Learning from Failure: Predicting Electronic Structure Calculation Outcomes with Machine Learning Models.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• High-throughput screening requires automated simulations for discovering new molecules and materials.
• Transition metal chemistry presents challenges due to complex simulations often needing human intervention and yielding null results.

Purpose of the Study:

• To develop machine learning models for predicting the success of computational simulations in transition metal chemistry.
• To enable fully automated chemical discovery by minimizing time spent on nonproductive simulations.

Main Methods:

• Training support vector machine and artificial neural network classifiers on chemical composition to predict simulation outcomes (geometry optimization, spin state deviation).
• Developing a convolutional neural network model using simulation output time series data for a dynamic prediction approach.
• Implementing a model uncertainty metric based on latent space distribution for improved prediction confidence.

Main Results:

• Static machine learning models achieved an area under the curve of at least 0.95 for predicting simulation success.
• Dynamic models showed improved generalization and predictability with increased simulation input length.
• The developed models effectively minimize computational time on nonproductive simulations.

Conclusions:

• Machine learning models can accurately predict computational simulation outcomes in transition metal chemistry.
• These models facilitate efficient exploration of chemical space and enable autonomous job control for materials discovery.
• The approach significantly advances the potential for fully automated discovery in complex chemical systems.