External Research, Blog

More and more materials researchers in industry and academia are starting to use data-driven methods to predict materials properties, but peer-reviewed, published data of experimentally validated predictions remain scarce. 

20 Examples of Success

Citrine’s External Research Department, collaborating with Anton Oliynyk of Manhattan College, compiled over 20 prominent examples of materials informatics predictions that have been validated by experiment or physics-based simulation. The work was published in Annual Reviews of Materials Research.

What makes them successful?

Across published case studies, several key ingredients for successful, validated materials informatics efforts were observed: 

  • Using Domain Knowledge
    Complementing materials data with existing physical knowledge (in the form of known empirical models or simulations) can improve the predictive capability of machine learning models
  • Featurization
    Informed representations of materials data (e.g., deriving elemental weighted averages of bulk moduli in alloys from composition, or bond angles from molecular structures) can improve model performance
  • Quality over Quantity in Data
    Quality and self-consistency of training data is more valuable than the sheer quantity of data

The outcomes from this project will advance energy-efficient, sustainable methods of chemical production across a broad range of industries.

Published Example from Citrine

One successful case is a collaboration between researchers at Citrine Informatics and Panasonic to predict molecules with high hole mobility for organic semiconductor applications. A homogenous, high-quality  training dataset of hole mobilities was generated by physics-based density functional theory (DFT) simulations. Predictive models for hole mobility were developed using physical features derived from the molecular structures. These models were then used to predict novel, high-mobility molecules, which were assessed by DFT and used to update the models. Within 35 iterations, this process resulted in a molecule with 25% higher hole mobility than the best in the training set, verified by DFT. The results of the project appeared in the Journal of Physical Chemistry A and the molecules are now patent pending.

Chart showing increased hole mobility in AI guided candidate materials.
Sequential learning results from a joint Panasonic and Citrine collaboration to discover new high hole mobility small molecules for radio-frequency identification (RFID) applications. This project (see case study) resulted in the discovery of materials that outperformed all initial training materials.

Find out more

You can read more about the Panasonic case study here.

You can find out how the Citrine Platform encourages domain knowledge integration here.