Application of Multi-Task Learning in Analysing the Methane Working capacity of Metal-Organic Frameworks

Abstract

The screening and rational design of metal-organic frameworks (MOFs) with optimal methane capture performance remain a critical challenge for environmental and energy applications. Existing research often emphasizes single-point adsorption capacity, overlooking the dynamic mechanisms relevant to pressure swing adsorption (PSA) required for practical use. To address these challenges, this study developed a multi-task learning (MTL) framework for high-throughput screening of methane adsorption in MOFs. The proposed model integrates attention and screening mechanisms to predict methane working capacity using a dataset of 252,352 MOFs characterized by geometric and chemical descriptors. Shared parameters were leveraged across adsorption prediction tasks at six pressure conditions, achieving high predictive accuracy, with R2 values of 0.992 and 0.962 for gravimetric and volumetric working capacities, respectively. SHAP analysis within the MTL context identified shared underlying mechanisms governing adsorption at both low and high pressures. Key descriptors influencing adsorption capacity included accessible pore volume and specific surface area, alongside chemically relevant atomic identity, covalent radius, and nuclear charge. Variations in feature importance from low to high pressures reflected shifts in methane adsorption mechanisms. This MTL model provides a novel approach to accelerate the discovery of MOFs with high working capacities, offering new insights into methane adsorption mechanisms in MOFs under varied pressure regimes.