Asst. Prof. Xia Kelin, a mathematician specializing in mathematical AI and its use in the molecular sciences. Photo credit – M.Fadly
The development of new materials, from bronze in ancient times to synthetic polymers in the modern day, has always played a major role in the advancement of human technology. Most research in materials science involves expensive and time-consuming trial and error, similar to Thomas Edison combing through thousands of materials to invent his light bulb. However, new methods based on artificial intelligence (AI) may be able to revolutionize this process by allowing scientists to predict the properties of materials before they are synthesized.
Assistant Professor Xia Kelin and Professor Sum Tze Chien, two researchers at the Nanyang Technological University, Singapore (NTU Singapore), have recently made an important advance in AI-assisted materials design. They applied two novel mathematical AI techniques, geometric data analysis and topological data analysis, to the design of an important class of materials known as organic-inorganic halide perovskites. This approach, which had never been attempted before, was found to predict material properties with much better accuracy and efficiency than earlier AI methods.
This breakthrough was the result of an interdisciplinary collaboration between Asst. Prof. Xia, a mathematician specializing in mathematical AI and its use in the molecular sciences, Prof. Sum, an expert in the physics of perovskite materials and their applications. Both are faculty members at NTU’s School of Physical and Mathematical Sciences (SPMS).
The work was reported in a paper published in the journal npj Computational Materials in September 2022.
Materials Science has a Data Problem
Modern materials research heavily relies on computational methods to guide the search for new materials. These methods are critically important because scientists typically face an immensely large number of possibilities to comb through. There are simply far too many candidate materials to synthesize and evaluate one-by-one even for the largest and best-equipped laboratories.
The most widely used and established approach, density functional theory (DFT), calculates the properties of a material by solving the fundamental equations of quantum mechanics, with minimal assumptions. For instance, DFT can be used to calculate a material’s “band gap”, which determines how well electric currents flow within the material. Unfortunately, such calculations require a great deal of computational power, and sometimes do a poor job of predicting experimental results.
In recent years, the rise of Big Data in numerous fields of science and technology has led to a seismic shift in computational materials research. Researchers are increasingly drawn to data-driven methods, which generate predictions not only using fundamental principles but also the analysis of a large amount of real data.
The new approach, which has been dubbed “materials informatics”, still faces numerous challenges. In particular, data-driven algorithms tend to struggle with “featurization”, which refers the problem of choosing the best pieces of information for predicting a material’s properties, from the ocean of raw data available to them.
AI Lights the Way
Professor Sum Tze Chien (left) and Assistant Professor Xia Kelin, two researchers at the Nanyang Technological University, Singapore (NTU Singapore). Photo credit – M.Fadly
Asst. Prof. Xia and Prof. Sum took on the challenge of improving materials informatics by focusing on organic-inorganic halide perovskites (usually just called “perovskites” for short), a new class of optoelectronic materials important for making the next generation of low-cost solar cells, light emitting devices, lasers, and radiation detectors etc.
They had the idea of applying two new mathematical AI techniques, called geometric data analysis and topological data analysis, which had recently been developed by mathematicians but had never been used to analyze perovskites.
The new techniques focus on “invariants” in a data set – features that do not change when the data is transformed, just as pulling at a knot does not untie it. Although the concept is highly abstract, algorithms based on this new approach often perform remarkably well, because of their ability to tease out subtle but important properties hidden within data sets.
The mathematical AI model developed by Asst. Prof. Xia and Prof. Sum predicts the properties of perovskite materials significantly better than competing state-of-the-art machine learning algorithms using traditional molecular descriptors. The benchmark dataset consisted of 1346 high-accuracy DFT calculation results. The team obtained a three-fold improvement in mean square error, a standard measure of prediction accuracy, in predictions of the band gap size.
The researchers also discovered that perovskite properties are influenced by a variety of factors that had never been acknowledged. For instance, an abstract quantity called the Ricci curvature was found to correlate extremely strongly with the size of the band gap (with a Person correlation coefficient of -0.849).
This ability to identify hidden relationships is a major advantage of geometric data analysis and topological data analysis. By comparison, deep learning, the AI method most commonly used in such tasks, acts as a “black box” and is unable to explain how it obtains its results. In the future, the team believes that the insights obtained using their method can be used to guide the development of new perovskite materials.
“The structure-function relationship is the key to unlocking the secrets of perovskite properties”, says Asst. Prof. Xia Kelin. “Our new approach provides a way to translate molecular properties into machine-understandable features. In the next step, we will use these results to design and discover new perovskite materials, which can then be used in practical photovoltaic devices.”
The work was supported by a Singapore Ministry of Education (MOE) Tier 2 grant awarded jointly to Assistant Professor Xia Kelin and Professor Sum Tze Chien.