Qinghe Gao
GRAPH MACHINE LEARNING FOR MOLECULAR PROPERTY PREDICTION AND PROCESS DESIGN
Characterizing molecular properties and designing chemical processes are two essential tasks in chemical engineering. Both tasks rely on domain knowledge, empirical experience, and carefully engineered heuristics, and both have been accelerated in recent years by computer-aided methods, including group con tribution, hierarchical decomposition, expert systems, and superstructure opti mization. Deep learning methods now offer new opportunities, with promising applications already reported in property prediction, flowsheet synthesis, process control, and digitization.
While developing standalone deep learning tools for individual applications has produced promising results, these tools often remain limited in their utility for practical engineering work. A typical engineering workflow involves multiple tools and the input of multiple experts, including a property prediction tool to screen candidate molecules, a thermodynamic model to estimate phase behavior, a flowsheet simulator to evaluate process configurations, and a techno-economic or environmental assessment to score the result. Therefore, a centralized system that integrates multiple tools, databases, and experts, and that can iteratively exchange information and feedback across them, is needed to genuinely support the engineer throughout the whole workflow. The overarching objective of this dissertation is therefore twofold: first, to develop a set of deep learning tools that improve on existing methods for molecular property prediction and process design; and second, to identify a methodology to combine these tools, together with existing ones, into an integrated system that supports the engineer across the design workflow rather than at isolated steps within it.
Realizing this objective is constrained by five key challenges: (i) scarce la beled training data, (ii) the choice of data representation, (iii) over-confident predictions, (iv) optimal decision-making, and (v) integration of models. The dissertation argues that graphs are the appropriate representation on which all five challenges can be addressed coherently. Molecules are graphs of atoms and bonds; flowsheets are directed graphs of unit operations and streams; and the engineering workflow that coordinates them can itself be modeled as a graph of interacting agents and information flows. Graph-based models, such as graph neural networks, exploit this shared structure directly, respecting the locality, topology, and permutation invariance of chemical systems.
Building on this idea, the dissertation develops graph-based methods along three tracks: molecular property prediction, process design, and integration. On the molecular side, graph neural networks are shown to be a viable alterna tive to quantitative structure-property relationships in predicting environmental
impact categories with molecular structure information, although categories gov erned primarily by process and supply-chain factors remain beyond the reach of molecule-only models. Self-supervised pretraining at the node, edge, and graph level is introduced as a solution to label scarcity in polymer property prediction, improving R² by more than 30% under extreme data scarcity and revealing that the learned representation is distributed across both the encoder and the subse quent layers. Bayesian uncertainty quantification via Parallel Stochastic Gradient Hamiltonian Monte Carlo is developed as a scalable alternative to ensembles and Monte Carlo dropout, producing sharper and better-calibrated posteriors on thermodynamic benchmarks. On the process design side, flowsheet synthesis is formulated as a sequential decision problem on a directed graph and solved by a hierarchical reinforcement learning agent. Since learning efficiency is a bottleneck in this setting, two transfer strategies are investigated: pretraining on shortcut simulators before fine-tuning on rigorous ones, and initializing from the latent space of a process variational autoencoder. Pretraining on shortcut simulators and transferring the encoder and actor to the rigorous setting cuts convergence time by roughly 64% against the no-transfer baseline, whereas reusing continuous action or critic networks across environments with mismatched reward shaping produces negative transfer. The results suggest that transfer learning serves as a carefully scoped accelerator for RL-based synthesis rather than a general remedy for learning inefficiency. Finally, multi-agent large language model systems are proposed as an orchestration layer above these specialized tools, surveying the required building blocks and outlining a concrete target for coupling molecular and process agents with embedded human oversight.
Returning to the five challenges, the contributions of this thesis collectively make tangible progress on each: graphs provide the relational representation, self-supervised pretraining and transfer learning mitigate data scarcity, Bayesian inference delivers calibrated uncertainty, hierarchical reinforcement learning addresses sequential decision-making, and multi-agent orchestration provides the integration layer. The overarching contribution is not any single method but the demonstration that graph-based learning, pursued consistently across scales, can serve as the connective tissue that begins to close the gap between molecular and process design.
Taken together, the contributions of this thesis suggest a research direction in which the three molecular models are integrated into a single general-purpose graph-based property predictor, in which reinforcement learning research for process design shifts its emphasis from optimization ability to inference general ization, and in which multi-agent orchestration provides the concrete mechanism by which molecular and process decisions are finally coupled into integrated de sign workflows. Looking further ahead, the thesis points toward a future in which
artificial intelligence is not an auxiliary tool invoked at specific steps of chemical engineering design but the substrate on which the workflow itself runs: molecular discovery, property prediction, process synthesis, and techno-economic and envi ronmental assessment coupled on a shared platform that formulates tasks, selects among tools and databases, and consults human experts when needed. The role of the engineer correspondingly shifts from constructing models and executing simulations individually to specifying objectives, interrogating trade-offs, and adjudicating between candidate designs that the system itself has generated and evaluated.
| Publicatiedatum | 20 november 2026 |
| Universiteit | Overig |
| Auteur | Qinghe Gao |
| ISBN nummer | 978-94-6518-488-3 |