2021-05-04

Scientists use active machine learning technology to improve photovoltaic panels

By yqqlm yqqlm

Scientists use active machine learning technology to improve photovoltaic panels

Research picture-1: Almost unlimited New discoveries in the space of AML

In the future-oriented research fields such as portable solar cells or rollable displays, organic semiconductor researchers are faced with considerable challenges when dealing with countless possible candidate molecules. test.

Fortunately, in the discovery task of such organic molecules with improved properties, more and more people have begun to consider the power of machine learning through computer simulation or experiment Related training.

Scientists use active machine learning technology to improve photovoltaic panels(1)

Research Picture-2: Limited OSC test space

However, with this possible number of small organic molecules, the scale may be as high as 10^33, making it almost impossible for us to actually generate enough data to reflect the diversity of materials.

The good news is that Professor Karsten Reuter, head of the Theory Department of the Fritz-Harber Institute, has just published their novel solution in the journal Nature Communications.

Scientists use active machine learning technology to improve photovoltaic panels(2)

Research Picture-3: Limited AML exploration in the test space

Active machine learning (AML) algorithms do not rely on learning from existing data, but through continuous iteration to determine the actual need to learn about the problem data.

Based on this, the scientists first simulated several smaller molecules to obtain data related to molecular conductivity (a measure of the practicality of solar cell materials).

Scientists use active machine learning technology to improve photovoltaic panels(3)

Research Picture-4: In The basic principle of AML discovery in almost unlimited space

Then the algorithm will determine whether the small modification of these molecules can derive practical characteristics, or it is uncertain due to the lack of similar data. In this case, the system will automatically request a new simulation, self-improve by generating new data, consider new molecules, and repeat this process continuously.

At present, scientists have shown how to effectively identify new and promising molecules, while algorithms are still continuing to explore the vast molecular space. As a result, we can sort out new types almost every week. The molecular structure helps to make the research and development of next-generation solar cells easier.