Automated screening and optimization of battery formation and cycling procedures

Abstract

A method for probing a multidimensional parameter space of a battery cell test protocol is provided, comprising defining a parameter space for a plurality of battery cells being tested, discretizing the parameter space, collecting a preliminary set of cells that are cycled to failure to sample strategies from the entire parameter space and including multiple repetitions of strategies, specifying resource hyperparameters, parameter space hyperparameters, and algorithm hyperparameters, selecting a random subset of charging strategies, testing the random subset of charging strategies until a number of cycles required for early prediction of battery life is reached, inputting the early predicted cycle data into an early prediction algorithm to obtain an early prediction, inputting the early prediction into an optimal experimental design (OED) algorithm to obtain a recommendation to run at least one next test, running the recommended test by repeating the random subset testing steps above, and validating the final recommended strategy.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of priority to U.S. Provisional Patent Application No. 62 / 573,561, filed on October 17, 2017, which is incorporated herein by reference. Technical Field

[0003] The present invention relates to battery optimization. More specifically, the present invention relates to a method for screening and optimizing within a multi-dimensional parameter space of battery cell formation and cycling protocols. Background Art

[0004] Battery optimization today is often performed by key tasks that include unoptimized empirical testing where test parameters are arbitrarily chosen. Unvalidated physiochemical models have been used that are inadequate for optimizing these processes. Limited implementations of data analysis have been attempted where only “summary” data are used, discarding important insights.

[0005] Optimizing battery protocols is a major bottleneck in the development and deployment of lithium-ion batteries. These protocols can include fast charging, where multi-step fast charging protocols can lead to longer cycle life and lower charging time; forming cycles (usually performed immediately after cell assembly in the factory), where complex forming protocols can reduce forming time (thereby reducing costs) while increasing cycle life and / or safety; and use conditions, where applications with flexible power consumption requirements can be optimized to maximize life.

[0006] Common challenges in these optimizations are high dimensionality and high manufacturing variations. Two approaches are currently being performed to explore these large parameter spaces: model optimization and grid search. Optimizing battery degradation models is an attractive approach due to its low cost (computational rather than physical); however, current battery models do not have sufficient complexity and accuracy to capture all relevant degradation modes as well as manufacturing variations. Grid search, which experimentally tests these setups multiple times, is accurate but expensive in terms of time, test equipment, and cells. Summary of the invention

[0007] To meet the needs in the art, a method for exploring a multidimensional parameter space and cycling protocol formed by a battery cell is provided, which includes defining a parameter space for a plurality of optimized battery cells, sampling strategies from the entire parameter space using a battery cycling instrument, and including one or more repetitions of at least one strategy, specifying hyperparameters, wherein the hyperparameters include resource hyperparameters, parameter space hyperparameters, and algorithm hyperparameters, selecting a subset of charging strategies, testing the subset of charging strategies until a desired accuracy is achieved, employing an optimal experimental design (OED) algorithm in a non-transitory computer readable medium to obtain a recommendation for running at least one next test, running the recommended test by repeating the random subset testing steps above, and validating the final recommended strategy.

[0008] According to one aspect of the invention, the parameter space includes the number of cycle steps, cycle time, state of charge (SOC) range, and minimum and maximum current, voltage, resistance and / or temperature boundaries for each cycle step.

[0009] In another aspect of the invention, the parameter space comprises a multi-step parameter space to optimize the formation cycling or charging rate within a series of defined ranges of SOC within a specified amount of time, wherein each step controls a percentage of each SOC range, wherein each SOC range is independent of the other SOC ranges, and wherein a final SOC range is the sum of all SOC ranges prior to the final SOC range.

[0010] In yet another aspect of the invention, resource hyperparameters include the number of available test channels and the number of batches.

[0011] In one aspect of the invention, the parameter space hyperparameters include the mean and standard deviation across the lifetime of all strategies, as well as the standard deviation of a single strategy over multiple tests.

[0012] In yet another aspect of the present invention, the algorithm hyperparameters include the degree of similarity between adjacent strategies in the parameter space, an exploration constant for controlling the balance between exploration and exploitation, and a decay constant of the exploitation constant in each round.

[0013] According to one aspect of the invention, a preliminary set of cells is configured to generate data to develop early predictive models, quantify means, standard deviations, and lifetime ranges over parameter space, and quantify intrinsic cell-to-cell variation of nominally identical cells cycled under nominally identical cycling conditions.

[0014] In one aspect, the invention also includes a multi-stage OED, wherein the multi-stage OED includes a first round and a second round of closed-loop testing, wherein the first round includes performing a preliminary classification of strategies into a low-life strategy group or a high-life strategy group, wherein quantitative prediction is not required.

[0015] In another aspect, the invention also includes dynamic early prediction, wherein the dynamic early prediction includes relaxing the size of the preliminary set of cells collected as more data is collected if the confidence of the prediction increases.

[0016] According to another aspect, the present invention further includes multi-cell sampling for each strategy within a test strategy round, wherein the multi-cell sampling targets one or more units of interest within the test strategy round. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of an autonomous closed-loop test procedure according to an embodiment of the present invention is shown.

[0018] Figure 2 A flow chart of an algorithm implemented using an implementation is shown. DETAILED DESCRIPTION

[0019] Lithium-ion batteries have a long time to market due to the extensive testing and optimization required for electrochemical protocols. The present invention provides a method to shorten the time to market for batteries, ultimately reducing costs. In one aspect, the present invention uses an adaptive optimal experimental design (OED) to accelerate the optimization of battery processes. According to one embodiment, the present invention includes a closed-loop test for battery optimization, combined with an "adaptive" optimal experimental design algorithm to quickly find the best battery program. The adaptive OED algorithm uses initial cycle data and the associated confidence interval of the prediction to combine the prediction of the battery life to determine whether to continue or stop the current test. If the test is stopped, it will determine the next strategy to be tested. When in use, the battery is cycled in a regulator, the life of each cell is estimated from the initial cycle results, and the adaptive OED algorithm intelligently decides when to stop the current test when the prediction has sufficient confidence, and the next strategy to be tested is automatically sent to the regulator.

[0020] Applications include any battery manufacturing step or use case that needs to be optimized. Some examples of processes to be optimized include fast charging, formation cycling, and state of health monitoring.

[0021] The present invention provides a completely data-driven optimization method, thereby avoiding these disadvantages. In one aspect, the method of automatically optimizing the battery process using adaptive optimal experimental design (adaptive OED) accelerates the battery development time by up to two orders of magnitude. The method of the present invention reduces the development cost and development time by several orders of magnitude. In addition, the present invention includes the intelligent use of early prediction and subsampling in the optimal experimental design.

[0022] Several variations are possible, such as incorporating additional data probes including electrochemical impedance spectroscopy (EIS) and strain monitoring, optimizing other conditions such as minimal swelling volume change, and complete closed-loop testing with integrated robotics for automated cell switching.

[0023] The present invention employs OED to efficiently probe the multidimensional parameter space of battery protocols. In this paper, efficiency is defined as low cost in terms of time and resources (test equipment and units). The process runs semi-automatically, meaning that after the algorithm is designed, only manual tasks (such as loading and unloading units) require human input. The parameter space can be discrete or continuous.

[0024] Turning now to the algorithm, a general method for optimizing a multi-step battery protocol is provided. Given a limited budget, the algorithm maximizes an objective function over a set of input parameters, specifically (but not limited to) the battery environment. For example, one objective includes finding a charging strategy that includes multiple current steps (input parameters) with a maximum lifetime value (objective function). The objective function may include a weighted combination of multiple objectives, such as charging time and lifetime.

[0025] Hoffman et al., “On correlation and budget constraints in model-based bandit optimization with application to automatic machine learning,” which is incorporated by reference in its entirety, describes in detail a Bayesian algorithm for fixed-budget optimization of an objective function with three unique design aspects. First, the correlation between input parameters allows one to infer the behavior of similar inputs without directly testing them, which reduces the number of tests required. In other words, the performance of an untested charging strategy can be inferred from the performance of the tested strategy. Second, the fixed budget setting limits the total number of tests required, which is a practical limitation often encountered in the real world. Third, the algorithm intelligently balances exploration and exploitation of parameter space. Exploration refers to learning more about previously unexplored areas of parameter space, while exploitation refers to learning more about promising areas of parameter space.

[0026] The present invention applies the teachings of Hoffman et al. in the context of battery optimization, while further incorporating new ideas to achieve this implementation. Specifically, the present invention incorporates two novel features into the Hoffman algorithm, which includes early prediction of battery life to further reduce optimization time. Herein, for a typical battery with a life of 3000 cycles, an accurate early prediction algorithm is provided that reduces the test time by 30 times using the first 100 cycles, although at the expense of additional sampling error (i.e., reduced prediction confidence). Secondly, the experiments are considered to be conducted in batches. Since battery testing equipment is typically designed for multiple tests in parallel, as well as practical issues such as cell loading / unloading processes, data management, and data quality issues, battery cycling is typically performed in parallel between many test "channels". Since multiple batches are typically required, the setup essentially runs in a "parallel-sequential" mode. Figure 1 A schematic diagram of this "closed loop" process is shown. First, a subset of batteries are tested. This data is fed into an early prediction algorithm, which in turn produces lifetime predictions. These predictions are then fed into an optimal experimental design algorithm, which makes recommendations on which strategies to test next. This process is completed until the budget is exhausted or the results are satisfactory.

[0027] Now turn to the algorithm implementation, such as Figure 2 As shown in the flowchart, in one embodiment, it is assumed that there is no prior knowledge of the parameter space, and in another embodiment, preliminary data is provided to the early prediction algorithm. Using non-transitory computer media, the algorithm first defines the parameter space, where the parameter space design includes the number of charging steps, the charging time, the state of charge (SOC) range, and the minimum and maximum current limits of each step. In an exemplary embodiment, a four-step parameter space is provided to optimize fast charging from 0% SOC to 80% SOC in 10 minutes. Each step controls a 20% SOC range, i.e., C1 ranges from 0% to 20% SOC, C2 from 20% to 40% SOC, and so on. With this configuration, the three parameters (C1, C2, and C3) are independent, while C4 is specified by a unique combination of C1, C2, and C3.

[0028] The next step is to discretize the parameter space. In one embodiment, the parameter space is sampled at 0.4C intervals for parameter values ​​below 6C and at 1C intervals for parameter values ​​above 1C. In summary, the parameter space in this example has 224 charging strategies.

[0029] A further step includes collecting a preliminary "batch" (data) of cells cycled to failure. This batch should sample strategies from across the parameter space and include multiple replicates of at least one strategy. This batch is used for several purposes, including generating data to develop early prediction models, quantifying the mean, standard deviation, and cycle life range over parameter space, and quantifying the inherent cell-to-cell variation of nominally identical cells that cycled with nominally identical cycling conditions.

[0030] Hyperparameters are then specified from here, where these hyperparameters fall into three categories, namely resource hyperparameters, parameter space hyperparameters, and algorithm hyperparameters. Resource hyperparameters include (a) the number of available test channels and (b) the number of batches (constrained by the available test time or number of units). Parameter space hyperparameters include (a) the mean and (b) standard deviation of the lifespan across all strategies, and (c) the standard deviation of a single strategy over multiple tests. These can be estimated from the preliminary batches in the above step. Algorithm hyperparameters include (a) gamma, (b) initial beta, and (c) epsilon. Gamma controls the degree of similarity between adjacent strategies in the parameter space. Beta represents the exploration constant, which controls the balance between exploration and exploitation. Epsilon represents the decay constant of beta per round; as the experiment progresses, epsilon shifts towards stronger exploitation (determined by β t =β0∈ t Given, where β t The three hyperparameters can be estimated by creating a simulator based on the range of lifetimes obtained in preliminary batches, testing all hyperparameter combinations on the simulator, and selecting the hyperparameter combination with the best performance (i.e., the one that most consistently obtains the true lifetime).

[0031] The invention also includes selecting a random subset of charging strategies without replacement. The size of this subset is determined by the number of available test channels. This is the first round of closed-loop testing.

[0032] The selected subset is then used to test the strategies until the number of cycles required for early prediction is achieved. In one implementation, the early prediction algorithm requires 100 cycles before returning a prediction.

[0033] The experimental data is then fed into the early prediction algorithm and predictions are obtained. The details of the early prediction method are beyond the scope of this work.

[0034] Next, the predictions are fed into the OED algorithm to get recommendations for the tests to run next. Hoffman describes in detail the Bayesian correlation between neighboring strategies and the exploration versus exploitation decision.

[0035] The proposed test is then run, where the above process is repeated until the number of cycles required for early prediction is achieved, where the repeated process includes testing the strategy, inputting data into the early prediction algorithm, and inputting the predictions into the OED algorithm.

[0036] The final step is to validate the suggested strategies, which involves confirming that the recommendations from the OED are satisfactory by testing multiple units using these strategies to failure.

[0037] Many variations of the OED algorithm are possible, including multi-stage OED, dynamic early prediction, and multi-unit sampling for each strategy within the same round. This paper defines multi-stage OED by considering two rounds of closed-loop testing. The first is to preliminarily classify the strategies into low-life group or high-life group. Since quantitative prediction is not required, the prediction algorithm can use a minimum cycle (i.e., 5), saving more time. The second stage will be performed as described above. Dynamic early prediction is defined as the increase in the accuracy of the early prediction algorithm as more data is collected. If the confidence of the early prediction exceeds a certain threshold, the "batch" setting can be relaxed; if the confidence of the prediction is high, the unit can be terminated early, saving additional time. Finally, multi-unit sampling for each strategy in the same round is implemented in the current setting, where at most one unit is tested per strategy per round. The algorithm can be modified to repeatedly test promising points within the same round.

[0038] The present invention has now been described in terms of several exemplary embodiments, which are intended in all respects to be illustrative and not restrictive. Thus, the present invention is capable of many variations in the detailed implementations that may be derived by one skilled in the art from the description contained herein. For example, one embodiment uses continuous OED, and another embodiment uses OED without any prior prediction. All such variations are considered to be within the scope and spirit of the present invention, as defined by the appended claims and their legal equivalents.

Claims

1. A method for optimizing a charging strategy for battery cell formation or cycling, the method comprising: a) defining a parameter space for a plurality of battery cells to be optimized; b) specifying hyperparameters, wherein the hyperparameters include resource hyperparameters, parameter space hyperparameters, and algorithm hyperparameters; c) determining a test sequence of a subset of the charging strategies, wherein each of the charging strategies includes a corresponding charging rate and the test sequence allows for repetition of the charging strategies; d) testing the subset of the charging strategies on the optimized plurality of battery cells using a battery cycling instrument until a number of cycles required for accurate life prediction is achieved; e) using an optimal experimental design (OED) algorithm to obtain a recommendation for running at least one next test; f) running the proposed test by repeating c)-e) above one or more times using closed-loop laboratory testing to iteratively find a lifetime optimized charging strategy according to the hyperparameters; and g) providing said lifetime optimized charging strategy as an output.

2. The method according to claim 1, wherein: The parameter space includes the number of cycle steps, cycle time, state of charge range, and minimum and maximum current, voltage, resistance and / or temperature boundaries for each of the cycle steps.

3. The method according to claim 2, wherein: The parameter space comprises a multi-step parameter space to optimize forming a cycling or charging rate within a series of defined ranges of the state of charge within a specified amount of time, wherein each of the steps controls a percentage of each of the state of charge ranges, wherein each of the state of charge ranges is independent of other state of charge ranges, and wherein a final state of charge range is a sum of all of the state of charge ranges preceding the final state of charge range.

4. The method according to claim 1, wherein: The resource hyperparameters include the number of available test channels and the number of batches.

5. The method according to claim 1, wherein: The parameter space hyperparameters include the mean and standard deviation across the lifetime of all the policies, as well as the standard deviation of a single policy over multiple tests.

6. The method according to claim 1, wherein: The algorithm hyperparameters include the similarity between adjacent strategies in the parameter space, an exploration constant for controlling the balance between exploration and exploitation, and a decay constant of the exploitation constant in each round.

7. The method according to claim 1, wherein: The preliminary cell groups are configured to generate data to develop early predictive models, quantify the mean, standard deviation, and lifetime range over the parameter space, and quantify the intrinsic cell-to-cell variation of nominally identical cells cycled under nominally identical cycling conditions.

8. The method according to claim 1 further comprises multi-stage said OED, wherein said multi-stage OED comprises a first round and a second round of closed-loop testing, wherein said first round comprises performing a preliminary classification of strategies into a low-life strategy group or a high-life strategy group, wherein no quantitative prediction is required, and wherein said second round comprises implementing the method of claim 1.

9. The method of claim 1 further comprising dynamic early prediction, wherein the dynamic early prediction comprises relaxing the size of the collected preliminary cell set as more data is collected if confidence in the prediction increases.

10. The method of claim 1, further comprising multi-unit sampling for each strategy within a test strategy round, wherein the multi-unit sampling targets one or more units of interest within the test strategy round.

Images