Adaptive-resolution chemical discovery strategies for precise and fast computer-aided transition metal complex design.
Abstract
Despite rapid advances and increasing prominence of first-principles, virtual high-throughput screening (VHTS) for catalyst and materials discovery, we know very little about the vast number of feasible compounds in chemical space (as little as 1 part in 1050 has been explored). The highly tunable electronic structure properties of transition metal complexes (i.e., variable spin, oxidation state, and coordination) that make them attractive for wide-ranging applicationsin energy storage and chemical- or electro-catalysis also make their rational design challenging. No unified VHTS tools have been developed for transition metal complexes, but such an effort should take inspiration from two mature VHTS targets where knowledge of suitable descriptors that define the chemical space neighborhood and freely available open source tools have enableddramatic advances: i) organic chemistry for therapeutic drug-design and ii) condensed matter materials science or heterogeneous catalysis. The unique challenge for transition metal complexes is that this chemical space is neither well-defined (e.g., through established descriptors) nor easily enumerable, and open source tools have not, until recently, been available. Another challenge is strong sensitivity of predicted properties to electronic structure method (e.g., density functional in DFT), a critical concern for all of VHTS, but especially in inorganic chemistry, that has been ignored in favor of simplicity and low computational cost. A new VHTS approach that anticipates and corrects for method sensitivity and enables efficient chemical space exploration is essential in transition metal complex design.Our long-term goal is to advance computationally driven, rational transition metal complex design. Our objective is to develop and to disseminate in user-friendly, open-source software malgorithms that simplify and accelerate chemical space exploration while making computational predictions robust. Systematic representation of chemical space will enable i) training machinelearning models and ii) predicting sensitivity to electronic structure method employed, which will drive adaptive optimization workflows that span from cheap, data-driven to expensive but accurate methods. Our preliminary data supports this effort: artificial neural network (ANN) models trained with our inorganic descriptors are predictive of first-principles transition metal complex properties as well as their sensitivity to the method employed. We will advance and demonstrate new algorithms and paradigms for inorganic discovery by 1) Building a descriptor and systematic model training toolkit for transition metal complexes, 2) Developing an algorithm for VHTS that spans from cheap, data-driven models to accurate, but expensive beyond-DFT methods on-the-fly in optimization, and 3) Harnessing new workflows to design transition metal complexes.The expected outcomes of the proposed work will be advanced algorithms for chemical space traversal/mapping and adaptive-theory optimization for the design of transition metal complexes distributed in an open source software package, empowering the broader chemical and materials science communities.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jul 26, 2018
- Source ID
- N000141812434
Entities
People
- Heather J. Kulik
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy