Rapid Detection of Strong Correlation with Machine Learning for Transition Metal Complex High-Throughput Screening

Abstract

<p>Despite its widespread use in chemical discovery, approximate density functional theory (DFT) is poorly suited to many targets, such as those containing open-shell, 3<i>d</i> transition metals that can be expected to have strong multi-reference (MR) character. For discovery workflows to be predictive, we need automated, low-cost methods that can distinguish the regions of chemical space where DFT should be applied from those where it should not. We curate over 4,800 open-shell transition-metal complexes up to hundreds of atoms in size from prior high-throughput DFT studies and evaluate affordable, finite-temperature DFT evaluation of fractional occupation number (FON)-based MR diagnostics. We show that intuitive measures of strong correlation (i.e., the HOMO–LUMO gap) are not predictive of MR character as judged by FON-based diagnostics. Analysis of independently trained machine learning (ML) models to predict HOMO–LUMO gaps and FON-based diagnostics reveals differences in metal- and ligand-sensitivity of the two quantities. We use our trained ML models to rapidly evaluate MR character over a space of ca. 187,000 theoretical complexes, identifying large-scale trends in spin-state-dependent MR character and finding small HOMO–LUMO gap complexes while ensuring low MR character. </p>

Machine learning to tame divergent density functional approximations: a new path to consensus materials design principles

“…(right) Box plot for the fraction of metal-local features (top) and the fraction of electronic features (bottom) for all 23 DFAs at each property. Following our previous work 60,107 , we have categorized χ, Z, O, and L as electronic features, with all remaining features categorized as geometric.…”

“…We employ two subsets of data, as curated in prior work 60 from five prior studies [33][34][35]107,115 that originally corresponded to a total of 2,828 mononuclear octahedral transition-metal complexes in equilibrium geometries obtained with gas-phase density functional theory (DFT). In comparison to the prior curation 60 (i.e., where the sets were referred to as MD1 and OHLDB), we refined the data further by de-duplicating structures with identical molecular graph, charge, and spin state across the two sets. This final filtering step followed the procedure for molecular graph identification described in Ref.…”

“…The filtering procedure applied in previous work 60 required that all B3LYP/LACVP* wavefunctions had limited spin contamination (i.e., ⟨S 2 ⟩ deviations from the expected S(S+1) value < 1.0 ). The number of calculations excluded by this filtering threshold is sensitive to the choice…”

“…As in prior work 60 , we use revised auto-correlations 35 (RACs) as descriptors for all our machine learning models. RACs are sums of products and differences of five atom-wise heuristic properties (i.e., topology, identity, electronegativity, covalent radius, and nuclear charge) on the 2D molecular graph.…”

Molecular excited states through a machine learning lens

“…101 As a solution to the limited data availability, ML models are often developed and trained on much more computationally affordable bandgaps or orbital energies from which bandgaps can be calculated. 85,95,98,[102][103][104][105][106][107][108][109][110][111][112][113][114][115] Such studies are also facilitated by the availability of many big databases with these properties. 100,[116][117][118][119][120] One should be aware that orbital energy gaps are a poor approximation for excitation energies.…”