The failure of available density functional (DFT) methods to compute accurate electronic structures and energies for complex systems is well documented. While newly developed dispersion corrected and long-range corrected density functionals offer improved results, the errors in thermochemical properties computed with DFT are often greater than 5 kcal/mol for many chemical processes. In this talk, we discuss a new fragmentation-based molecular representation framework “FragGraph” for QM/ML methods involving embedding fragment-wise fingerprints onto molecular graphs. Our model is specifically designed for delta machine learning (Δ-ML) with the central goal of correcting the deficiencies of approximate methods such as DFT to achieve a high level of accuracy that is currently only achievable with sophisticated and computationally intensive methods such as coupled cluster theory.<br/><br/>Our framework is based on a judicious combination of ideas from fragmentation, error cancellation, and a state-of-the-art deep learning architecture. Broadly, we develop a general graph-network framework for molecular machine learning by incorporating the inherent advantages prebuilt into existing error cancellation methods such as the generalized Connectivity-Based Hierarchy. More specifically, we develop an attributed graph representation through a fragmentation-based encoding of local chemical environments in the form of fragment-wise molecular fingerprints. In addition, we incorporate electronic descriptors from DFT calculations in the Δ-ML model to achieve high accuracy for the calculation of electronic properties such as redox potentials. Another key feature of our FragGraph model is the interpretability which allows the breakdown of the total Δ-ML corrections into individual group contributions from the structural building blocks comprising the graph network. Our QM/ML framework and its performance for theoretical thermochemistry, redox potentials, and <i>p</i>Ka calculations are analyzed and assessed in this presentation.

Coupled cluster accuracy at DFT cost: Interpretable graph-network-based machine learning models via molecular fragmentation

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.