Research Theme #2: Autonomous discovery of new designs in the chemical process systems by addressing artificial intelligence (AI)
As the use of machine learning approaches in chemical engineering has become mainstream, excitement has grown around the potential—for robust and cost-effective hardware automation in the design of new processes. Interpretable results from modern AI, and statistical machine learning models to extract new insights from large quantities of chemical engineering information—has changed the pessimistic viewpoint of automated discovery: “machines will never be able to make real revolutions”[i],[ii]. We have already seen the confluence of these factors that have developed “Robo-chemist”[iii] which is able to find organic synthesis routes at the level of human chemists; also “Robo-operator[iv]” which is hybrid platform of brain-based cognitive science chip and machine learning-based chip to enable human-level control. These proof-of-concept studies have demonstrated that AI-assisted autonomous product and process design methodology can be realized in near future. Interestingly, petrochemical companies such as ExxonMobil and Dow are making bold investments in academia to develop novel machine learning and artificial intelligence approaches to solve problems, leading to reductions in operating costs[v],[vi].
My group will contribute to the effort to use self-exploration AI (Figure 2) to bridge the gap throughout systematic platform technology, experiments, automation, and modeling that ultimately perform autonomous discovery of new process designs in chemical industries. During my previous research programs, I have focused on developing fundamental platform technologies including automatic process synthesis framework8, improving interpretability of machine learning11,[i],[ii], and robust design and control4,[iii],[iv] to perform elementary-level automatic technoeconomic analysis for screening extensively large processes alternatives (Phase 1). To elaborate technologies in Phase 1, we will create domain-specific physical informed machine learning model which reflect knowledge from chemical engineering, process data, and expert heuristics applied to neural networks that obey physics and other real-life constraints. Then we will address the problem of how to perform optimal experimental design (OED), an approach to assist the AI in efficiently selecting experiments and modeling for discovering new designs.
Even with rigorous multiscale models and experimental techniques with an optimally planned design workflow, exhaustive enumeration of all possible combinations of design alternatives and operating conditions is infeasible. To overcome this issue, we will utilize feature (e.g. chemical properties) extraction to minimize the number of expensive simulations and experimental data. In Phase 2, the way that we will navigate the feature space in a discovery nonlinear representation is automatic feature learning via inference and generative models based on variational autoencoder (VAE) and generative adversarial neural network (GAN) (Figure 2f). Trained networks generate processes and molecules that have target properties: “reverse engineering” without additional simulations and experiments. As a significant advance of research theme #1 and theme #2, I propose a broadly applicable digital-twins platform that reflects actual physical system dynamics (Figure 2c). To illustrate the closed-loop workflow of a digital-twins platform, we will develop a digital systems built on a cyber-physical environment via collecting and analyzing automated physical process systems incorporating multiscale simulations and experimental data. Then the system will decide the optimal design variables such as catalyst/device/process that to secure economic feasibility. Finally, physical systems will validate and transfer suggested optimal design candidates to cyber systems to update parameters. Iterating this cycle over various catalyst/device/process candidates will lead to the rapid discovery of new process design.
A key obstacle of this research program defined in Phase 3 is how to develop an AI that understands all types of structured/unstructured data while securing data privacy throughout collaborators without loss of performance. To do this, my research group will develop a chemical engineering-specific ontological informatics infrastructure that standardizes data representations to integrate and manage all types of information as AI understandable (Phase 3)[v],[vi]. At the final step of the research theme #2, we will apply “Robo-designer” to discover new process designs that are not intuitive to design. The role of humans in this AI-assisted discovery is clear: implement high-level inspiration that never be conducted by machine.
This project will require active collaboration throughout process systems engineering, computational science, and experimental groups. Initially, work will focus on formally extending previous work of automated process synthesis and deep learning technology for better interpretability. During Phase 2, I will actively collaborate with experimental groups to build robot experimentalists to verify multi-step workflow in digital-twins platform. Finally, autonomous design discovery methodology will be demonstrated by actual process development projects, in particular non-traditional sustainable processes, with experiment groups in Ewha Womans University and industrial collaborators, which is directly connected to research theme #3 as follows.
[i] Lee, W. J., Na, J., Kim, K., Lee, C. J., Lee, Y., & Lee, J. M. (2018). NARX modeling for real-time optimization of air and gas compression systems in chemical processes. Computers & Chemical Engineering, 115, 262-274.
[ii] Lee, Y., Na, J., & Lee, W. B. (2018). Robust design of ambient-air vaporizer based on time-series clustering. Computers & Chemical Engineering, 118, 236-247.
[iii] Kim, K., Na, J., Kim, J.-W., Harinath, E., Jiang, M., Lee, J.-M., Trout, B. K., Braatz, R. D. (2019) Continuous manufacturing of thin films for pharmaceutical applications. Part I: surrogate model-based experimental design and maximum-likelihood parameter estimation. In Preparation (pre-printed).
[iv] Kim, J.-W., Na, J., 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 III: stochastic model-based predictive control via polynomial chaos expansion. In Preparation (pre-printed).
[v] Venkatasubramanian, V., Zhao, C., Joglekar, G., Jain, A., Hailemariam, L., Suresh, P., … & Reklaitis, G. V. (2006). Ontological informatics infrastructure for pharmaceutical product development and manufacturing. Computers & chemical engineering, 30(10-12), 1482-1496.
[vi] Venkatasubramanian, V. (2019). The promise of artificial intelligence in chemical engineering: Is it here, finally?. AIChE Journal, 65(2), 466-478.
[i] Anderson, P. W., & Abrahams, E. (2009). Machines fall short of revolutionary science. Science, 324(5934), 1515-1516.
[ii] Jensen, K. F., Coley, C. W., & Eyke, N. S. (2019). Autonomous discovery in the chemical sciences part I: Progress. Angewandte Chemie International Edition.
[iii] Segler, M. H., Preuss, M., & Waller, M. P. (2018). Planning chemical syntheses with deep neural networks and symbolic AI. Nature, 555(7698), 604.
[iv] Pei, J., Deng, L., Song, S., Zhao, M., Zhang, Y., Wu, S., … & Chen, F. (2019). Towards artificial general intelligence with hybrid Tianjic chip architecture. Nature, 572(7767), 106.
[v] https://corporate.exxonmobil.com/Research-and-innovation/University-partnerships
[vi]https://corporate.dow.com/en-us/science-and-sustainability/working-together/universities/carnegie-mellon
