From surface area to functionality: data-driven insights into MIL-100(Fe) synthesis for enhanced dye removal...

A traditional MOF design often maximizes generic metrics such as BET surface area and crystallinity, assuming they universally predict performance. In this study, we present a machine-learning-guided optimization framework for MIL-100(Fe) from experimentally synthesized samples. All materials were synthesized via an acid-free, water-based hydrothermal route. We trained small-data ML models to link synthesis parameters, including temperature, time, metal-to-ligand molar ratio, and ion concentration, to key properties comprising surface area, total pore volume, average crystallite size, crystallinity, yield, and methylene blue (MB) removal. SHAP analysis showed that time and the metal-to-ligand molar ratio dominated dye removal, whereas surface area was more sensitive to temperature and time. The most accurate model, Gaussian process regression, was coupled with a genetic algorithm (GA) to optimize synthesis for property-specific targets. Through optimization, the BET-optimized sample increased the surface area from the highest baseline value in the initial experimental dataset, 1748 to 1841.9 m² g^-1, corresponding to a 5.37% relative increase. The MB-optimized sample increased MB removal from the highest baseline value in the initial experimental dataset, 88.6% to 98.3%, corresponding to a 9.7 percentage-point improvement and a 10.9% relative increase. The MB removal optimized sample with a surface area of 1274.3 m² g^-1 and 17.4% crystallinity delivered the highest MB uptake of 98.3%, corresponding to the highest adsorption capacity under the benchmark test conditions. While the optimized sample for surface area reached 1841.9 m² g^-1 (about 44% higher) with 34.9% crystallinity but achieved only 85.1% removal, about 13% lower than that of the optimized MB sample, indicating that even substantial increases in surface area do not govern adsorption performance. A qualitative t-SNE embedding of the descriptor space shows that the optimized samples occupy distinct neighborhoods, elucidating that MOF synthesis should be tailored to the target application rather than a single metric such as surface area.