Machine learning could redefine understanding of covalent bonds, say researchers
5 June 2019 | 0
Researchers from AMBER, the SFI Research Centre for Advanced Materials and BioEngineering Research, the School of Physics and the CRANN Institute, at Trinity College Dublin, have developed a new method to for scientists to find out what happens within chemical and biochemical reactions by using articial intelligence.
The researchers have found a way to bypass the traditional way of modelling the atomic world by teaching a computer the underlying physics and chemistry associated with a covalent bond.
Machine learning has enabled the research team to make a breakthrough in modelling – meaning that, through artificial intelligence, computers used to model materials can learn by themselves by reviewing the available data.
This could prove very useful to model experiments for sectors like the aerospace industry where it is difficult and costly to identify and test prototype materials that maintain their properties under very high pressure and temperature.
All materials, including living beings, are made of atoms – the smallest building blocks of the material world. Many models currently exist that can predict what will happen when molecules form covalent bonds – which is a bond that forms when different atoms share electrons.
In order to model what will happen when covalent bonds are formed, scientists use the fundamental equations of quantum mechanics. However this generally requires significant computing power and can take a considerable amount of time to complete.
Prof Stefano Sanvito, a professor in the School of Physics and director of the CRANN Institute at Trinity College, explained: “There are a range of numerical techniques, called first principles methods that scientists traditionally use to simulate how materials behave at the atomic level. These require us to solve the fundamental equation of quantum mechanics.
“While these simulations are usually highly accurate, they need lot of computational resources to complete. In our work we have constructed a range of models that avoid the need for solving the Schrodinger equation, but achieve an identical level of accuracy.
“Using machine learning, which is a branch of artificial intelligence research, it allows us to simulate any material at the atomic level in a shorter amount of time than traditional methods. We have invented a novel way to systematically construct atomistic models for materials, which are as accurate as the computationally expensive first principles approach.”