Graph machine learning for molecular property prediction and design

Molecules with optimal properties are essential for chemical engineering. However, the search for promising molecules with desired properties – which can also lead to more efficient chemical processes – is often limited by missing property data, leading to the need for predictive models. Based on a graph representation of molecules with atoms as nodes and bonds as edges, graph machine learning (ML) has recently emerged as a powerful approach for predicting molecular properties and exploring the chemical space. In this dissertation, we therefore utilize graph ML to advance the identification of optimal molecules for chemical engineering applications. We first develop graph neural networks (GNNs) to predict molecular properties that are highly relevant for chemical engineering. Our developed GNN models provide highly accurate predictions of pure component properties, such as normal boiling points and biodegradability, and mixture properties, e.g., activity coefficients. The GNNs are applicable to a wide spectrum of molecules and can be readily transferred to predict other properties of interest. To further enhance the predictive quality of GNNs, we incorporate thermodynamic relations into the model architecture and training. Specifically, we propose thermodynamics-informed GNNs that learn thermodynamics through regularization during model training, and thermodynamic-consistent GNNs that predict fundamental thermodynamic potentials, such as the Gibbs free energy, from which related properties can be deduced using automatic differentiation. Using activity coefficients as a prime example, we demonstrate that the GNNs provide thermodynamic consistent predictions with increased accuracy and generalization capabilities, paving the way for combining ML with thermodynamics.Targeting the design of molecules with desired properties, we develop and apply a graph ML computer-aided molecular design (CAMD) framework. The framework combines GNNs with generative graph ML and optimization in a modular way. We use generative models, i.e., variational autoencoders (VAEs) and generative adversarial network (GAN), to learn a continuous molecular space that enables strategic sampling of novel molecules using optimization approaches, such as Bayesian optimization (BO) and genetic algorithms. The properties of these molecules are then predicted by GNNs. Thereby, we provide a data-driven CAMD framework that enables automated design of molecules based on available property data. We apply our framework to the design of high-octane fuels and identify well-known octane enhancers as well as promising new fuel candidates, one of which we investigate in engine experiments, demonstrating an important step towards ML-driven molecular discovery. Building on our graph ML CAMD framework, we extend both generative ML models and optimization in molecular design. That is, we propose a generative graph transformer model, called GraphXForm, that constructs molecular graphs with desired properties by sequentially adding atoms and bonds to an initial structure in a self-improving loop. We apply GraphXForm to the design of solvents for liquid-liquid extraction processes, outperforming state-of-the-art generative ML methods while allowing for the consideration of molecular structure constraints and thus increasing flexibility in molecular design. We further propose an optimization-based CAMD approach by formulating ML-based molecular design as mixed-integer linear program to identify molecules with global optimal predicted properties, which is highly promising to increase the sample efficiency in molecular discovery. Overall, we provide predictive and generative graph ML methods to identify molecules with desired properties for energy and chemical systems. This dissertation thus advances both ML and the molecular scale in chemical engineering, while the process scale can be integrated in future work.