Discovery of Energy Storage Molecular Materials Using Quantum Chemistry-Guided Multiobjective Bayesian Optimization

Abstract

Redox flow batteries (RFBs) are a promising technology for stationary energy storage applications due to their flexible design, scalability, and low cost. In RFBs, energy is carried in flowable redox-active materials (redoxmers) which are stored externally and pumped to the cell during operation. Further improvements in the energy density of RFBs necessitates redoxmer designs with wider redox potential windows and higher solubility. Additionally, designing redoxmers with a fluorescence-enabled self-reporting functionality allows monitoring of the state of health of RFBs. To accelerate the discovery of redoxmers with desired properties, state-of-the-art machine learning (ML) methods, such as multiobjective Bayesian optimization (MBO), are useful. Here, we first employed density functional theory calculations to generate a database of reduction potentials, solvation free energies, and absorption wavelengths for 1400 redoxmer molecules based on a 2,1,3-benzothiadiazole (BzNSN) core structure. From the computed properties, we identified 22 Pareto-optimal molecules that represent best trade-off among all of the desired properties. We further utilized these data to develop and benchmark an MBO approach to identify candidates quickly and efficiently with multiple targeted properties. With MBO, optimal candidates from the 1400-molecule data set can be identified at least 15 times more efficiently compared to the brute force or random selection approach. Importantly, we utilized this approach for discovering promising redoxmers from an unseen database of 1 million BzNSN-based molecules, where we discovered 16 new Pareto-optimal molecules with significant improvements in properties over the initial 1400 molecules. We anticipate that this active learning technique is general and can be utilized for the discovery of any class of functional materials that satisfies multiple desired property criteria.

Bayesian optimization with known experimental and design constraints for chemistry applications

“…5a) for which the reduction potential E red , solvation free energy G solv , and maximum absorption wavelength λ abs were computed with DFT. 69 This example application thus simulates a DFT-based computational screen 70–72 that aims to identify redoxmer candidates with self-reporting features and a high probability of synthetic accessibility.…”

Bayesian optimization with known experimental and design constraints for chemistry applications

“…5a) for which the reduction potential E red , solvation free energy G solv , and maximum absorption wavelength λ abs were computed with DFT. 68 This example application thus simulates a DFT-based computational screen [69][70][71] that aims to identify redoxmer candidates with self-reporting features and a high probability of synthetic accessibility.…”

The most compact search space is not always the most efficient: A case study on maximizing solid rocket fuel packing fraction via constrained Bayesian optimization

“…• Could regularization terms or optimization algorithms geared towards ill-posed inverse problems mitigate degeneracy issues? 13 For Negative Integrated Posterior Variance acquisition function usage in Ax, see https://github.com/facebook/ Ax/issues/930…”

“…BO has been used to create and adaptively refine surrogate models for physics-based simulations whether acting directly as the surrogate model [2,4,5,9,[13][14][15][16][17][18][19][20][21] or tuning hyperparameters of a surrogate model [3,22] among other applications such as experimental discovery [7,9,11] and crystal structure prediction [23][24][25][26]. A review of Bayesian optimization applied to materials science in general is given in Kotthoff et al [27].…”