Electrolyte recommender system for batteries using ensemble Bayesian optimization
Introduction
Bayesian optimization (BO) is a powerful optimization technique that can be used for unknown target functions and is robust against noisy observations (Frazier and Wang, 2016, Močkus, 1974, Shahriari et al., 2016). Hence, BO has been applied to a wide variety of applications, including optimizing the hyper-parameter settings of deep neural networks (Falkner et al., 2018, Swersky et al., 2014), robot control systems (Antonova et al., 2017, Cornejo Bueno et al., 2018, Marco et al., 2017), material designs (Ju et al., 2017, Todorović et al., 2017), and drug discovery (Negoescu et al., 2011). Recently, BO has been also applied to applications related to not only simulations but also real experimental systems in materials science (Dave et al., 2020, Li et al., 2017, Schweidtmann et al., 2018). Furthermore, although there are commercial products such as automated liquid distribution systems, which can apply to electrolytes in batteries, optimization algorithms (such as BO) are not used in these systems.
In BO, Gaussian process (GP) regression is often used to derive an expressive, non-parametric surrogate function that represents the unknown target function but is more flexible and tractable (Snoek et al., 2012). This approach has been utilized in the exploration–exploitation trade-off of BO (Schulz et al., 2018). However, according to Schulz et al. (2016), a mismatch between the smoothness properties of the kernel function and the unknown target function can reduce the convergence speed of BO. In addition, the convergence speed in BO differs depending on the type of acquisition function (Hennig and Schuler, 2012, Srinivas et al., 2010). Thus, a suitable acquisition function must be selected depending on the unknown target functions. To resolve such issues, Hoffman et al. (2011) proposed a method of portfolio allocation of acquisition functions, named GP-Hedge, and demonstrated its effectiveness compared to the best acquisition function. This method can be used as a BO as a part of an ensemble, but further hyper-parameter tuning of GP-Hedge is needed. For BO, hyper-parameter tuning of the acquisition function is needed, which increases experimental costs.
To address these issues, we previously developed an ensemble BO (EBO) (Ohno, 2018), consisting of several BO algorithms (hereafter referred to as BOs) with different kernel functions, acquisition functions, and associated hyper-parameters. A single common database is used for all kernel functions within an EBO. Hence, an empirical Gram matrix of the GP for each BO in the ensemble is based on the common database. However, in the absence of a priori knowledge about the target functions, suitable kernel and acquisition functions cannot be selected. In such cases, EBO can be conducted using an ensemble of several BOs with different kernel and acquisition functions to match the characteristics of the target functions as closely as possible. In addition, for practical experiments, iterations of experiments are often restricted due to limitations such as tank capacity and the cost of materials. Here, the characteristics of EBO are summarized as follows:
- 1.
Easy to use practically due to its simple structure.
- 2.
Enlarges the search space in combination with different search strategies, which can improve the training of surrogate functions.
- 3.
Reduces the computational costs associated with hyper-parameter tuning due to the use of acquisition functions with distinct parameter values.
In 2, the search strategy is determined by the combination of acquisition functions and kernel functions. Differences in the performance of results, due to different search strategies, can enlarge the search space. In addition, low-performance results are actively gathered for the purpose of training BOs and for use in analyzing target phenomena of interest. This characteristic is not seen in previous work such as Hoffman et al. (2011). Characteristic 3 can be achieved by setting different hyper-parameter values to each acquisition function. This is available for the limited iterations for the experiments. Recently, Wang, Clement et al. (2018) proposed an EBO to tackle high-dimensional and large-scale parameter-search problems by using a Mondrian process to partition the search space and assign a different BO in the ensemble in order to search each partitioned space. However, this approach did not resolve the above-mentioned issues. Furthermore, although the partitioning of the search space may mitigate the mismatch in smoothness locally, it does not sufficiently deal with the mismatch issue.
As a feasibility study of EBO, we develop a recommender system for electrolytes, in which an EBO optimizes the electrolytes used in Li-ion batteries, which is challenging due to the large number of possible combinations of solution ratios for the electrolytes as well as the noise in the observations. The proposed recommender system is designed to identify the solution (solvents and salt) compositions that maximize ionic conductivity, and to reduce costs associated with hyper-parameter tuning and the selection of the acquisition functions and kernel functions. Thus, in practical situations, scientists who may not have expert knowledge of optimization techniques can easily set up and utilize the proposed system. Experimental optimizations were conducted for four different solutions and the results obtained demonstrated the effectiveness of the EBO. Moreover, the experimental results provided additional insight in the form of a useful mixing rule for maximizing the ionic conductivity with the four solutions used here. This rule states that a small amount of a solvent with a low dielectric constant and high viscosity should be included in the mixture of solutions in order to attain high ionic conductivity. Furthermore, the behavior of each BO in the ensemble was investigated after carrying out the optimization process in the system, and it was revealed that the convergence to the maximum ionic conductivity differed depending on the type of kernel and acquisition functions used. This suggests that the proposed EBO system can be generalized to other optimization problems.
The remainder of this paper is organized as follows. Prior related work is described in Section 2. In Section 3, we describe the proposed EBO algorithm in detail. An overview of the electrolyte recommender system is provided in Section 4 along with the experimental results and an analysis of the findings. In Section 5, we discuss the limitations of the present method and obtained results. Finally, a summary and conclusion, as well as future work, are presented in Section 6.
Section snippets
Related work
The electrolyte recommender system, as described in the following sections, can be viewed as a closed-loop experimental system (Houben & Lapkin, 2015). The optimization of this system involves parallel experiments. Thus, in this section, we review BO algorithms that are suitable for parallel experiments and closed-loop experimental systems in materials science.
EBO
Fig. 1 shows an EBO framework. This framework is useful for optimization problems in cases when there is little knowledge about the target function because it would be effective to use various search strategies. Here, the search strategy is followed by the specifications represented by a 3-tuple: type of acquisition function, hyper-parameter value of the acquisition function, and type of kernel function. Each BO (i.e., worker) searches for different spaces according to the different search
Development of a system to recommend electrolytes for batteries
Fig. 2 shows a schematic diagram of the electrolyte recommender system that is designed to determine an optimal mixing ratio for the four solutions to maximize the ionic conductivity of the electrolyte. Here, the solutions contain ions and exhibit electrically conductive properties. In this system, the electrolyte is a mixture of four or fewer solutions. Five electrochemical cells were used to conduct parallel experiments. For each electrochemical cell, four pumps were required to inject the
Discussion
Theoretical analysis: Proposition 1 was presented as a theoretical foundation for the effectiveness of EBO. The proposition states that the effect (positive or negative) of an ensemble depends on the number of observations. In the previous studies described in Section 2, the studies using batch modes of BO and EBO (Contal et al., 2013, Desautels et al., 2014, Wang, Clement et al., 2018) did not address such effects of the ensemble. However, as described in Section 3.4, the assumptions in
Conclusion
In this study, we presented an algorithm for EBO and as a feasibility study of EBO, developed a system to recommend a battery electrolyte (the optimal composition ratio for four solutions) with maximized ionic conductivity. The EBO comprises multiple BO algorithms with different kernel and acquisition functions. An aim of EBO is to make search space large, including not only high performance results but also low performance results. The experimental results confirmed that the EBO significantly
CRediT authorship contribution statement
Hiroshi Ohno: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Writing - original draft, Visualization. Akitoshi Suzumura: Software, Validation, Resources, Writing - reviewing & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
The authors sincerely thank all of the anonymous reviewers for their useful comments and suggestions to improve the manuscript, Dr. Kenshuke Takechi and Dr. Kaito Miyamoto for helpful discussions, and Dr. Shin Tajima and Ms. Yumi Masuoka for their help with the experimental preparation.
References (48)
- et al.
Bayesian Optimization of a hybrid system for robust ocean wave features prediction
Neurocomputing
(2018) - et al.
Autonomous discovery of battery electrolytes with robotic experimentation and machine learning
Cell Reports Physical Science
(2020) - et al.
Automatic discovery and optimization of chemical processes
Current Opinion in Chemical Engineering
(2015) - et al.
Progress in electrolytes for rechargeable li-based batteries and beyond
Green Energy & Environment
(2016) Empirical studies of Gaussian process based Bayesian optimization using evolutionary computation for materials informatics
Expert Systems with Applications
(2018)- et al.
A tutorial on Gaussian process regression: Modelling, exploring, and exploiting functions
Journal of Mathematical Psychology
(2018) - et al.
Machine learning meets continuous flow chemistry: Automated optimization towards the Pareto front of multiple objectives
Chemical Engineering Journal
(2018) - et al.
Review on modeling of the anode solid electrolyte interphase (SEI) for lithium-ion batteries
NPJ Computational Materials
(2018) - et al.
Deep kernels for optimizing locomotion controllers
- et al.
Closed-loop optimization of fast-charging protocols for batteries with machine learning
Nature
(2020)
Using confidence bounds for exploitation-exploration trade-offs
Journal of Machine Learning Research
Batch Bayesian optimization via simulation matching
Exploration enhanced expected improvement for Bayesian optimization
Efficient multiobjective optimization employing Gaussian processes, spectral sampling and a genetic algorithm
Journal of Global Optimization
A tutorial on Bayesian optimization of expensive cost functions, with application to active user modeling and hierarchical reinforcement learning
Networking chemical robots for reaction multitasking
Nature Communications
Parallel Gaussian process optimization with upper confidence bound and pure exploration
SDO: A statistical method for global optimization
Parallelizing exploration-exploitation tradeoffs in Gaussian process bandit optimization
Journal of Machine Learning Research
Bayesian Optimization for materials design
Multi-objective optimization for materials discovery via adaptive design
Scientific Reports
Controlling an organic synthesis robot with machine learning to search for new reactivity
Nature
A kernel two-sample test
Journal of Machine Learning Research
Cited by (8)
Influence of imbibition temperature on the extraction stage of the cane sugar production
2023, Journal of Food EngineeringFinding a novel electrolyte solution of lithium-ion batteries using an autonomous search system based on ensemble optimization
2022, Journal of Power SourcesCitation Excerpt :A set of BOs in parallel were utilized with various types of kernel and acquisition functions, including their hyper-parameters [26] and the combination of kernel and acquisition functions were set to three by three, as listed in Table 1. The squared exponential kernel (SE), the 0th and the 2nd Matern kernels (MAT0 and MAT2, respectively) were set as kernel functions whilst the Gaussian process upper confidence bound (GP-UCB) and the expected improvement (EI) were set as acquisition functions [25]. To obtain the composition with the maximum observed physical property, our automated experiment was performed according to Algorithms (Algorithm 1 and 2) in the supplementary material.
How to build an effective self-driving laboratory
2023, MRS BulletinA Novel Optimized Tourism Itinerary Recommender System: A Modified Capacitated Vehicle Routing Problem Approach
2023, 2023 IEEE International Conference on Industrial Engineering and Engineering Management, IEEM 2023