Project 39 : Divide and Conquer : Local Gaussian Processes to design Covalent Organic Frameworks for Methane Deliverable Capacity

Authors

1. Nikhil Kumar Thota (Johns Hopkins University) (https://github.com/T-NIKHIL)
2. Maitreyee Sharma Priyadarshini (Johns Hopkins University) (https://github.com/msharmap)
3. Yiran (Gigi) Wang (Johns Hopkins University) (https://github.com/gigiwang08)
4. Jarett Ren (Johns Hopkins University) (https://github.com/jren0))

---

Method

In this project, we explore the use of local Gaussian Process models to accelerate materials discovery when the search spaces are very large. We evaluate the performance of the framework on a covalent organic framework (COF) dataset that consists of 69,840 2D and 3D COFs [1]. This dataset replicates some real-world scenarios wherein the search space to explore is very large. In this test, we used an initial training dataset comprising 5% of the total search space. These COF structures are designed for methane storage and our optimization target here is the deliverable capacity (v STP/v) of the COF structure. We employ gaussian process surrogates with zero prior mean function and Matern kernel as the covariance function.

Gaussian Process (GP) has been a popular choice of surrogate model in Bayesian Optimization due to its flexibility and uncertainty quantification. However, training a Gaussian Process involves several matrix inversions, which can dramatically scale up the computational cost as more data is obtained via Bayesian Optimization. Gaussian Process has a runtime complexity of $O(n^3)$, where n is the number of training samples. Given its poor scalability, the application of GPs to high-dimensional problems with several thousand observations remains challenging. In this project, we aim to reduce the computational cost of GP-based Bayesian Optimization by breaking a global GP model into several local GP models. These local GP models will run in parallel, accelerating the optimization problem. We have also designed a new acquisition function to aggregate predictions from local GPs and select the next points to explore with a tunable parameter to adjust exploration vs. exploitation. We hypothesize that our method will significantly accelerate the runtime of Bayesian Optimization, and enable us to explore more points in the COFs dataset which cannot be done with a standard global GP model.

What to know more click here

References

[1] Mercado, R.; Fu, R.-S.; Yakutovich, A. V.; Talirz, L.; Haranczyk, M.; Smit, B. In Silico Design of 2D and 3D Covalent Organic Frameworks for Methane Storage Applications. Chem. Mater. 2018, 30 (15), 5069–5086. https://doi.org/10.1021/acs.chemmater.8b01425.

---

User guide

1. Install the conda environment bo-hackathon from the bo-hackathon.yml file. Type the following command in the terminal : conda env create -f bo-hackathon.yml

2. The inputs to the code must be provided in code_inputs.py

3. The jupyter notebook for running the code is located under src/BO.ipynb

4. The dataset for training the model is located under data/properties.csv

5. The results from training the model are located under bo_output

---