Foundations of Molecular-to-Systems Design: Frameworks for Data-Driven Multiscale Process Systems Engineering

Multiscale process systems engineering across molecular to infrastructure scales is essential to realize sustainable technologies. For example, the promise of nanoengineering has been slow to manifest because materials such as novel membranes are rarely evaluated in the context of a process at scale to assess sustainability, feasibility, and economics of their application in realistic conditions. In this dissertation, I propose a molecular-to-systems engineering framework to address this challenge. This framework is built on four pillars which are rooted in computational science and data-driven methodologies, and are general and widely applicable to virtually all facets of sustainable technology development:

1. Materials-process targeting uses structure-property relations, thermodynamic laws, and conservation equations quantify the feasibility limits of integrated processes. This paradigm may be used to identify the scale-up potential of technologies based on novel materials such as adsorptive membranes.

2. Data-driven design of experiments utilizes nonlinear parameter estimation and leverages time-series data from dynamic experiments to fit differential-algebraic equation models for the system. This technique, in collaboration with experimental scientists, enables the efficient building of predictive models via faster turnaround in experiments.

3. Superstructure optimization provides a mathematical framework to search through several alternative flowsheet configurations simultaneously to find the optimum topology. In this work, this tool is used to quantify complex tradeoffs across the materials- and the systems-scale to guide property targets for the development of novel materials.

4. Bayesian uncertainty quantification using hybrid models combine partially known (assumed) mechanistic models with a data-driven Gaussian processes (GP) and random noise to account for epistemic (model-form) and aleatory (paramertric) uncertainties. In conjunction with stochastic programming, this works demonstrates the superior performance of Bayesian hybrid models for data-driven decision-making under uncertainty.

In this dissertation, I describe the development of the materials-process targeting models, superstructure optimization framework, and Bayesian uncertainty quantification methods to advance the state of art in multiscale process systems modeling. I briefly discuss how this work provided the impetus for data-driven design of dynamic experiments which sped up membrane characterization experiments. I conclude with a discussion of the future directions of this work and highlight the generality of the framework by identifying potential industrial applications of the proposed framework.