Machine learning in theoretical and computational chemistry

Contact person: Thomas Bondo Pedersen       
Keywords: Theoretical chemistry, machine learning, partial differential equations, curse of dimensions, data scarcity    
Research group: Hylleraas Centre for Quantum Molecular Sciences
Department of Chemistry    
 

Chemistry plays a crucial role in addressing several significant global challenges. Achieving the green transition, discovering life-saving drugs, and developing energy-saving and environmentally friendly materials all heavily rely on advancements in theoretical and computational chemistry. The realm of theoretical chemistry encompasses two primary aspects: the development of efficient computational tools and the utilization of those tools for chemical discovery and analysis. However, this field faces obstacles that can be described as the dual curse of dimensions, resulting from the intricacies of many-body dynamics and the vastness of chemical space.

The advent of machine learning and data-driven science is revolutionising computational and theoretical chemistry, enabling remarkable progress in all fields of chemistry. However, the field encounters a distinctive combination of challenges encompassing complexity and a severe lack of available data. Consequently, theoretical chemistry presents unparalleled opportunities for the advancement of methodologies in machine learning and computer science.

Research topics:

  • Graph Neural Networks for chemical properties and computations:
  • Novel GNNs for molecular, materials and chemical reaction properties. In particular leveraging the graph categories of heterogenous, multilayer, and hypergraphs
  • GNNs in Differentiable Programming (see below)
  • Machine learning models for the many-body Schrödinger equation:
  • A high-dimensional PDE with complicated symmetry constraints
  • Machine learning force-field models for coarse-grained molecular dynamics
  • Data readiness problem in chemistry:
  • Active learning and Active Data Collection, where data are expensive quantum chemical model calculations
  • Generative models for both chemical properties and manybody wavefunctions
  • Generative models in chemistry
  • Variational Autoencoders for Genetic Algorithms: E.g., learning the genes used by a GA
  • Active Learning strategies for Bayesian Genetic Algorithms
  • Differentiable Programming in chemistry
  • Augment quantum chemistry models with Artificial Neural Networks
  • Convex Regression
  • Learning irregular convex functions over Banach spaces
     

Research team:

Mentoring and internship will be offered by a relevant external partner.