Machine Learning-Driven Synthesis Optimizations in Thin Films and Nanoparticles

Metal-Organic Frameworks (MOFs) have attracted much attention in the materials science community due to their high variability combined with well-defined structure. They are constructed from metal nodes and organic linker molecules to form regular, porous networks. [1]<br/>Their preparation is influenced by several interdependent variables (e.g., metal and linker sources, concentration, solvent, modulators, reaction time and temperature, among others). But unfortunately, their pronounced influence, especially on nucleation, crystallisation processes and the resulting porous crystalline structure, is not yet understood. [2] In the last decade, machine learning (ML) methods have proven to be target-oriented and efficient for solving complex problems that cannot be solved with conventional approaches, such as the discovery and optimisation of synthetic parameter spaces for the production of specific MOFs. [3][4]<br/>The idea of this research project is to use machine learning to establish the development of optimised synthesis processes and parameters. In the first part we will demonstrate the feasibility of ML optimization for several nanoparticle systems with the aim to achieve very small particle sizes with low distribution. The second, much more complex optimization is dedicated to the targeted setting of a specific orientation of a thin film with high crystallinity. This involves a great deal of understanding about the influences on the system as well as highlighting the supporting features of automated synthesis.<br/>Since it is usually too difficult for humans to vary more than one parameter at a time, machine learning algorithms are the perfect choice here to simultaneously optimize a variety of synthesis parameters that are changed simultaneously.<br/>For this purpose, ranges must first be defined for the synthesis parameters within which they are to be varied. In the first instance the parameters are statistically spread for a defined number of experiments. After the execution and characterisation of each experimental result, new parameters are generated by a genetic algorithm, which recombines the previous parameters based on the evaluation of the results.<br/>This step can be repeated as often as desired until the results are close enough to the desired goal. In the final step, all generated data is evaluated for relevance of the parameters again using a machine learning tool. This provides a lot of information both for biasing variables for similar systems and to understand the influences of the variables on the system.<br/>Since the simultaneous variation of several parameters is possible, the time and money saved cannot be dismissed as an advantage. In addition, mathematically sound data are created at the end, which can be referred to for further experiments. And last but not least, if the possibility existed within the chosen parameters, a system optimised to the desired parameters emerged from this method.<br/><br/>References and acknowledgements<br/>[1] Kitagawa, S.; Kitaura, R.; Noro, S. Functional Porous Coordination Polymers. Angewandte Chemie International Edition 2004, 43 (18), 2334–2375.<br/>[2] Moosavi, S. M.; Chidambaram, A.; Talirz, L.; Haranczyk, M.; Stylianou, K. C.; Smit, B. Capturing Chemical Intuition in Synthesis of Metal-Organic Frameworks. Nature Communications 2019, 10 (1), 539.<br/>[3] Maik Jablonka, K.; Ongari, D.; Mohamad Moosavi, S.; Smit, B. Big-Data Science in Porous Materials: Materials Genomics and Machine Learning arXiv e-prints [Online], 2020.<br/>[4] Moliner, M.; Román-Leshkov, Y.; Corma, A., Machine Learning Applied to Zeolite Synthesis: The Missing Link for Realizing High-Throughput Discovery. Acc. Chem. Res. 2019, 52 (10), 2971-2980.