Semi-empirical methods like Parameterization Method 6 (PM6) are valued for their efficiency in computational speed compared to ab initio methods such as Density Functional Theory (DFT). Using empirical data from diverse molecules, these methods offer broad applicability but may lack specificity in accuracy for particular chemical investigations. This research aims to enhance the precision of semi-empirical methods by refining their parameters to more closely align with specific molecular properties such as HOMO-LUMO gaps and Mulliken Spin Density, with a particular focus on a dataset consisting of dicopper complexes [1,2].
By utilizing DFT calculations as a benchmark, this study adjusts PM6 parameters to diminish the discrepancies between semi-empirical predictions and ab initio outcomes. The adjustment process employs different local and global optimization algorithms like the Nelder-Mead method or the Basin Hopping algorithm, which are tested for their efficiency in fine-tuning parameter settings.
While this tailored approach significantly improves PM6s performance on targeted properties, it may alter its efficacy on other attributes not considered in the parameter adjustment. Notably, the optimization process does not require comprehensive DFT data across the entire dataset but only a representative subset.
This investigation not only advances the understanding of parameter optimization in semi-empirical methods but also contributes to the broader effort of creating training sets for machine learning approaches with optimized semi-empirical parameters [3,4].
Philipp Baltruschat