SciPy 2023

A Graph-Neural Network-Based model for rapid prediction of Thermal Transport in Metal-Organic Frameworks
07-14, 13:55–14:25 (America/Chicago), Grand Salon C

Metal-Organic Frameworks (MOFs) have vast potential for gas adsorption, but their practical use hinges on their ability to dissipate thermal energy generated during adsorption. Here, we performed the first high-throughput screening of thermal conductivity in over 10,000 MOFs using molecular dynamics simulations. Next, we developed a graph neural network (GNN) based model to swiftly predict the diagonal components of the thermal conductivity tensor for accelerated materials discovery. Attendees will gain insights into how GNNs can be trained to predict material tensor properties, benefiting both the materials science and machine learning communities.


Metal-organic frameworks (MOFs) are a promising class of porous materials that have potential applications in various areas, including gas storage and separations. However, effective thermal energy management in MOFs is critical to enhancing their performance in these applications. Unfortunately, there is still a lack of understanding regarding the structure-property relationships that govern thermal transport in MOFs.

In order to provide a data-driven perspective on these relationships, a large-scale computational screening study was conducted to investigate the thermal conductivity of MOFs. This study utilized classical molecular dynamics simulations to calculate the thermal conductivities of 10,194 hypothetical MOFs generated using the Topology-Based Crystal Constructor (ToBaCCo) code developed in Python. These MOFs comprised 1,015 different topologies, along with 40 types of organic edge building blocks and 38 inorganic and organic nodular building blocks.

The study discovered that high thermal conductivity in MOFs is favored by high densities, small pores (<10 Å), and four-connected metal nodes. Moreover, it identified 36 MOFs with ultra-low thermal conductivity (<0.02 W/mK) primarily due to their extremely large pores (~65 Å). Additionally, the study uncovered six hypothetical MOFs with exceptionally high thermal conductivity (>10 W/mK).

To handle a large number of MOFs screened, an algorithm was developed to adaptively determine the appropriate plateaued interval of the thermal conductivity vs. correlation time curve based on a set of criteria. The search strategy utilized for finding the optimal plateaued interval involved iteratively performing linear fitting to data segments of 2 ps in length at 1 ps increments if the data was between 0 and 10 ps, and segments of 10 ps length at 5 ps increments if the data was beyond 10 ps. The normalized slopes and normalized average oscillation amplitudes were then calculated with respect to the average thermal conductivity for each of those data segments.

Using the 10,194 MOF-thermal conductivity data, a range of state-of-the-art graph neural network-based models, including CGCNN, iCGCNN, MEGNet, DimeNet++, ALIGNN, and others, were trained for the rapid prediction of thermal conductivity in MOFs. Finally, the model that demonstrated the best performance on the test data was applied to screen the Computation-Ready, Experimental (CoRE) MOF database, resulting in the identification of experimentally viable MOF structures with potentially exceptional thermal transport properties.

This talk will discuss the ToBaCCo hypothetical MOF crystal generation algorithm, various state-of-the-art GNN architectures, and their implementation in PyTorch. This presentation will be of interest to the wider material science community, particularly those with a passion for deep learning models. The findings of this study have the potential to enhance our understanding of thermal transport in MOFs, paving the way for the development of more efficient MOFs for gas storage and separation applications.

Currently, I am pursuing a Ph.D. degree in Chemical Engineering at the University of Pittsburgh with an expected completion date of January 2024. My current research at Pitt focuses on understanding nanoscale thermal transport physics in Metal-Organic Frameworks (MOFs), a class of porous materials, which have been heralded as revolutionary materials for gas adsorption applications. In my research, I use high-performance computing, deep learning, and computational materials science/chemistry techniques.