A topological deep learning framework for graph representation: application to metal-organic frameworks

The performance of metal-organic frameworks is critically determined by their topology, yet traditional methods struggle to automatically and accurately extract topological features.In this paper propose a novel topology-aware message-passing graph neural network that integrates persistent homology into the message passing process via a differentiable gating mechanism.This integration allows the network to dynamically modulate information flow along topologically critical paths, simultaneously capturing local chemical environments and global pore topology.We evaluate our model on two public datasets: the CoRE MOF 2019 dataset for topology classification and the quantum MOF database for CO2 adsorption energy regression.The proposed approach achieves 87.5% classification accuracy (9.4% absolute improvement over the state-of-the-art graph attention network) and reduces the regression root mean square error to 0.25 eV.By enabling more accurate prediction of MOF properties, this work can accelerate high-throughput screening and the discovery of high-performance porous materials.