Aware of Uncertainty Graph Unit Regression). Our model combines graph neural networks and Gaussian processes to create a flexible, efficient, symmetry-aware, translation, and rotation-invariant predictor with inbuilt uncertainty quantification. This predictor is then used as a surrogate for a data-efficient Bayesian Optimization scheme to determine the optimal adsorption positions. This pipeline determines the optimal position of large and complicated clusters with far fewer iterations than current state-of-the-art approaches. Further, it does not rely on hand-crafted features and can be seamlessly employed on any molecule without any alterations. Additionally, the pooling properties of graphs allow for the processing of molecules of different sizes by the same model. This allows the energy prediction of computationally demanding systems by a model trained on comparatively smaller and less expensive ones.