Machine Learning for Materials Science: 2D Materials Discovery and Design
Nowadays, we are witnessing a tremendous increase in data generation and complexity enabled by advances in experimental, theoretical, and computational developments. This availability of data, associated with new tools and technologies capable of storing and processing that data, culminated in the so-called data-driven science, known as the fourth science paradigm. One of the most prominent areas of artificial intelligence (AI), called machine learning (ML), aims to autonomously identify correlations and patterns in data sets, allowing the extraction of knowledge and insights from it.
However, only recently has the materials science community introduced its application, as to employ these strategies many technical details must be carefully evaluated.
Here in this thesis, in Part I, we show studies combining materials simulations with experimental and theoretical efforts. Specifically, we employ this combination for the pair distribution function (PDF), a technique that elucidates nanomaterials' structure and thus reveals the connection with corresponding properties. We also explore the possibility for the realization of two-dimensional (2D) amorphous topological insulators (TIs) and 2D higher-order TIs (HOTIs), confirming their robustness and proposing the spin Hall conductivity signature to discover novel 2D HOTIs.
In Part II, we show how AI-driven approaches to computational materials science can be exploited to discover and design novel 2D materials for different applications. Specifically, we use machine learning techniques to identify thermodynamically stable 2D materials, which is the first essential requirement for any application. The proposed approach enables the stability and topology evaluation of novel 2D compounds for further detailed investigation of promising candidates, using only composition properties and structural symmetry, without the need for information about atomic positions. Then, we illustrate the applicability of the stable materials, by performing screening of electronic materials suitable for photoelectrocatalytic water splitting.
Finally, in Part III, we show two directions of AI applications that bring all elements together, closing the feedback loop, by using and generating data for real-life situations. One is using nanomaterials' sensors experimental data, towards diverse applications such as cancer diagnostics and chemicals quality control, to discover simple linear equations with high predictive power from these challenging cases of small data volume. The second shows how to employ active learning, which directs the data collection at every stage, to guide the design in the immense space of 2D rotated heterostructures, a new field called twistronics, finding the best properties of interest (in this case, flat electronic bands).
We conclude by highlighting that nowadays data- and AI-driven research is not only viable but increasingly important for materials science, with its own challenges and exciting possibilities for the future.