Efficient Data Utilization in Training Machine Learning Models for Nanoporous Materials Screening

Machine learning is becoming a key tool in the study of nanoporous materials and promises to play a continuous crucial role in the discovery of new materials in the foreseeable future. It is important to keep in mind that machine learning is a “data hungry” approach, whose success in any field is predicated on the ability of the pertinent research community to generate and use data as efficiently as possible. In the past ten years, the nanoporous materials community has seen an explosion in the availability of data, mostly due to the application of molecular simulation to calculate adsorption properties in nanoporous materials databases. Thus, the prediction of adsorption properties has served as a natural “testbed” for the application of machine learning approaches to nanoporous materials discovery. However, it is important to put things into perspective and note that while “big data” in other areas (e.g., social media) refers to billions of datapoints, in the nanoporous materials community, data generation (even for adsorption) has rarely hit the million datapoints. The latter holds true despite research groups worldwide pushing their computational resources to the limit. From this perspective, machine learning applications in the field of nanoporous materials have been explored within a context of “data scarcity.” This chapter uses select machine learning efforts to predict adsorption properties from the past eight years (going through them somewhat chronologically) as a point of reference to discuss different topics pertinent to the various decisions that need to be made when attempting to efficiently train a machine learning model to predict a nanoporous material property. An attempt is made throughout this chapter to consistently bridge these different decisions to how they could affect the efficacy with which data is used in model training. The first half of the chapter focuses on the most basic decisions we face when developing a machine learning model to predict material properties. For instance, how to represent the material (i.e., descriptor selection), which materials to use to train the model, and what kind of model to train to make the predictions. The second half of the chapter focuses on more advanced strategies adopted in recent years, seeking to more directly address the data scarcity issue. These strategies include but are not limited to, transfer learning and active learning. The lessons learned during the past eight years are starting to come together, to the point where a single machine learning model can predict adsorption in nanoporous materials as distinct as zeolites, metal–organic frameworks (MOFs), and hyper-cross-linked polymers [1]. But most of these lessons have been learned through the study of MOFs, which is why this chapter primarily focuses on these materials. Finally, we hope that while the discussion of machine learning approaches in this chapter is “anchored” to the examples of adsorption property predictions, our attempt to present the rationale behind different model training aspects or approaches stripped down to their basics can make the insights provided in this chapter somewhat application agnostic and useful for the reader interested in the prediction of nanomaterial properties other than adsorption.