Research Theme #1: Integrated multiscale modeling and simulation of materials and process systems design: “How to quantify the enterprise-wide profit sensitivity of catalyst activity?”
In 2018, NASA announced its future vision for 2040. They predict that multiscale modeling and simulation approaches will enable cost-effective, rapid, and revolutionary design of fit-for-purpose materials, components, and systems in the manufacturing industry[i]. This multiscale model-based design methodology will allow for the generation of optimal designs that consider synergistic benefits from combining length and time scales from atomic properties to enterprise-wide factors simultaneously—from the beginning of the design workflow. Furthermore, stages of the product/process development lifecycle are seamlessly joined across various communities: DFT (atomic scale) through MD (molecular scale) in the chemistry/physics community, KMC (molecular scale) through CFD (fluid scale) in the catalyst community, and process scale through enterprise scale in the process community. A unified multidisciplinary design platform guarantees greater understanding and certification of a novel design. In particular, the chemical industry has paid great attention to this technology to reduce reliance on experimentation for model verification (~50% of time and cost of model-based engineering)[ii].
In spite of improvements in computational science and high-performance computing to process design[iii], modern science is incapable of realizing fully integrated multiscale modeling and simulation to address the complex multi-scale design problem1. Progress is limited by persistent computational challenges that must be overcome to interlock different length and time scales and to develop efficient optimization methods. As shown in Figure 1, key challenges are the underdevelopment of coupling methods to link length and time scale, difficulty quantifying how uncertainty propagates throughout different scales, and lack of a robust optimization methodology that bridges across not only scales but also design phases. Currently, it is not possible to quantify the enterprise-wide profit sensitivity of a catalyst’s activity. To address this problem, building on my broad modeling and simulation experience across the chemistry[iv],[v], catalyst[vi],[vii], and process systems communities[viii],[ix], my research group will develop multiscale modeling and simulation methodologies to more accurately develop a complex multiscale model-based process design and optimization.
We will first conceptualize physics-based models that interconnect quantum-molecule-fluid-process systems, reducing the dimensionality of each scale will allow us to consider multiple scales simultaneously, and the coupling of multiple boundary conditions (Figure 1b,d,e). To determine which variables, have to be coupled to yield self-consistent results, we have to understand the impact of complex microscopic dynamics on macroscale dynamics1. To rapidly screen for various combinations of interfacing variables, high throughput supercomputing resources will be used. Quantifying the propagation of uncertainty throughout different scales allows us to estimate the effects of multiscale models for key uncertainty sources, thereby robust design and optimization can be performed. Uncertainty quantification will be incorporated in the methodology through my recently developed Bayesian inference workflow for computationally intensive models (Figure 1a)5,[x].
As the goal is to perform
simultaneous optimization, we will develop a black-box optimization technique that
handles computationally expensive multi-disciplinary design problems (Figure
1c). This approach will enable us to simultaneously optimize a number of design
objectives with minimal computation. By redefining multiscale models from different
disciplines (not scales) as decomposed optimization problems and extract the
maximum “physicochemical information” for constructing prior knowledge of system
behavior through information theory. This expressive surrogate model will be developed
with machine learning; in particular I will use a variational autoencoder (VAE)
based on deep learning technology[xi], to extract
latent variables to be used for design variables.
[i] Liu, X., Furrer, D., Kosters, J., & Holmes, J. (2018). Vision 2040: a roadmap for integrated, multiscale modeling and simulation of materials and systems.
[ii] http://stephanopoulos-symposium.mit.edu/wp-content/uploads/sites/30/2017/06/04-Process-Modeling-Where-to-Next-Pantelides.pdf
[iii] Zhang, T., Sahinidis, N. V., Rosé, C. P., Amaran, S., & Shuang, B. (2019). Forty years of Computers and Chemical Engineering: Analysis of the field via text mining techniques. Computers & Chemical Engineering, 129, 106511.
[iv] Na, J., Kim, J.-W., Kim, K., Harinath, E., Jiang, M., Lee, J.-M., Trout, B. K., Braatz, R. D. (2019) Continuous manufacturing of thin films for pharmaceutical applications. Part I: mathematical modeling and probabilistic uncertainty analysis. In Preparation (pre-printed).
[v] Na, J., Park, S., Bak, J. H., Lee, D., Yoo, Y., Kim, I., Park, J., Lee, U., Lee, J.M. (2019) Bayesian parameter estimation of aqueous mineral carbonation kinetics. Industrial & Engineering Chemistry Research. 58(19), 8246-8259.
[vi] Nguyen, D.L.T., Lee, C.W., Na, J. et al. (2019) Understanding suppressed hydrogen evolution using surface layers for selective CO production in electrochemical CO2 reduction, Applied Catalysis B: Environmental, under review.
[vii] Na, J., Kshetrimayum, K. S., Lee, U., Han, C. (2017) Multi-objective optimization of microchannel reactor for Fischer-Tropsch synthesis using computational fluid dynamics and genetic algorithm. Chemical Engineering Journal, 313, 1521-1534.
[viii] Na, J. et al. (2019) General technoeconomic analysis for electrochemical coproduction coupling carbon dioxide reduction with organic oxidation. Nature communications, accepted.
[ix] Na, J., Jung, J., Park, C., Han, C. (2015) Simultaneous synthesis of a heat exchanger network with multiple utilities using utility substages. Computers & Chemical Engineering. 79, 70-79.
[x] Na, J., Bak, J.H., Sahinidis, N.V.* Bayesian inference for computationally intensive models via adversarial networks and low-complex surrogate model. 2019, pre-printed.
[xi] Na, J., Jeon, K., & Lee, W. B. (2018). Toxic gas release modeling for real-time analysis using variational autoencoder with convolutional neural networks. Chemical Engineering Science, 181, 68-78.
