An Iterative Machine Learning Approach for Discovering Unexpected Thermal Conductivity Enhancement in Aperiodic Superlattices

Machine learning algorithms and materials informatics-based methods have shown increasing effectiveness in engineering applications ranging from optimization of materials properties to guiding design of nanostructures with targeted transport characteristics. Although machine learning methods have shown varying degrees of success in finding optimal solutions when the governing physics is well understood, it is still difficult to use these methods to challenge conventional wisdom and discover new physics due to the traditional “interpolative” nature of machine learning. However, the absence of human bias in machine learning methods makes them highly attractive in searching for materials and nanostructures showing unexpected physical behavior, since the possibility of the search getting constrained within the subspace of known physics due to human intervention is eliminated. In this work, we demonstrate such potential of machine learning algorithms by implementing an adaptive neural network approach that is able to discover unexpected thermal transport behavior in binary periodic and aperiodic superlattices (SLs). The randomization of layer thicknesses of periodic SLs to form aperiodic SLs or random multilayers (RMLs) has been previously believed to lead to a significant reduction in lattice thermal conductivity from the periodic SL due to Anderson localization of coherent phonons. However, it has not yet been elucidated whether certain random distributions of SL layer thicknesses can actually lead to higher thermal conductivity than that of the periodic SLs. Due to the large number of possible RML structures in the design space, a systematic and efficient approach is required to probe the existence of these low-probability-of-occurrence RML structures. Using a convolutional neural network (CNN) based thermal conductivity prediction tool that allows us to rapidly screen a large number of candidate RML structures, we demonstrate the existence of RMLs showing unexpected thermal conductivity enhancement, which contradicts the previously well understood hypothesis. In order to develop an “extrapolative” prediction tool, we adopt an iterative approach to generate a dataset containing structural features which lead to locally enhanced thermal transport in RMLs and include them as additional training sets in each iteration. As a result, our CNN can accurately predict the high thermal conductivities of RMLs that are absent from the initial training dataset, which allows us to identify the previously unseen RML structures exhibiting exceptional thermal transport behavior. The identified RML structures show increased coherent phonon contribution to thermal conductivity owing to the presence of closely spaced interfaces with reduced apparent interfacial thermal resistance. Our work describes a general purpose “extrapolative” machine learning approach for probing low-probability-of-occurrence exceptional solutions within an extremely large subspace and discovering the underlying physics.