If the properties of materials can be reliably predicted before they are made, then the process of developing new products for a huge range of industries can be streamlined and accelerated. Unfortunately, the relevant data is often too ambiguous for conventional human interpretation. But increasingly, AI is enabling scientists and engineers to make sense of data that is indecipherable by human researchers alone.
For example, the research was described in the journal Advanced Intelligent Systems. Connecting spectral data to the properties of a material such as electron conductivity, density, stability and optical characteristics — has remained prohibitively ambiguous. However, the Tokyo team used machine learning to reveal information hidden in the spectra of 22,155 organic molecules. The spectra of the molecules and related descriptors were then input into the system. The descriptors are features that can be directly measured in experiments and can therefore be determined with high sensitivity and resolution.
This method is highly beneficial for materials development because it has the potential to reveal where, when, and how certain material properties arise. A model created from the spectra alone was able to successfully predict what are known as intensive properties. However, it was unable to predict extensive properties, which are dependent on the molecular size. Therefore, to improve the prediction, the new model was constructed by including the ratios of three elements in relation to carbon (which is present in all organic molecules) as extra parameters to allow extensive properties such as the molecular weight to be correctly predicted.
This machine learning treatment of core-loss spectra provides accurate predictions of extensive material properties, such as internal energy and molecular weight. The link between core-loss spectra and these extensive properties has previously never been made; however, artificial intelligence was able to unveil the hidden connections. Going forward, this approach might also be applied to predict the properties of new materials and functions. The researchers believe that this model will become a very useful tool for the high-throughput development of materials in a wide range of industries.