Edit | x
Over the past few years, data-driven machine learning (ML) technology has become a powerful tool for designing and discovering advanced materials. However, due to the need to consider precursors, experimental conditions, and reactant availability, material synthesis is often much more complex than properties and structures, and few calculations can be achieved experimentally.
In order to solve these challenges, a research team from Southeast University and Zhejiang Normal University proposed a general framework that integrates high-throughput experiments, prior knowledge of chemistry, and machine learning techniques such as subgroup discovery and support vector machine to guide the experimental synthesis of materials, which can reveal the structure-property relationships hidden in high-throughput experiments and quickly screen out materials with high synthetic feasibility from the vast chemical space.
By applying the proposed method to solve the challenging and subsequent synthesis problems of two-dimensional silver-bismuth (AGBI) organic-inorganic hybrid perovskites, the success rate of synthesis feasibility is increased by four times compared with traditional methods. This study provides a practical approach to solve multidimensional chemistry acceleration problems using small datasets from typical laboratories with limited available experimental resources.
The study, titled Universal Machine Learning Aideed Synthesis Approach of Two-dimensional Perovskites in a Typical Laboratory, was published in Nature Communications on January 2, 2024.
*Link: The discovery of advanced functional materials can help humanity address the major global challenges facing it. However, material synthesis is a typically complex, multidimensional challenge that requires experts to evaluate various reaction conditions, such as precursors, additives, solvents, concentrations, and temperatures.
Due to the availability and inherent limitations of chemical precursors and experimental instruments, synthetic chemists can only evaluate a fraction of these conditions during standard optimization activities in a typical and simple laboratory. Again, the exploration of conditions often depends on predefined optimal designs, limited literature on solid-state synthesis reactions, and the experience of chemists.
The past decade has witnessed significant efforts in the discovery of new materials using data-driven technologies, particularly machine learning. However, the use of these techniques to guide the experimental synthesis of new materials is still limited.
Recently, closed-loop automated synthesis frameworks based on machine learning techniques and robotic experiments have been shown to be effective in accelerating the experimental synthesis process, but the experimental cost is high. In addition, many time-consuming experiments can only provide small-scale datasets, which is incommensurate with traditional machine learning methods due to the inherent sparsity and imbalance of available data. Therefore, it is particularly important to develop a framework that integrates machine learning techniques and small-scale experiments to rapidly accelerate the process of materials synthesis for expansion into the field of new materials.
Two-dimensional organic-inorganic hybrid perovskites (2D HOIP) have emerged as one of the most promising functional materials with enhanced environmental stability, excellent optical properties, diverse electronic properties, and easy access and cost-effective fabrication. Inspired by its superior properties, there is a growing interest in developing new, stable, and environmentally friendly 2D HoIP materials.
To date, the design and discovery of novel 2D perovskites have largely relied on traditional trial-and-error methods. With millions of experimentally usable organic molecules and dozens of inorganic frameworks, the unexplored chemical space contains a large number of potentially novel 2D HoIPs, making searches based on traditional trial-and-error methods slow and expensive.
The study demonstrates the synthetic feasibility of two-dimensional silver-bismuth (AGBI) iodide perovskites, which have been proposed for applications in photodetectors, light-emitting diodes, and X-ray imagers. The researchers have developed a framework that combines small-scale high-throughput experiments, quantification of the spatial and topological properties of organic precursors, and machine learning techniques to rapidly screen 2D HOIP with high synthetic feasibility.
Figure 1: Screening framework for two-dimensional silver-bismuth (2D AGBI) iodide perovskites. （*
The quality and quantity of training datasets are the cornerstones for developing high-performance machine learning models. Considering the previously reported organic spacers employed in 2D perovskites, as well as chemical intuition and the commercial availability of amines, 79 promising amines were selected for 2D AGBI iodide perovskite synthesis.
Figure 2: Summary of synthesis results from high-throughput experiments. （*
The results of high-throughput experiments showed that only 13 organic spacers could form a 2D AGBI iodide perovskite structure, and the chemist's intuitive success rate was 164%
In view of the interaction between the inorganic layer and the organic spacer of two-dimensional perovskites, a set of information features for quantifying the spatial and topological properties of organic precursors has been developed. With the help of the sub-emergence method, regions that are more conducive to the formation of two-dimensional AGBI iodide perovskites were derived.
Figure 3: Visualizing the synthephalability of 80 compounds using material descriptors. （*
Then, an equation was obtained by applying ML technology to quantitatively assess the feasibility of the synthesis of 2D AGBI iodide perovskites and** 344 of the 8406 organic spacers had the potential to form 2D AGBI perovskites.
Figure 4: Results and insights from the ML model. （*
A further explainable ML technique, Shapley Additive Explanations (SHAP) analysis, highlights the importance of the molecular topology of organic spacers for 2D AGBI perovskite formation. In the end, 8 of the 13** 2D AGBI iodide perovskites with high synthetic feasibility were successfully synthesized, indicating that the success rate of ML-guided 2D Agbi iodide perovskites can reach 615%, which is much higher than the success rate based on chemical intuition (16.)4%）。
Figure 5: Screening of 2D AGBI iodide perovskites with high synthetic feasibility and experimental validation. （*
In conclusion, this study not only provides a practical method for the rapid discovery of promising advanced functional materials, but also provides a general machine learning-assisted synthesis framework that combines powerful capabilities with physicochemical explainability.
Note: The cover is from the Internet and has nothing to do with the study.