Reinforcement learning for group contribution methods

Focus topic:	Artificial Intelligence
Type of funding:	Individual funding programmes
Funded institution:	Rheinland-Pfälzische Technische Universität Kaiserslautern-Landau

The Prisma fund enables Prof. Dr. Fellenz and Prof. Dr. Jirasek, both junior professors at RPTU Kaiserslautern, to carry out preliminary work for a joint project proposal to the DFG.

Goals

The project aims to show that reinforcement learning can also be used successfully in thermodynamics. The aim is to develop group contribution methods for properties of pure substances and mixtures.

Reinforcement learning (RL) is a machine learning technique used to train software to make decisions in order to achieve optimal results. Group contribution methods are a method for estimating substance data in which only the properties of a few dozen structural groups need to be known, instead of requiring the properties of several million substances for the calculation of substance mixture properties.

In a first feasibility study, simple thermodynamic group contribution methods for the vapor pressure of pure substances are considered. The aim is to develop and set up an RL system that makes suggestions for the definition of structure groups, trains and evaluates the model on experimental literature data and ultimately learns the optimal decomposition of substances into structure groups.

Professor Jirasek and Professor Fellenz bring together experience in the development of thermodynamic models and expertise in the development of (deep) machine learning models.