High resolution raw data, unique cycle metrics – UHPC has valuable applications in development, quality, and machine learning lifetime prediction.
For any company dependent on the battery supply chain, whether sourcing cells for electric vehicles, qualifying novel materials or standing-up giga-factories, assessing cell quality, performance, and lifetime typically takes months to years of costly testing and is the single biggest bottleneck to energy storage innovation. But what if testing time was cut to a matter of weeks?
Degradation mechanisms in chemistries such as lithium-ion are non-trivial, often occurring at minuscule rates of reaction, making them difficult – or impossible – to capture on traditional spec cell testing equipment, requiring long-test times before differences are evident. This is where Ultra-High Precision Coulometry (UHPC) cell testing equipment excels, and data-driven predictive models thrive.
Filling the gaps between traditional cell testing and complex and/or destructive analytical techniques, UHPC offers quantifiable insights into electrochemical mechanisms which are otherwise invisible using other techniques. This article explores how cell and auto OEMs, materials companies, and world-class institutions use UHPC to cut an order of magnitude off their testing times, and how high-fidelity UHPC data unlocks opportunities for predictive Machine Learning (ML) methods.
UHPC was pioneered by Prof. Jeff Dahn in 2010 with the goal of accurately detecting low-rate electrochemical degradation in lithium-ion cells.1,2 In 2013 Dr. Chris Burns and Dr. David Stevens, both mentored by Prof. Dahn, founded NOVONIX and commercialized the first market-ready UHPC cell testing product, which has evolved and spread across the battery industry over the last 12 years.
UHPC equipment has been engineered to source and measure voltage and current to the highest precision and accuracy possible, allowing researchers to understand specific electrochemical degradation mechanisms during cell testing and measure metrics such as coulombic efficiency accurate to 10s of ppm. These mechanisms involve cell materials such as anodes, cathodes, electrolytes, etc., exchanging electrons at electrode surfaces in complex reactions. As degradation occurs, the associated processes which create charge balance in the cell while undergoing testing cause tiny differences in charge and discharge capacity. These processes are shown in Figure 2 and include, but are not limited to:
UHPC data can add a previously unattainable level of insight into cell chemistry and processes leading to end of life. These data therefore have important implications in the field of predictive analytics for batteries.
A variety of methods are used to predict the lifetime of cells under various conditions. These include:
Each method has advantages and disadvantages. For example: empirical models can be iterated quickly but do not generalize well, qualitative ranking provides a holistic picture of cell degradation but requires subject-matter expertise and excessive time, physics-based models give detailed performance predictions but require extensive parametrization, and ML models provide broad predictions at low cost but typically require large training data sets.
UHPC is historically reported in the literature adopting qualitative approaches, making head-to-head comparisons of similar systems such as different electrolyte additives in the same cells, comparisons of electrode materials, or direct measurements of specific mechanisms such as
lithium plating or self-discharge.3-5 However, the insights from UHPC can complement other modeling methods such as ML and may pave a way for more accurate prediction capabilities.
There is often a strong correlation between early electrochemical signatures and eventual cell failure.6 ML models thrive on exactly the kind of high-resolution UHPC data that captures those first, almost imperceptible signs of degradation. By training models on thousands of cells whose subtle early-cycle fingerprints are paired with their eventual end-of-life metrics, models can learn to map “day-one” electrochemical features to long-term performance outcomes. This ability to forecast lifetime after only a handful of cycles can shorten qualification and validation loops from months-years to days-weeks and turn UHPC into a launchpad for data-driven battery R&D.
Figure 3 depicts one method of how UHPC could be used for ML lifetime prediction. This method was employed during the case-study in this article.
A recent study performed by NOVONIX and SandboxAQ demonstrated the potential value of UHPC for ML cell cycle life predictions. The goal was to develop a model that was generalizable across various degradation signatures based on approximately 1 month of UHPC testing.
Cylindrical cells from three manufacturers were tested using various voltage ranges and temperatures. All cells contained high-Ni positive electrodes, two types contained silicon-graphite blend negative electrodes, and one cell type contained a graphite-only negative electrode. Testing conditions were selected to cover specific degradation mechanisms based on the known cell compositions:
Cells were cycled in triplicates for each test condition (cell type, temperature, voltage range); on a NOVONIX UHPC system at C/10-C/10 constant current cycling for approximately 4 weeks, and on a common R&D-spec system for up to 2500 cycles (~2 years) at C/3-C/3 (CCCV charge to C/20).
The capacity retention of long-term cycling data was used as the target metric for ML predictions. Various capacity retention thresholds were considered, for example 90%, 85%, and 80%. To construct features that capture cell degradation, residuals of differential capacity (dQ/dV) curves between two cycles were computed from the paired UHPC data. Figure 5 shows
that using merely 6 UHPC cycles, the number of cycles to 85% capacity retention was predicted to be within 46 cycles for cell types not used in model training. These results are a dramatic improvement over traditional methods used in the literature, such as early discharge capacity trends and changes in low resolution voltage features. 6-8
How well can UHPC-derived features generalize to arbitrary data sets? The same features described above, developed on a small, curated data set were directly applied to a data set of over 4000 cells composed of various chemistries, including both Ni-based and LFP positive electrodes, and graphite and silicon-containing negative electrodes, from a variety of vendors. Each testing condition contained UHPC/long-term paired cells. The same ML pipeline described above was applied to this data set, with features constructed from the residual UHPC differential capacity curves.
By directly applying the ML feature generation approach developed on a small, curated data set to a data set 100 times larger, with merely 25 UHPC cycles as input, the number of cycles to 85% capacity retention was predicted to within 108 cycles.
The success of this method to generalize arbitrary data sets is due to how electrochemical processes that lead to cell degradation over a long-time scale can be encoded in early-life
features with high-fidelity data. The precision and accuracy of UHPC makes this ML approach possible.
The results shown in this article have important implications on all aspects of battery development and commercialization, spanning material evaluation to warranty estimation. Utilizing UHPC channels to complement a suite of standard cyclers can significantly impact decisions and progress using short-term, high-throughput precision tests and predictive analytics. New measurement techniques and equipment in combination with advancements in ML and AI have significant implications for the battery lab of the future. Paired with proper data aggregation, labeling, and organization, investments in R&D will go further and faster.
Stephen Glazier is the Director, Cell Technology at NOVONIX Battery Technology Solutions in Halifax, Nova Scotia, Canada, overseeing R&D Services prototyping programs and Ultra-High Precision Coulometry technology development. Prior to joining NOVONIX in 2018, Stephen completed his MSc and PhD under the supervision of Dr. Jeff Dahn at Dalhousie University.
Marc Cormier is a battery data analytics consultant for SandboxAQ. Marc has experience developing enterprise software solutions for battery R&D data, novel analytics methods, and applying quantum mechanical methods to study battery materials. He completed a PhD under the supervision of Dr. Jeff Dahn at Dalhousie University.
Dr. Ty Sours is a senior researcher in the AI Simulation Group at SandboxAQ. He has a Ph.D. in chemical engineering from the University of California, Davis, where he studied the application of multiscale atomistic models for materials discovery. At SandboxAQ, Tyler is focused on developing tools to integrate AI and simulation with experimental data to accelerate materials and process optimization.
