(502f) Toward Universal Prediction of Adsorption Isotherms in Metal-Organic Frameworks Using Deep Learning Algorithms

Machine learning is emerging as a tool that can further accelerate the screening of material databases over traditional molecular simulation. For example, machine learning models can be trained to predict the adsorption properties of metal-organic frameworks (MOFs) for applications such as gas storage and chemical separations. However, previous efforts to predict adsorption using machine learning have focused on developing algorithms using simulated data for a specific application at a single operating condition. Therefore, if the need emerges to evaluate materials for a different operating condition (e.g. different pressure, temperature or gas composition), or for a different application (involving different adsorbates) extensive molecular simulation needs to be done again to enable the training of a new algorithm. To overcome this issue, in this contribution we demonstrate the viability of predicting adsorption loadings for different adsorbates at different conditions using a single model.

To do so, we demonstrate the prediction of full isotherms (in the 1-100 bar range) of small adsorbate gases in a sample MOF database. These predictions were made possible by introduction of adsorbate properties as part of the training data (in addition to MOF properties). Notably, this approach allows our model to predict the loading of adsorbates not included in the training set. We trained our model on a database of 2,000 topologically and chemically diverse MOFs using simulated data for nine different atomic and diatomic adsorbates. Models were trained by a “leave two out” approach. Briefly, in this approach, performance was evaluated by using a version of the model trained on only eight of the adsorbates and a subset of the MOFs to predict the loading of the “left-out” adsorbate in the “left-out” MOFs. That is, no adsorbate or MOF used to train each model version was included in the data used to validate that version. We will discuss the descriptor set (among five tested) that yields the most accurate loading predictions.

Among “as is” algorithms, we found multilayer perceptrons (MLPs) to outperform other models. However, to accomplish the desired accuracy, we found that a hierarchical “committee of nets” approach was appropriate. Briefly, the “committee of nets” consists of a top-level MLP that makes loading predictions using as input the predictions of multiple individual base-level MLPs along with the descriptors used to train these base models. Given its accuracy and scope, our final model can be coupled with traditional mixture-adsorption prediction approaches to rapidly estimate (orders of magnitude faster than traditional molecular simulations) the adsorption loadings of diverse gas mixtures relevant to diverse chemical separations in MOF databases. The combination of these capabilities with our Topologically Based Crystal Constructor (ToBaCCo) code for computational synthesis of MOF prototypes could allow high throughput screening of MOFs at previously inaccessible scales.