(169cw) Integrating Machine Learning with Evolutionary Algorithms to Design and Discover High-Performing MOFs for Methane Adsorption Using Building Blocks and Crystal Information

Metal-Organic Frameworks (MOFs) are an alluring class of porous crystalline materials. These materials are built from repeated coordinated units of metal clusters and organic linkers. By intentionally selecting what molecular building blocks to use to fabricate these materials, certain properties can be tuned for certain applications. One particular application of interest is gas adsorption. MOFs can be fabricated with high surface area and porosity as well as tunable pore size and functionality which are all desirable for gas adsorption. Methane is of specific interest as it is the main component of natural gas. Adsorbed natural gas (ANG) is a safer and more energy efficient storage method than conventional compressed natural gas (CNG) or liquified natural gas (ANG). Natural gas has been touted as a transition fuel from gas powered to electric vehicles yet these traditional storage techniques are not economically feasible for vehicular applications. Identifying high performing methane adsorbents would greatly reduce this barrier. The vast number of molecular building blocks available also results in a near limitless number of MOFs that could theoretically be fabricated. There could very likely be an unrealized MOF that exceeds current methane adsorption performance.

The goal of this work was to construct and screen high-performing metal-organic frameworks for methane storage. This feat was accomplished through the development of an integrated genetic algorithm (GA) and random forest (RF) machine learning algorithm (GARF). GARF is able to build hypothetical MOFs and predict their methane adsorption using only their molecular building blocks and crystal structure information. While it is infeasible to experimentally fabricate the near limitless number of possible MOFs, MOFs can be constructed and evaluated in silico. By using GARF to specify the building blocks and crystal information, a hypothetical MOF crystal can be generated. Comprehensive molecular simulations can then be employed to calculate methane adsorption. While faster than experiments, simulations are hindered by long convergence times and large computational cost. Machine learning (ML) offers a faster method for predicting methane adsorption. The downside with ML is that there is often a compromise with accuracy. Pardakhti et al. discovered that using structural and chemical properties as descriptors for a RF machine learning algorithm led to highly accurate methane adsorption predictions (1). In addition to accurately predicting adsorption, high-throughput screening of MOFs requires an intelligent method for generating MOF crystals. Evolutionary algorithms can be used to evolve high-performing solutions to problems using the principles of recombination and mutation. GAs, a class of evolutionary algorithms, encode solutions in a data structure known as a chromosome. These chromosomes are allowed to evolve until a solution meeting the desired criteria is found or some other termination criteria is met. The chromosome consists of the building blocks and crystal information needed to build a MOF crystal. In order to function efficiently, GAs require an efficient accurate fitness function to evaluate the chromosomes. In this case, the GARF algorithm was used to replace comprehensive molecular simulations with RF machine learning. GARF was able to screen 250,000 MOFs in mere minutes on a personal computer significantly reducing the computational time while also effectively screening these hypothetical MOFs for their methane adsorption. The resulting decrease in both time and computational cost would be invaluable for materials discovery.

Using solely molecular building blocks and crystal information, the GARF algorithm was able to successfully evolve high-performing MOFs for methane adsorption. Figure 1 illustrates the progression from building blocks to methane adsorption. First, GARF uses the building blocks and crystal information which includes framework number and topology to generate a hypothetical MOF crystal. From this crystal, a chemical formula can be extracted which can be used to calculate several chemical properties. The crystal information and chemical information are then used as descriptors to predict six structural properties. These structural properties along with the chemical properties are used as descriptors to predict methane adsorption. To train the methane adsorption and six structural property random forest algorithms, 50,000 hypothetical MOFs (hMOFs) from the MOFXDB database were used (2). This database contains values for structural properties and methane adsorption calculated via comprehensive molecular simulations along with crystallographic information files (CIFs) detailing their crystal structure. Using this training data, we achieved R2 values above 0.92 for the six structural properties and methane adsorption and relatively low mean absolute error percentage (MAPE) values justifying the use of the RF algorithms for as the fitness evaluator for the genetic algorithm. The GARF model was then able to evolve high-performing MOFs with chromosomes matching MOFs in the top 100 of the hMOF database. After extracting an ensemble of high-performing evolved MOFs generated by GARF, we were able to elucidate information about the ideal chemistries and building blocks for generating MOFs with high methane adsorption. This will allow us to intelligently select new building blocks to add to GARF to evolve novel MOFs. The GARF algorithm rapidly and accurately screens MOFs using minimal input information and evolves high-performing candidate MOFs for methane adsorption opening up a new avenue for materials discovery.

REFERENCES

1. Pardakhti, M., Moharreri, E., Wanik, D., Suib, S. L. & Srivastava, R. Machine Learning Using Combined Structural and Chemical Descriptors for Prediction of Methane Adsorption Performance of Metal Organic Frameworks (MOFs). ACS Comb. Sci 19, 34 (2017).
2. Bobbitt, N. S. et al. MOFX-DB: An Online Database of Computational Adsorption Data for Nanoporous Materials. J. Chem. Eng. Data 68, 483–498 (2022).