Notes on Successful Projects

Synthesizing and testing new material samples can take months and cost tens of thousands of dollars. Physics-based modeling techniques, such as Density Functional Theory (DFT) are cheaper and quicker. However, these require a large amount of computational power and can still take days. Running a machine learning model takes seconds to minutes and can predict properties over a vast design space. For this reason, AI helps scientists efficiently narrow down the number of materials candidates to model and later synthesize.


A successful example of this process is our work with Panasonic. Citrine worked with their modelling team to develop a new soluble organic semiconductor that could be used in spin coating.

Erin Antono, a member of our Data Science Team, worked closely with Dr. Nobuyuki Matsuzawa, the Senior Principle Scientist at Panasonic, to combine machine learning and physics-based modeling to identify high-performing organic semiconductors, a novel approach to R&D for the Panasonic team.

Erin Antono, Data Science Team Manager

It was really rewarding to see the new Sequential Learning methodology we had been working on be so successful.

Erin Antono

The Challenge – where to look

One of the most challenging aspects to the project was defining the design space – the set of molecules for which property predictions would be made. The team’s initial approach used a subset of materials identified by the Harvard Clean Energy project, but this did not identify promising candidates. Erin and Dr. Matsuzawa put their heads together to try a new approach. Dr. Matsuzawa knew that molecular building blocks called furans, benzenes, and thiophenes showed the high hole mobility needed when fused together in particular shapes. If he could list a set of these building blocks and define the rules about how they fuse together, then Erin and the Citrine team could build an algorithm to explore all possible combinations systematically with AI.

creating a novel molecule design space

The Results

This was a breakthrough approach. The team was now looking in the right place, and the AI model surveyed 500,000 possible semiconductors. Citrine’s AI models recommended a handful of candidates for Dr. Matsuzawa’s team to model.

The team added the modeling results to the data set and retrained the AI model.  This process improved the quality of the AI model and narrowed it down to promising areas.  In total, Dr. Matsuzawa’s team modelled 196 candidates, and 4 of these are patent pending IP for Panasonic. One candidate in particular had a hole mobility 25% higher than previously available materials.

Panasonic materials results funnel

Meeting in person – a great team

Dr. Matsuzawa and Erin co-presented a paper on this success at an MRS meeting. Together they demonstrated the utility of combining AI and molecular modeling.

This work demonstrates the utility of using the Sequential Learning methodology to design experiments for the discovery of novel materials.

Nobuyuki Matsuzawa, Panasonic

You can get more details on the project by downloading our technical case study.