Automated Analysis of DFT Output Files for Molecular Descriptor Extraction and Reactivity Modeling

TL;DR

The paper tackles the challenge of mapping molecular structure to reactivity by combining DFT-based descriptors with Hammett-like linear free energy analysis. It introduces DFTDescriptorPipeline, an automated, open-source workflow that parses Gaussian log files to extract electronic, vibrational, and steric descriptors (including HOMO/LUMO, dipole, isotropic polarizability, NBO anchor-based features, and Sterimol parameters) and assembles them into a design matrix for multivariate linear regression. The authors validate the approach across four case studies, reporting LOOCV-based $Q^2_ ext{LOO}$, $R^2$, and RMSE metrics to assess predictive power and interpretability, with several cases yielding strong structure–reactivity correlations and others highlighting limits. The framework emphasizes reproducibility, extensibility, and interpretability, aiming to lower barriers for nonexperts to integrate quantum-chemical data into data-driven molecular design and potentially enable closed-loop discovery in catalysis, materials science, and medicinal chemistry. Overall, the work provides a scalable, transparent platform that translates ground-state quantum descriptors into actionable insights for predicting reaction outcomes and properties.

Abstract

Understanding the relationship between molecular structure and chemical reactivity or properties is fundamental to rational molecular design. Linear free energy relationships (LFERs), particularly Hammett analysis, have long served as powerful tools in organic chemistry. Recently, these approaches have been enhanced by incorporating computationally derived parameters, enabling broader applicability across diverse molecules and reactions. To facilitate and scale this process, we present DFTDescriptorPipeline, a fully automated workflow for extracting quantum chemical descriptors from Gaussian log files and constructing structure-property and structure-reactivity relationships using multivariate linear regression (MLR) models. We validate the workflow across four case studies, including photoswitchable molecules and catalytic reactions. In each case, the models provide interpretable results, demonstrating the versatility of this approach and its relevance to a wide range of chemical contexts. We anticipate that this platform will serve as a generalizable framework for integrating quantum chemical calculations into data-driven molecular design.

Figures (12)

• Figure 1: (a) Classical Hammett analysis linking substituent constants to reactivity. (b) DFT-based featurization incorporating steric and electronic parameters. (c) The present work: DFTDescriptorPipeline automates descriptor extraction, screening, and multivariate linear regression modeling.
• Figure 2: End-to-end pipeline overview of DFTDescriptorPipeline: input preparation, automated descriptor extraction, substituent matching, model construction, and visualization.
• Figure 3: Definition of the seven anchor atoms (a–g) used for NBO analysis. Atom c (C1) is the carbonyl carbon; atom e (C2) is the connecting aryl carbon.
• Figure 4: Calculation of Sterimol parameters ($L$, $B_1$, and $B_5$) along the C1--C2 axis. Bondi radii are applied to define atomic boundaries.
• Figure 5: Thermal Back-reaction of Azoarene photoswitches.