(678i) Inverse Design with Data-Driven Materials-Process Co-Optimization: Post-Combustion Carbon Capture As a Case Study

Inverse design is a computational materials discovery approach that aims to generate materials design with desired properties by optimizing the generative process on the properties of interest [1]. The concept is appealing as researchers can design materials aiming at achieving desired properties without experimentally synthesizing and testing any and all candidate materials explored during the design iterations. By reducing experimental synthesis and testing efforts, it can save researchers significant time and cost. A challenge in performing inverse design, however, is the vast design space of materials, which often includes infinite possible options to be explored. There are two key factors to discovering outstanding materials among the options; first, how to effectively identify the next candidate material based on the data acquired previously; and second, how to efficiently predict the property of interest for the candidate (which might have never been synthesized before) in every iteration of the design process. In fact, even high-fidelity molecular simulations might be too time-consuming, and hence ruled out, in this context.

In this work, we attempt to not only design materials through inverse design, but to also design related chemical processes through materials-process co-optimization. More specifically, we design metal-organic frameworks (MOFs) through a set of structural descriptors, along with pressure swing adsorption (PSA) process parameters targeting post-combustion carbon capture applications as a case study. To this end, we integrate multiple approaches, including machine learning, derivative-free optimization (DFO), and accelerated simulation through surrogate modeling, to tackle the above two key factors in this application.

Post-combustion carbon capture is an important approach to reduce CO₂ emissions from various sources, such as coal-fired power plants, gas-fired power plants, and the petrochemical industry. This type of carbon capture can be approximated to the separation of N₂ and CO₂ [2, 3]. One of the most established post-combustion carbon capture processes is PSA. However, to bring out excellent performance, the design of an inherently dynamic process like PSA can be complicated, as it involves many process parameters, such as reflux ratios, the pressure and time for each step of a cycle, and flux of feed. Furthermore, packing the PSA beds with efficient adsorbents is also paramount. Here, MOFs are widely considered as promising adsorbents for CO₂ separation [2]. MOFs are an emerging class of porous materials with highly tunable secondary building units. The excellent tunability facilitates a great number of possible MOFs and thus grants us a vast design space. In this case study, our goal is to design the best MOF and PSA combination that can bring out excellent carbon capture performance.

In order to provide experimental synthesis efforts with tangible information about the material structure, we deviate from previous works that attempt to optimize MOFs via means of their adsorption isotherm curves and specifically try to optimize directly the detailed geometric descriptors of MOFs. To tackle the challenge, our group previously established a workflow to obtain a surrogate model that efficiently predicts the corresponding adsorption behavior for given MOF descriptors. The surrogate model can predict the adsorption behavior even when the given MOF descriptors don’t correspond to any existing MOF—which tends to be the norm during an inverse design process. By embedding the aforementioned surrogate model within a PSA simulator [4], we can efficiently evaluate the carbon capture performance for proposed materials iteratively. Next, we formulated an optimization model to allow us to identify material descriptors along with optimal process parameters that optimize a process key performance indicator such as capture productivity (i.e., amount of CO₂ captured per unit time and mass of adsorbent), while meeting constraints on secondary performance indicators as well as bounds to enforce that our surrogate model is not extrapolated. The optimization itself is performed via a DFO methodology.

Our results demonstrate that DFO can efficiently explore the design space based on data it previously acquired and converge with satisfying solutions. The best-found design from our optimization, including the MOF descriptors and the process parameters, shows an 80% enhancement in the capture productivity objective, compared to the baseline (UTSA-16 combined with its optimized PSA process). Moreover, the best-found design possesses a 27% enhancement in the purity and an 18% enhancement in the recovery of the product, with only a modest 8% increase in the total energy requirement (i.e., energy required per cycle per unit mass of CO₂ captured). Furthermore, considering the potential deviation between surrogate model prediction and ground-truth values, we conducted a rigorous sensitivity analysis that entails 5000 perturbed samples to evaluate the co-optimized materials-process combination against plausible carbon capture performance realizations. The results show that the optimized design is robust to such potential deviation under a 95% confidence interval.

In conclusion, the optimized MOF descriptors and PSA process parameters from our inverse design study can further provide materials synthesis experts and process developers with information about the design of MOFs and PSA processes that hold great promise for post-combustion carbon capture performance.

Reference

1. Yao, Z., et al., Inverse design of nanoporous crystalline reticular materials with deep generative models. Nature Machine Intelligence, 2021. 3(1): p. 76-86.
2. Taddei, M. and C. Petit, Engineering metal–organic frameworks for adsorption-based gas separations: from process to atomic scale. Molecular Systems Design & Engineering, 2021. 6(11): p. 841-875.
3. Bui, M., et al., Carbon capture and storage (CCS): the way forward. Energy & Environmental Science, 2018. 11(5): p. 1062-1176.
4. Yancy-Caballero, D., et al., Process-level modelling and optimization to evaluate metal–organic frameworks for post-combustion capture of CO2. Molecular Systems Design & Engineering, 2020. 5(7): p. 1205-1218.