Multi-fidelity Bayesian optimization of covalent organic frameworks for xenon/krypton separations

Abstract

Our objective is to search a large candidate set of covalent organic frameworks (COFs) for the one with the largest equilibrium adsorptive selectivity for xenon (Xe) over krypton (Kr) at...

Role of multifidelity data in sequential active learning materials discovery campaigns: case study of electronic bandgap

“…larger total budgets), because the costs incurred early in the campaign become too high to offset the benefit of a more predictive high-fidelity ML model. It is interesting to note, perhaps by coincidence, that in the study by Gantzler et al [51], which used multifidelity Bayesian optimization for discovering covalent organic frameworks for gas separations, they found an optimum discovery campaign used 38 low-fidelity and 9 high-fidelity measurements, an acquisition ratio of 38/9 = 4.2, close to our ideal value of 5 discussed above. In their work, the cost ratio (given in terms of calculation time) of low-vs. high-fidelity was 15 min vs. 230 min, a ratio of 15/230 = 6.5%, again quite similar to our cost ratio cutoff of about 5% denoting when the use of multifidelity data becomes most beneficial.…”

Multiscale Modeling of Physical Properties of Nanoporous Frameworks: Predicting Mechanical, Thermal, and Adsorption Behavior

“…We expect such multiscale modeling strategies to play an important role in the materials discovery pipeline, combined with experimental feedback and the use of optimal experimental design, or approaches such as multifidelity Bayesian optimization. 34 More recently, some authors have proposed using deep learning methods in order to predict elastic properties across the entire chemical space of crystals, a harder problem because it needs to take into account both geometrical and chemical considerations. Some authors have used this approach to predict scalar quantities, such as Mazhnik et al 35 who targeted hardness and fracture toughness in a study in order to identify new superhard materials.…”

Accelerating In Silico Discovery of Metal–Organic Frameworks for Ethane/Ethylene and Propane/Propylene Separation: A Synergistic Approach Integrating Molecular Simulation, Machine Learning, and Active Learning

“…However, despite the plethora of screening studies − , available in the literature, there is no consensus on the different aspects that influence the accuracy of these ML models, such as the quality and size of descriptors and training data, the most appropriate ML model, etc. Hence, there is a need to adopt alternative approaches like transfer learning , and active learning (AL) − that are highly efficient and avoid the drawbacks of previously employed methods. Recently, Fanourgakis et al conducted a thorough investigation which brought attention to the extrapolation issues inherent in supervised ML models, and suggested the introduction of artificial MOFs in the training data to improve the efficacy of ML models.…”

“…They further expanded their findings by replacing descriptors with high-and low-fidelity simulations to identify prime candidates for xenon/ krypton separation. 38 Jablonka et al 43 brought attention to the value of pareto-optimal materials and devised an AL-based system to rapidly find them, demonstrating their utility for dispersant-related applications.…”

“…Deshwal et al, on the other hand, used the framework of Bayesian optimization to screen out the best COFs from a pool of ∼70,000 COFs by evaluating only 140 COFs. They further expanded their findings by replacing descriptors with high- and low-fidelity simulations to identify prime candidates for xenon/krypton separation . Jablonka et al brought attention to the value of pareto-optimal materials and devised an AL-based system to rapidly find them, demonstrating their utility for dispersant-related applications.…”