MatCreatioNN: Machine Learning Guided Computational Discovery of Photocatalysts for Environmental Applications

Research Approach

The rational design of photocatalysts is often hindered by the high computational cost of traditional screening and the lack of comprehensive experimental data. This paper addresses these challenges by developing an integrated machine-learning pipeline, MatCreatioNN. The researchers used reinforcement learning to generate 120,000 potential MOF structures, which were then filtered through a sequential Crystal Graph Convolutional Neural Network (CGCNN) funnel. This funnel evaluated candidates across 13 key descriptors, including band-gap suitability, CO2/H2O selectivity, adsorption energy, and structural stability, effectively reducing the computational burden by over fourfold.

Key Findings

The framework successfully identified two top-performing candidates—a Cr-based and a Zn-based MOF—that demonstrated superior photocatalytic fitness compared to established benchmarks like PCN-224(Zr). These materials showed improved light absorption and redox properties, which are critical for efficient CO2 conversion. Furthermore, post-hoc analysis identified specific structural motifs, such as the N262 metal cluster, that appear frequently in high-performing candidates, providing a potential design rule for future photocatalyst development. Simulated X-ray diffraction patterns confirmed that these predicted structures are likely synthesizable.

By accelerating the discovery of efficient and durable photocatalysts, this work provides a scalable, data-driven pathway for environmental remediation and CO2 valorization. The use of a multi-objective funnel ensures that the selected materials are not only catalytically active but also economically and environmentally viable. This approach bridges the gap between theoretical design and practical application, offering a foundation for the experimental realization of advanced MOFs in energy-related transformations.