Autonome Entdeckung in den chemischen Wissenschaften, Teil I: Fortschritt

Abstract

Connor W. Coley graduierte mit B.S. in Chemieingenieurwesen am California Institute of Technology und erhielt seinen M.S.C.E.P. und Ph.D. in Chemieingenieurwesen am MIT.Sein Forschungsschwerpunkt liegt auf den Mçglichkeiten, durch Data Science und Automation im Labor die wissenschaftliche Entdeckung in den chemischen Wissenschaften zu rationalisieren. Natalie S. Eyke studierte Verfahrenstechnik an der University of Michigan und schloss 2014 mit dem Bachelor of Science ab. Danach ging sie zu Merck & Co.Inc. an das Chemical EngineeringResearch & Development Department und arbeitete an der Prozessentwicklung fürW irkstoffe. Seit 2017 promovierts ie am MIT unter Anleitung von Klavs F. Jensen und William H. Green. Ihre Forschung konzentriert sich auf die Kombination aktiver maschineller Lerntechniken mit Hochdurchsatzexperimentenfüre in verbessertes Reaktions-Screening. Klavs F. Jensen hat die Warren-K.-Lewis-Professur fürChemical Engineeringa nd Materials Science and EngineeringamMIT inne und ist Codirektor des dortigen Pharma-AI-Konsortiums. Er erhielt seinen MSc in Chemieingenieurwesen von der Technischen Universitätvon Dänemark. Er promovierte in Chemieingenieurwesen an der University of Wisconsin-Madison. Zu seinen Interessen gehçren On-Demand-Mehrstufensynthesen, Methoden fürdie automatisierte Synthese sowie maschinelles Lernen fürdie chemische Synthese und die Interpretation großer chemischer Datensätze. Abbildung 1. Die drei großen Kategorien von wissenschaftlicher Entdeckung, die in diesem Aufsatz beschriebenw erden:E ntdeckungvon Materie, Prozessenu nd Modellen.

Automated, Accelerated Nanoscale Synthesis of Iminopyrrolidines

“…The automation aspect of our work synergizes and is of equal importance to synthesize a large chemical space that is clearly beyond human capabilities. [17] Moreover, it reduces errors, increases speed and safety, leads to better reproducibility, and more efficient and cheaper workflow. [18] Finally, the big data generated can be potentially leveraged by machine-learning software.…”

Benchmarking the Acceleration of Materials Discovery by Sequential Learning

“…One method of accelerating materials research is through integration of automated experiments [1][2][3][4] that are guided by artificial intelligence (AI) 5,6 . Specifically, AI sampling strategies 7,8 hold great promise for resource-constrained activities such as materials research due to their potential to minimize the number of experiments necessary for achieving a desired objective. 9 .…”

“…Recent demonstrations in chemistry include optimization of the drug-likeness and synthesizability of small molecules, 15 the efficiency of organic light emitting diode molecules, 16 . SL methods 7,8 have also been paired with physical experiments to improve the efficiency of materials discovery as demonstrated by a factor of 2 to 5 reduction in the number of experiments required to discover efficient thermoelectric materials, superconductors, steels with high fatigue strength; 17 and to discover new Pb-free BaTiO 3 (BTO) based piezoelectrics with large electrostrains. 18 .…”

“…SL has even been paired with robotic experiments to create a fully autonomous organic reaction search-ing system for exploring chemical reactivity and synthesis of new molecules, 19 and optimization of organic photovoltaics. 20 A recent review 7,8 highlights the breadth of demonstrated applications of SL in the chemical sciences, though there have been relatively limited demonstrations in solid state materials science, where the autonomous optimization of carbon nanotube growth is a seminal example. 21,22 SL paired with robotic experiment has the potential to greatly accelerate experimental discovery of new materials and chemicals, 23 which points to the importance of benchmarking these methods and understanding their behavior in different materials research settings.…”

“…The update could be via an incremental update rule or retraining of the model with the incrementally expanded training data. 7 The incremental update rules used in the former setting restricts the ML models to Bayesian based approaches. In addition, incremental update rules typically focus on improving overall prediction of the model, making this setting less flexible for various research objectives.…”