Systems and methods for battery smart sensing using a virtual reference electrode

Abstract

Systems and methods directed to a battery management system (BMS) for controlling charging of a battery are provided. Parameters corresponding to a state of the battery are received and provided to a trained machine learning model. The machine learning model generates a virtual reference electrode. A predicted anode voltage of the battery is determined using the virtual reference electrode. The system for charging the battery is controlled via the charging protocol using the predicted anode voltage of the battery.

Description

2024-209-2 / 790482.00539SYSTEMS AND METHODS FOR BATTERY SMART SENSING USING A VIRTUAL REFERENCE ELECTRODECROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. Application No. 63 / 684,679, titled “SYSTEMS AND METHODS FOR BATTERY SMART SENSING USING A VIRTUAL REFERENCE ELECTRODE,” filed August 19, 2024, which is hereby incorporated by reference in its entirety.STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH

[0002] N / ABACKGROUND

[0003] Charging protocols rely on battery management systems (BMS) that translate measurable parameters (e.g., cell voltage, current) into estimates of “safe” or “unsafe” operation, adjusting the control parameters as necessary. However, existing BMS are largely blind to catastrophic failure modes (e.g. metallic Li plating). For example, standard measurables give limited physical insight into battery degradation mechanisms. Thus, conventional charging protocols are highly empirical and rigid and have no way to account for catastrophic failure modes, such as may be caused by metallic Li plating, or the like. For example, while cell voltage is a common measurable related to battery health and state of charge, it does not provide physical insight into when or if metallic Li is plating on the anode, making it difficult for a BMS to define battery safety limits based on cell voltage. These limitations restrict current BMS to either overly conservative modes of battery operation or more aggressive protocols at the risk of failure modes going undetected.

[0004] Thus, there is an ongoing need for systems and methods for improved battery management and control.SUMMARY OF THE DISCLOSURE

[0005] The present disclosure provides systems and methods that overcome the aforementioned drawbacks. In particular systems and methods are provided for controlling or managing batter operation using a virtual reference electrode or other virtual or similar data andQB\98068473.1 12024-209-2 / 790482.00539 control. As will be described, the systems and methods provided herein yield improved overall operation, reduced need for conservative control for failure avoidance, and reduced potential for failure, as compared to traditional BMS.

[0006] In accordance with once aspect of the disclosure, a battery management system (BMS) for controlling charging of a battery can include a battery, a processor, and a memory. The processor can be configured to control systems for charging the battery. The memory can have instructions stored thereon that, when executed, can cause the processor to receive a plurality of parameters corresponding to a state of the battery. The plurality of parameters can be provided to a trained machine learning model. The machine learning model can generate a virtual reference electrode. A predicted anode voltage of the battery can be determined using the virtual reference electrode. The system can be controlled for charging the battery using the predicted anode voltage of the battery.

[0007] In accordance with another aspect of the present disclosure, a computer readable storage medium having instructions stored thereon is provided. When the instructions are executed by a processor, they can cause the processor to carry out steps that can include receiving a plurality of parameters corresponding to a state of a battery. The plurality of parameters can be provided to a trained machine learning model. A virtual reference electrode can be generated via the machine learning model. A predicted anode voltage of the battery can be determined using the virtual reference electrode. Charging of the battery can be controlled using the predicted anode voltage of the battery.

[0008] The foregoing and other aspects and advantages of the present disclosure will appear from the following description. In the description, reference is made to the accompanying drawings that form a part hereof, and in which there is shown by way of illustration one or more exemplary versions. These versions do not necessarily represent the full scope of the disclosure.BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The following drawings are provided to help illustrate various features of non-limiting examples of the disclosure, and are not intended to limit the scope of the disclosure or exclude alternative implementations.

[0010] FIG. 1 schematically illustrates a battery management system for virtual anode voltage prediction in accordance with some configurations.QB\98068473.1 22024-209-2 / 790482.00539

[0011] FIG. 2 is a flow diagram illustrating an example process for controlling a charging protocol of a battery based on a predicted anode voltage in accordance with some configurations.

[0012] FIG. 3 is a flow diagram illustrating an example process for machine learning model training for anode voltage prediction in accordance with some configurations.

[0013] FIG. 4A illustrates an example plot of data collected from conventional 2-electrode cells in accordance with some configurations.

[0014] FIG. 4B illustrates example plots of lithiation modes detected using a 3 -electrode geometry in accordance with some configurations.

[0015] FIG. 4C illustrates an example workflow for building a virtual reference electrode illustrate an example feature selection workflow and demonstration of performance for anode voltage prediction in accordance with some configurations.

[0016] FIGS. 5A illustrates an example feature selection workflow and demonstration of a performance for an anode voltage prediction model in accordance with some configurations.

[0017] FIG. 5B illustrates an example plot of an overall performance associated with a training and testing of a neural network model in accordance with some configurations.

[0018] FIG. 5C illustrates example plots of a multi-cycle performance in accordance with some configurations.

[0019] FIG. 6A illustrates an example plot associated with experimental validation results of an Al-powered virtual reference electrode in accordance with some configurations.

[0020] FIG. 6B illustrates an example plot associated with experimental validation results of an Al-powered virtual reference electrode in accordance with some configurations.

[0021] FIG. 6C illustrates an example plot associated with experimental validation results of an Al-powered virtual reference electrode in accordance with some configurations.

[0022] FIG. 7A illustrates an example plot of cell voltage based-control data in accordance with some configurations.

[0023] FIG. 7B illustrates an example plot of anode voltage-based control data in accordance with some configurations.

[0024] FIG. 7C illustrates an example plot of capacity retentions in accordance with some configurations.

[0025] FIG. 7D illustrates an example plot of anode voltage versus a state of charge in accordance with some configurations.QB\98068473.1 32024-209-2 / 790482.00539

[0026] FIG. 7E illustrates an example graphite anode surface in accordance with some configurations.

[0027] FIG. 7F illustrates an example graphite anode surface in accordance with some configurations.DETAILED DESCRIPTION OF THE PRESENT DISCLOSURE

[0028] Provided herein are systems and methods which may be used to predict an anode voltage of a battery using a virtual reference electrode. The systems and methods provided herein will be described with reference to the figures, forming an instrumental yet non-limiting reduction to practice of the specification.

[0029] The detailed description set forth below in connection with the appended drawings is intended as a description of various configurations and is not intended to represent the only configurations in which the subject matter described herein may be practiced. The detailed description includes specific details to provide a thorough understanding of various embodiments of the present disclosure. However, it will be apparent to those skilled in the art that the various features, concepts and embodiments described herein may be implemented and practiced without these specific details. In some instances, well-known structures and components are shown in block diagram form to avoid obscuring such concepts.

[0030] Charging protocols rely on battery management systems (BMS) that translate measurable parameters (e g., cell voltage, current) into estimates of “safe” or “unsafe” operation, adjusting the control parameters as necessary. However, some BMS can be blind to failure modes (e.g. metallic Li plating), as standard measurables may give limited physical insight into battery degradation mechanisms, thus making conventional charging protocols empirical and rigid. For example, while cell voltage is a common measurable related to battery health and state of charge, it does not provide physical insight into when or if metallic Li is plating on the anode, making it difficult for a BMS to define battery safety limits based on cell voltage. These limitations restrict current BMS to either overly conservative modes of battery operation or more aggressive protocols at the risk of failure modes going undetected.

[0031] A smart BMS that can intelligently optimize battery operation may utilize reliable sensing of parameters that directly relate to the physics of battery failure. For example, undesired metallic Li plating at the graphite anode (a dominant material in some Li-ion batteries) isQB\98068473.1 42024-209-2 / 790482.00539 thermodynamically favored below an anode voltage (AV) of 0 V vs. Li metal. To avoid this failure mode that may lead to battery fire or even explosion, the AV can serve as a physically interpretable parameter that precisely defines safety limits of normal operation (i.e., Li intercalation above 0 V vs. Li metal) and catastrophic failure (i.e., Li metal plating below 0 V vs. Li metal). In conventional BMS, the AV may only be measured using a 3rdreference electrode, which can increase battery manufacturing complexity. Other physically interpretable parameters that detect Li plating (e.g., pressure, volume & thickness) may face a similar challenge. While simulating AV profiles using physics-based models is possible, accurately capturing the complex battery degradation modes coupled to chemical, thermal, and mechanical heterogeneities that evolve across a battery’s lifetime remains computationally and fundamentally challenging. For example, the accuracy of such physics-based approaches to predict AV across battery lifetime is rarely reported or validated beyond the 5thcycle, further highlighting the difficulty in predicting these parameters as batteries age. Therefore, the systems and method described herein address the need to accurately predict the AV continuously throughout battery operation without direct measurement.Hardware Configurations and Embodiments

[0032] Referring to FIG. 1, an example battery management system 100 that can be used to provide anode voltage prediction and battery charging control is illustrated. The server 120 may include additional, different, or fewer components than those illustrated in FIG. 1 in various configurations. The server 120 may perform additional or different functionality than the functionality described herein. Also, the functionality (or a portion thereof) described herein as being performed by the server 120 may be performed by another component, distributed among multiple devices (e.g., as part of a cloud service or cloud-computing environment), combined with another component (e.g., another component of the system 100), or a combination thereof.

[0033] In the battery management system 100, a battery 105 is monitored using one or more inputs, such as a sensor assembly 110, each of which transmits a signal over a cable 112 or other communication link or medium to a server 120. Moreover, charging protocols may be applied between the battery 105 and the memory 135 using a cable 112 or other communication link using the methods described herein. The server 120 includes an electronic processor 125, a communication interface 130, and memory 135. The electronic processor(s) 125, the memory 135,QB\98068473.1 52024-209-2 / 790482.00539 and the communication interface 130 may communicate wirelessly, over one or more communication lines or buses, or a combination thereof.

[0034] In some examples, the sensors 110 can include two electrodes, such as a working electrode and a counter electrode, or the like. The sensors 110 can generate respective signals by measuring parameters of the battery 105 such as an overall cell voltage, a capacity, or the like. The signals are then processed by one or more electronic processors 125. In some embodiments, the one or more electronic processors 125 can be any suitable hardware processor or combination of processors, such as a central processing unit (CPU), a graphics processing unit (GPU), and so on. For clarity, a single block is used to illustrate the sensors 110 shown in FIG. 1. It should be understood that the sensors 110 shown are intended to represent one or more sensors and adapted to receive signals from the battery 105. Various combinations of numbers and types of sensors, as mentioned, are suitable for use with the battery management system 100.

[0035] In some embodiments of the system shown in FIG. 1, all of the hardware used to receive and process signals from the sensors 110 are housed within the same housing. In other embodiments, some of the hardware used to receive and process signals is housed within a separate housing. In addition, the server 120 of certain embodiments includes hardware, software, or both hardware and software, whether in one housing or multiple housings, used to receive and process the signals transmitted by the sensors 110.

[0036] In some examples, the communication interface 130 may include a transceiver that communicates with the battery 105 or connected computing device (e.g., a user device) over a communication network and, optionally, one or more other communication networks or connections. The memory 135 can include a non-transitory, computer-readable storage medium. In some examples, the electronic processors 125 can be is configured to retrieve instructions and data from the memory 135 and execute the instructions (e.g., to implement the functionality of the electronic processors 125 described herein). In some examples, the server 120 and the electronic processors 125, as illustrated, represent a plurality of servers (and a plurality of processors) implementing the functionality of the server 120 described herein (e.g., providing the functions as a cloud-based system or service). Accordingly, although functions may be described as performed by the electronic processors 125, the electronic processors 125 may include multiple processors, whether co-located or geographically distributed, that perform one or more of these functions.QB\98068473.1 62024-209-2 / 790482.00539

[0037] As illustrated in FIG. 1, the memory 135 may store one or more machine learning models 140 and a virtual electrode module 145. Each machine learning model of the machine learning model(s) 140 may be using the method described below with respect to FIG. 3. In some examples, one or more of the machine learning models 140 may include one or more trained neural networks with multiple layers. In some examples, one of more of the machine learning models 140 may include one or more trained, multi-layer transformer neural networks (e.g., trained neural networks having multiple layers and a transformer architecture). In some examples, the one or more machine learning models 140 may be used to generate the virtual electrode module 145. The virtual electrode module 145 may include instructions that are retrieved and executed by the electronic processor(s) 125 to perform the functionality of the virtual electrode module 145. Accordingly, when the virtual electrode module 145 is described herein as performing functions, this description may be a shorthand manner of describing the functions of the electronic processor(s) 125 executing the instructions of the virtual electrode module 145. Accordingly, these functions of the virtual electrode module 145 may also be described as being functions of the electronic processor(s) 125 (and / or one or more connected user devices).

[0038] The virtual electrode module 145 may obtain data, such as battery state parameters, from the battery 105, sensors 110, the electronic processor(s) 125, the communication interface 130, or a combination thereof. In some examples, the analysis performed by the virtual electrode module 145 may cause the battery management system 100 to dynamically regulate one or more charging parameters of the battery 105. For example, a current output by a charger connected to the battery 105 may be adjusted based on an anode voltage predicted by the virtual electrode module 145 and / or the machine learning model(s) 140.

[0039] The memory 135 may include additional, different, or fewer components in different configurations. Alternatively, or in addition, in some configurations, one or more components of the memory 135 may be combined into a single component, distributed among multiple components, or the like. Alternatively, or in addition, in some configurations, one or more components of the memory 135 may be stored remotely from the server 120, or, in a remote database, another server, a remote user device, an external storage device, or the like.Example ProcessesQB\98068473.1 72024-209-2 / 790482.00539

[0040] Turning to FIG. 2, a process 200 for controlling a charging state of a battery based on a predicted anode voltage in accordance with the present disclosure, is illustrated. As described below, a particular implementation can omit some or all illustrated features / steps, may be implemented in some embodiments in a different order, and may not require some illustrated features to be implemented in all embodiments. In some examples, an apparatus (e.g., computing device, processor with memory, a battery management system, etc.) can be used to perform example process 200. However, it should be appreciated that any suitable apparatus or system for carrying out the operations or features described below may perform process 200.

[0041] In some aspects, the process 200 may begin at process block 205, where battery state parameters are received. In some examples, the battery state parameters may be measurements directly measured from a battery, such as parameters measured from battery 105 using sensors 110, as described above with respect to FIG. 1. For example, the battery state parameters may include a cell voltage, a cell current, a cell capacity, or the like. In some examples, the parameters measured from the battery may be obtained using a 2-electrode battery cell system in connection with FIG. 1, which does not include a third, reference electrode.

[0042] At process block 210, the battery state parameters are provided to a trained machine learning model, such as the machine learning model(s) 140 in connection with FIG. 1. In some examples, the machine learning model was trained in connection with processor 300 of FIG. 3. For example, the machine learning model may include one or more types of machine learning models, such as a feedforward neural network (FNN), a convolutional neural network (CNN), a recurrent neural network (RNN), or the like.

[0043] At process block 215, a virtual reference electrode is generated via the trained machine learning model. In some examples, the virtual reference electrode may be saved as the virtual electrode module 145, as described above with respect to FIG. 1. The virtual reference electrode may be configured to use only measurements obtained using a 2-electrode sensor system (i.e., sensors 110) coupled to the battery and output a predicted anode voltage for a corresponding battery cell. In some examples, during model implementation, the sensors 110 of the battery 105 do not include a third, reference electrode. Therefore, the virtual reference electrode may not require measurements from a third, reference electrode (in a 3 -electrode battery cell system) to output the predicted anode voltage as soon as the machine learning model(s) 140 are trained. InQB\98068473.1 82024-209-2 / 790482.00539 some examples, the virtual reference electrode may be generated for a specific battery, a type of battery, or be applicable to any battery.

[0044] At process block 220, a predicted anode voltage of the battery is determined using the virtual reference electrode. In some examples, the predicted anode voltage may be a continuous output during a charging process. For example, the trained machine learning model or the virtual reference electrode may predict an anode voltage of the battery throughout a charging process of the battery when a battery is connected to a charger or charging station. In some examples, the predicted anode voltage may be determined using a plot that compares the predicted anode voltage to a cell capacity of the battery, or any other parameter received at process block 205.

[0045] At process block 225, a battery management system controls a charging protocol of the battery based on the predicted anode voltage. In some examples, the battery management system (such as battery management system 100 described above with respect to FIG. 1) may execute and change a smart charging protocol based on the predicted anode voltage. For example, the protocol may change a current or voltage output of a connected charging to maintain a predicted anode voltage, prevent Lithium metal plating, mitigate capacity degradation associated with lithium metal plating, prevent short circuits, prolong cycling stability, or the like.

[0046] FIG. 3 is a flow diagram illustrating an example process 300 for training a machine learning model to predict anode voltages, in accordance with some aspects of the present disclosure. As described below, a particular implementation can omit some or all illustrated features / steps, may be implemented in some embodiments in a different order, and may not require some illustrated features to be implemented in all embodiments. In some examples, an apparatus (e.g., computing device, processor with memory, a battery management system, etc.) can be used to perform example process 300. However, it should be appreciated that any suitable apparatus or system for carrying out the operations or features described below may perform process 300.

[0047] In some aspects, the process 300 may begin at process block 305, where 3-electrode cell data is obtained from a plurality of battery cells. In some examples, the 3-electrode cell data can include data obtained from one or more batteries of different types and during varying charging conditions in order to mimic multiple battery operation conditions. For example, the 3-electrode cell data may include data collected during various charging rates such as a charging under constant current conditions and / or constant voltage conditions, as well as varying discharging stepsQB\98068473.1 92024-209-2 / 790482.00539 and / or stepped charging conditions. Moreover, in some examples, the 3-electrode cell data may be obtained from multiple batteries with varying battery current densities.

[0048] At process block 310, a plurality of features is extracted from the 3-electrode cell data. In some examples, the plurality of features can include a cell voltage, a charging current, a battery cell capacity, corresponding time stamps, a power, an overpotential, a cell throughput, or the like. In some examples, a processor (e.g., the electronic processor(s) 125) may further process the data to determine engineered features, such as derivatives of one or more features.

[0049] At process block 315, anode voltages corresponding to the plurality of features are obtained, and feature selection may be applied to access features correlated with anode voltage(s). In some examples, the anode voltages may be obtained concurrently to the 3-electrode cell data at process block 305. For example, the anode voltage may be obtained directly from a reference electrode in a 3-electrode sensor system.

[0050] At process block 320, a training set is created based on the anode voltage and the corresponding features. In some examples, a portion of the plurality of features and corresponding anode voltages may be selected to create the training set. For example, the portion may be selected to reduce large featuring pooling, as well as mitigate overfitting.

[0051] At process block 325, a machine learning model is trained using the training set. For example, process 300 can build and evaluate a machine learning model to predict an anode voltage of a cell based only on measurements obtained from a 2-electrode system. In some examples, process 300 can train the machine learning model further based on features obtained from a 3- electrode system comprising a physical reference electrode. In some examples, the machine learning model can further predict Li plating conditions based on the predicting anode voltage (i.e., the virtual electrode module 145). Once the machine learning model is trained, it may be saved to the memory 135 of the server 120, with respect to FIG. 1, described above.Examples and Experiments

[0052] Described below are experimental setups and validations of the disclosed system and methodology. Here, an example of a virtual reference electrode that can dynamically and continuously monitor the AV of a 2-electrode (2E) cell throughout its lifetime without the need for a physical 3rdreference (FIG. 4C) is described. This virtual reference electrode is powered by deep learning models that are trained and validated on a large dataset (9.7 M datapoints) of 3-electrodeQB\98068473.1 102024-209-2 / 790482.00539(3E) coin cells cycled under a variety of conditions. This artificial intelligence (Al) powered virtual reference electrode takes only features available in 2E cells and outputs the predicted AV. These deep learning models are trained on measured data that is closer to the ground truth. An overall root mean squared error (RMSE) of 0.022 V with mean absolute error (MAE) of 0.017 V when averaged across this test set is reported, with no degradation in accuracy between early and late cycles. The accuracy of this virtual reference electrode is further validated experimentally with scanning electron microscopy (SEM), which indicates the absence of Li metal plating at a predicted AV near the thermodynamic threshold (i.e., 0.01 V vs. Li metal) and the presence of Li metal plating at a predicted AV just below the thermodynamic threshold (i.e., - 0.02 V vs. Li metal). With this Al-powered virtual reference, a smart BMS that can respond dynamically to predicted failure modes without the need for additional sensing equipment is developed, allowing us to push battery performance to the materials limit without compromising safety or stability. In particular, while low N / P ratios can enable increased energy density, increased likelihood of Li metal plating typically limits minimum N / P ratios to 1.1 or higher. When cycling 2E coin cells with low a N / P ratio (1.0), a ~7x enhancement in cycle life is observed using this smart charging protocol (e.g., BMS equipped with virtual reference) compared to a conventional charging baseline (i.e., constant-current-constant-voltage, CC-CV). This improvement is enabled by this virtual reference electrode’s ability to detect and avoid Li metal plating throughout battery cycling. This concept of a virtual reference electrode can be broadened to include other internal battery parameters that have physical meaning but are difficult to measure at scale (e g., pressure volume). Together, such virtual sensors powered by deep learning are no longer blind to catastrophic battery failures and may usher in a new era of battery health sensing, transforming how battery safety and optimization are managed.

[0053] FIG. 4A shows an absence of anode voltage information in conventional 2E cells. FIG. 4B shows two lithiation modes on graphite anode can be detected in 3E geometry. Further illustrated in FIGS. 4A-B is a reversible Li intercalation with positive anode voltage, as well as an example of catastrophic Li plating with negative anode voltage. FIG. 4C shows a workflow of building a virtual reference electrode, from 3E cell geometry and AV data collection to model training, and real-time prediction (i.e., WE: working electrode; CE: counter electrode; RE: reference electrode).QB\98068473.1 112024-209-2 / 790482.00539

[0054] To generate data that can accurately predict the AV throughout a battery’s lifetime, a 3E coin cell architecture is leveraged that enables high throughput and consistent battery cycling with simultaneous measurement of AV. In this study, lithium iron phosphate (LFP) and graphite (Gr) were selected as a model battery chemistry to build this 3E dataset. Battery failure and health are particularly difficult to predict in Gr||LFP chemistries since the overall cell voltage is mostly flat throughout battery operation. Using the 3E coin cell architecture, 55 Gr||LFP cells were cycled under a diverse array of charging conditions, generating ~10M data points of AV. Specifically, the “constant-current” (CC) regime of a typical constant-current constant-voltage (CCCV) charging protocol was cycled while keeping the CV regime and discharging steps consistent. The CC regime in this dataset spans from current densities of 0.79 mA cm’2to 2.37 mA cm’2and also contains cells with “stepped CC”, in which the charging current is incrementally decreased throughout the CC regime to mimic a more advanced charging profile. Together, this initial dataset captures a broad range of typical battery operating conditions with which this Al model can be trained to predict the corresponding AV throughout the charging process.

[0055] To maximize the utility of this Al-powered virtual reference electrode, an Al model was developed with high accuracy and robustness, while retaining interpretability and compatibility with large datasets. Indeed, 3E battery cycling at a commercial scale can contain billions of data points, which would further enhance the accuracy of this Al model if it can scale with the dataset. To build this Al model, battery domain knowledge was first leveraged to generate 65 features (all measurable or calculable in 2E cells) that are related to AV. This large pool of features encompasses (1) direct measurable parameters (e.g., cell voltage, current, capacity), (2) engineered features (e.g., throughput, overpotential), and (3) lagged features (e.g., measurable and engineered features from seconds prior), all of which may be salient in predicting AV. After specifying the training and test set (78% / 22% split), random forest recursive feature elimination (RF-RFE) was used on the training set to reduce the large feature pool and mitigate overfitting, resulting in 5 selected features that are then used in training and testing of this Al model (FIG. 5A). An ensemble of 5 feedforward neural networks demonstrated the highest accuracy out of all the model architectures tested, with an overall root mean square error (RMSE) of 0.022 V on the test dataset, a mean absolute error (MAE) of 0.017 V (FIG. 5B) (2.2 M data points), and a minimum single-cycle RMSE of 0.006 V (FIG. 5C). In some examples, these error metrics remain robust throughout battery lifetime without a decay in accuracy and highlight the ability of this AlQB\98068473.1 122024-209-2 / 790482.00539 model to continuously predict AV. Indeed, the single-cycle RMSE of cycle 20 (0.013 V) is comparable to that of cycle 100 (0.016 V), predictions validated by this 3E direct measurement that have not been demonstrated in the past. Considering this relatively small dataset compared to those used to train foundation models (e.g., large language models), the high prediction accuracy exhibited by this virtual reference electrode is quite promising with room for even further improvement. Much like how the performance of large language models scales with the data size, the accuracy of this virtual reference electrode increases with larger and more diverse datasets.

[0056] Beyond simply building an Al model that is black box, understanding how and why this resulting model functions effectively will further bolster confidence in its predictive capabilities. From battery domain knowledge, certain features strongly correlate with AV. In particular, cell voltage is embedded in the very definition of AV (i.e., difference between cathode voltage and cell voltage) and thus should be an input feature for this model. Additionally, the electrochemical state of graphite (i.e., degree of lithiation) thermodynamically relates to the AV and can be estimated by the instantaneous, single-cycle cell capacity, which provides an additional feature that is predictive for anode potential. Interestingly, all 5 model features selected by the RF- RFE process directly contain either cell voltage and / or capacity, suggesting that this Al model might be able to “learn” the underlying physics of predicting AV that aligns with this domain knowledge. For example, the top feature selected with a relative importance of 60% was Energy, a feature that captures the holistic impact of both cell voltage and capacity on AV through its very definition (i.e., energy is the product of voltage and capacity). Lagged features like Voltage-8 (i.e., voltage value “8” seconds prior to the present measurement) were found to help with predicting AV. Collectively, Voltage-8, Energy-8, and Power-8 have a relative feature importance of 31%, indicating that the present AV prediction depends on these parameters measured (or calculated) 8 seconds in the past.

[0057] FIG. 5A shows 3E cell data that was first split into training and test sets (78% / 22%), followed by the creation of 65 features based on battery-domain knowledge. Finally, 5 features were selected via RF-RFE. FIG. 5A shows a 2D density plot of predicted AV versus measured AV in all cells in test set. FIGS. 5C-E show predicted and measured AV comparison at cycle 20 and cycle 100, as well as cycle 64.

[0058] Using this virtual reference electrode in this Gr||LFP 2E cell, the AV can be continuously monitored during charging without additional sensing equipment and perform post-QB\98068473.1 132024-209-2 / 790482.00539 mortem analysis of the graphite surface at various charging states. At predicted AV just above the Li plating threshold (i.e., 0.01 V predicted AV, shown in FIG. 6B), this scanning electron microscopy (SEM) experiments show that the graphite surface is smooth and similar to pristine graphite, indicating that Li metal plating has not occurred (FIG. 6B). At later stages of charging where this virtual reference predicted a negative AV (-0.02 V predicted AV, shown in FIG. 6C), filamentary and mossy structures were observed on the graphite surface, which are direct markers of Li metal plating (FIG. 6C). Without additional sensing equipment in conventional 2E cells, such a catastrophic failure mode would go undetected during operation and likely compromise the safety and stability of the battery. Instead, 2E cells equipped with this virtual reference electrode are able to reliably detect the onset of Li metal plating similar to 3E cells with a physical reference. This capability to continuously monitor and avoid catastrophic failure without physical sensing equipment is significant, making it now possible to intelligently optimize battery performance without the need for any changes in battery manufacturing or chemistry.

[0059] FIG. 6A further shows an absence of anode voltage in 2E system without virtual reference electrode. FIG. 6B illustrates an ability of predicting anode voltage in 2E system with virtual reference electrode (virtual 3E system). Cell voltage & predicted AV plots under FIG. 6B show safe condition, while d shows an unsafe condition. SEM images of the top surface of graphite anode at predicted AV are shown in c (0.01 V) and e (- 0.02 V).

[0060] A smart charging protocol was designed for 2E cells equipped with this virtual reference electrode controlled by anode voltage safety, aiming to prevent Li metal plating by maintaining the predicted AV above 0 V vs. Li metal (FIG. 7B). This mitigates capacity degradation associated with Li metal reactivity and safety risks associated with short circuits, ultimately prolonging the cycling stability without requiring changes in battery chemistry or architecture. This smart BMS starts battery charging at a constant current density of 2.37 mA cm’2that dynamically decreases in decrements of 0.237 mA cm'2each time the real-time, predicted AV reaches 0.02 V vs. Li metal. Charging continues until the last constant current step falls to nominal 0.948 mA cm'2. The safety threshold was set at 0.02 V as a buffer considering the mean absolute error of this virtual reference electrode is 0.017 V. For comparison, a 3E cell was cycled, equipped with a physical reference electrode, using a conventional 2E CCCV charging protocol controlled by cell voltage safety (FIG. 7A) with the same discharging conditions and battery chemistry as this smart charging protocol. This physical reference electrode allows for theQB\98068473.1 142024-209-2 / 790482.00539 measurement of the AV profiles during conventional CCCV charging and the comparison of them with the predicted AV profiles of this 2E smart BMS cells. Since lowN / P ratios often compromise battery safety and stability to enable higher energy densities, a lowN / P ratio (~ 1.0) Gr||LFP battery chemistry was used as a challenging stress test for this smart BMS to effectively prevent Li metal plating and extend cycle life. This smart BMS enabled a ~7x enhancement in the cycle life of Gr||LFP compared to the conventional CCCV baseline (FIG. 7C), with areal capacities (~2.5 mA cm'2) normalized between the two protocols. Post-mortem analysis reveals that this performance enhancement is a direct result of mitigating Li metal plating (FIGS. 7E-F), whereas metallic Li is observed both optically and with SEM in fully discharged CCCV cells at end-of-life (80% capacity retention) (FIG. 7F), the surface of the graphite anode remains smooth and pristine in smart BMS cells after 336 cycles (FIG. 7E). Moreover, Li metal plating in this CCCV cell occurs as early as the 4thcycle, during which the measured AV drops below 0 V vs. Li metal (FIG. 7C). This result highlights the challenge with cycling low N / P ratio cells that are prone to Li metal plating, as both the current and cutoff voltage for this CCCV protocol would not trigger metallic Li plating. This catastrophic failure mode would have remained silent and undetected with conventional measurables like overall cell voltage and capacity decay, which largely appear to be nominal within the first 30 cycles of CCCV.

[0061] FIGS. 7A-F illustrate an example implementation of a smart charging protocol enabled by virtual reference electrode. FIGS. 7A-B are schematic illustrations of smart charging and CCCV charging, based on cell voltage safety and anode voltage safety. FIG. 7C illustrates capacity retention of smart charging (green) versus conventional CC-CV charging (brown), while FIG. 7D illustrates plots of measured AV in CC-CV (shown in solid brown line), predicted AV in smart charging (shown in solid green line), and cell voltage of both protocols (dashed brown line for CCCV, dashed green line for smart charging). FIG. 7E shows SEM images of the top surface of graphite anode after end-of-life (80% capacity retention) for FIG. 7E, smart charging, and FIG. 7F, CC-CV charging.

[0062] Various embodiments, configurations, materials, devices, systems, methods, and techniques are disclosed herein. With respect to the devices and systems described above, certain alternative components and materials are described, none of which are intended to be limiting or required. The description of components of such devices and systems is intended to be illustrative only, and neither a minimum nor limit of the types of components that could be used in variousQB\98068473.1 152024-209-2 / 790482.00539 embodiments hereof. Similarly, the methods described herein are explained with reference to optional steps and modifications, none of which are intended to be limiting or required. The methods described herein can be performed using hardware such as (or including) the devices and systems described herein but need not be implemented through such hardware except in specific examples that identify the use of such hardware.

[0063] In the foregoing specification, implementations of the disclosure have been described with reference to specific example implementations thereof. It will be evident that various modifications may be made thereto without departing from the broader spirit and scope of implementations of the disclosure as set forth in the following claims. The specification and drawings are, accordingly, to be regarded in an illustrative sense rather than a restrictive sense.QB\98068473.1 16

Claims

2024-209-2 / 790482.00539CLAIMSWhat is claimed is:

1. A battery management system (BMS) for controlling a charging protocol of a battery, the BMS comprising: a battery; a processor in configure to control systems for charging the battery; a memory in communication with the processor and having instructions stored thereon that, when executed, cause the processor to: receive a plurality of parameters corresponding to a state of the battery; provide the plurality of parameters to a trained machine learning model; generate, via the machine learning model, a virtual reference electrode; determine a predicted anode voltage of the battery using the virtual reference electrode; and control the system via the charging protocol for charging the battery using the predicted anode voltage of the battery.

2. The BMS of claim 1, wherein the systems for charging the battery include a charger connected to the battery.

3. The BMS of claim 2, wherein the memory having instructions stored thereon that, when executed, further cause the processor to adjust a current output of the charger based on the predicted anode voltage.

4. The BMS of claim 3, wherein adjusting the current output of the charger prevents lithium metal plating of the battery.

5. The BMS of claim 1, further comprising a two-electrode sensor system coupled to the battery, wherein the plurality of parameters corresponding to the state of the battery comprises data obtained from the two-electrode sensor system.QB\98068473.1 172024-209-2 / 790482.005396. The BMS of claim 5, wherein the data obtained from the two-electrode sensor system comprises at least one of a cell voltage, a cell current, or a cell capacity.

7. The BMS of claim 1, wherein the machine learning model has been trained using a plurality of features extracted from a three-electrode sensor system and a plurality of corresponding anode voltages obtained using a physical reference electrode of the three- electrode sensor system.

8. The BMS of claim 1, wherein the machine learning model was trained by: obtaining three-electrode sensor system data measured from one or more battery cells; extracting a plurality of features from the three-electrode sensor system data; obtaining a plurality of anode voltages corresponding to each feature of the plurality of features; creating a training set based on the plurality of anode voltages and the plurality of features; and training the machine learning model using the training set.

9. The BMS of claim 8, wherein the plurality of features comprises at least one of a cell voltage, a charging current, a battery cell capacity, an overpotential, or a cell throughput.

10. A computer readable storage medium having instructions stored thereon that, when executed by a processor, cause the processor to perform steps comprising: receive a plurality of parameters corresponding to a state of a battery; provide the plurality of parameters to a trained machine learning model; generate, via the machine learning model, a virtual reference electrode; determine a predicted anode voltage of the battery using the virtual reference electrode; and control a charging of the battery using the predicted anode voltage of the battery.QB\98068473.1 182024-209-2 / 790482.0053911 . The computer readable storage medium of claim 10, wherein controlling the charging of the battery includes changing a current or a voltage supplied by a charger coupled to the battery to maintain a predicted anode voltage.

12. The computer readable storage medium of claim 11, where in the current or the voltage is changed to prevent lithium metal plating of the battery.

13. The computer readable storage medium of claim 10, wherein the plurality of parameters corresponding to the state of the battery comprises data obtained from a two- electrode sensor system.

14. The computer readable storage medium of claim 13, wherein the data obtained from the two-electrode sensor system comprises as least one of a cell voltage, a cell current, or a cell capacity.

15. The computer readable storage medium of claim 10, wherein the instructions further cause the processor to: determine, via the machine learning model or the virtual reference electrode, a predicted lithium plating condition based on the predicted anode voltage of the battery.

16. The computer readable storage medium of claim 10, wherein the machine learning model has been trained using a plurality of features extracted from a three-electrode sensor system and a plurality of corresponding anode voltages obtained using a physical reference electrode of the three-electrode sensor system.

17. The computer readable storage medium of claim 10, wherein the machine learning model was trained by: obtaining three-electrode sensor system data measured from one or more battery cells; extracting a plurality of features from the three-electrode sensor system data; obtaining a plurality of anode voltages corresponding to each feature of the plurality of features;QB\98068473.1 192024-209-2 / 790482.00539 creating a training set based on the plurality of anode voltages and the plurality of features; and training the machine learning model using the training set.

18. The computer readable storage medium of claim 17, wherein the plurality of features comprises at least one of a cell voltage, a charging current, a battery cell capacity, an overpotential, or a cell throughput.QB\98068473.1 20

Images