Machine-learning-assisted materials discovery using failed experiments.

Area of Science:

• Materials Science
• Inorganic Chemistry
• Machine Learning Applications

Background:

• Inorganic-organic hybrid materials, including metal-organic frameworks (MOFs) and perovskites, are synthesized via hydrothermal and solvothermal methods.
• The formation mechanisms of these materials are not fully understood, leading to reliance on exploratory synthesis.
• Data-driven and simulation approaches offer alternatives to experimental trial-and-error in materials discovery.

Purpose of the Study:

• To develop a machine-learning model for predicting the crystallization outcomes of templated vanadium selenites.
• To utilize historical 'dark' reaction data (failed syntheses) to train a predictive model.
• To identify new conditions for the successful formation of organically templated inorganic materials.

Main Methods:

• Collected data from archived laboratory notebooks detailing failed hydrothermal syntheses ('dark' reactions).
• Augmented raw notebook data with physicochemical property descriptions using cheminformatics.
• Trained a machine-learning model on the combined dataset to predict reaction success.

Main Results:

• The machine-learning model successfully predicted conditions for new organically templated inorganic product formation with an 89% success rate.
• The model outperformed traditional human strategies in predicting successful hydrothermal synthesis outcomes.
• Inverting the model provided new hypotheses regarding conditions favorable for product formation.

Conclusions:

• Machine learning, trained on historical synthesis data, is a powerful tool for accelerating the discovery of inorganic-organic hybrid materials.
• This approach significantly enhances the efficiency and success rate of discovering novel materials compared to conventional methods.
• The predictive model not only guides synthesis but also generates new scientific understanding of material formation.