Battery degradation modeling
Three ways to model battery degradation, and when each one is worth the effort.
From empirical capacity-fade curves to mechanism-level SEI and plating models, the full spectrum is available through PyBaMM-based simulation in Ionworks.
The test matrix is too big
Cycle life testing is slow. A single aging study at 1C and 25 °C can take 18 months or more to reach 80% capacity retention, and that result tells you nothing about what happens at 2C, 45 °C, or a different SoC window. Multiply by the number of conditions you actually care about, and the test matrix becomes unmanageable.
Modeling fills the gap between the conditions you can test and the conditions you need to predict. “Degradation model” means very different things depending on who you ask. A data scientist training XGBoost on cycling features and an electrochemist parameterizing SEI diffusion kinetics are solving the same problem at different levels of fidelity, cost, and risk.
The framework below organizes degradation modeling into three tiers. Each tier answers a progressively deeper question: how much capacity is lost, what is being lost, and why it is being lost.
Three tiers of degradation modeling
| Tier 1: empirical / ML | Tier 2: degradation modes | Tier 3: degradation mechanisms | |
|---|---|---|---|
| Core question | How much capacity fades? | What is being lost (LLI, LAM)? | Why is it being lost (SEI, plating, cracking)? |
| Fidelity | Low | Medium | High |
| Data cost | Low | Medium | High |
Pick the tier that matches your data budget and the question you need answered. Many teams operate across two or three tiers simultaneously, using empirical models for fleet monitoring while running mechanism models for protocol design.
Tier 1
Empirical and ML models
What they predict. Capacity fade (Q vs. cycle number) and resistance increase as functions of cycling conditions. Input-output models with no representation of internal cell state.
Methods. Standard regression and ML techniques applied to cycling features: elastic net, XGBoost, Gaussian processes, and classical power-law fits. The landmark work by Severson et al. (2019, Nature Energy) demonstrated that features extracted from the first 100 cycles can predict cycle life with high accuracy across a population of cells, using variance in the discharge voltage curve as a key predictor.
Power-law models (Q = Q₀ · n^b) remain common for quick extrapolation within a known operating regime. Gaussian processes add uncertainty quantification, which matters when you are making warranty decisions.
When to use. Tier 1 models work best when you have large datasets on a known cell chemistry and format, and you are interpolating within tested conditions. They fail outside the training distribution. A model trained on 1C/25 °C data cannot reliably predict behavior at 3C/45 °C, because it has no representation of the mechanisms that change with rate and temperature.
In Ionworks. Ionworks ingests cycling data from Maccor, Neware, Arbin, BioLogic, and Novonix, organizing raw files around cell instances and computed cycle metrics (discharge capacity, coulombic efficiency, resistance at fixed SoC). The Python SDK exposes these structured datasets for custom ML model training, so teams can build Tier 1 models directly on clean, version-controlled data without writing their own parser for each cycler format.
Tier 2
Degradation mode models
What they predict. LLI, LAM (anode and cathode separately), and resistance growth as explicit state variables. Rather than treating capacity fade as a single number, degradation mode models decompose the loss into its constituent parts.
Diagnostic methods. DVA and ICA are the primary non-destructive diagnostics. Shifts and amplitude changes in dQ/dV or dV/dQ peaks map to specific modes: a shrinking graphite staging peak indicates LAM at the anode, while a uniform voltage curve shift points to LLI. Half-cell teardown data provides ground truth for validating these assignments.
When to use. When you need to know which electrode is limiting life. Degradation mode analysis separates calendar aging contributions (typically LLI-dominated) from cycle aging contributions (which may include LAM). The limitation is that knowing “LLI is the dominant mode” does not tell you whether LLI comes from SEI growth, lithium plating, or both.
In Ionworks. PyBaMM tracks LLI and LAM as state variables within its electrochemical models. Ionworks parameterization workflows fit mode evolution trajectories to experimental data from standard test protocols, using train/test splits to validate predictive accuracy.
Tier 3
Degradation mechanism models
What they predict. Root-cause physics. Tier 3 models represent the electrochemical and mechanical processes that produce LLI and LAM, enabling prediction under conditions that have never been tested.
SEI growth
Electrolyte reduction at the anode forms a passivation layer that continues to grow throughout cell life. In the diffusion-limited regime, SEI thickness follows sqrt(t) kinetics: rapid initial growth that slows as the layer thickens. SEI growth dominates calendar aging and is accelerated by high temperature and high SoC (both of which increase the driving force for electrolyte reduction at the anode surface).
PyBaMM implements three SEI growth submodels: reaction-limited, solvent-diffusion limited, and electron-migration limited. Each captures a different rate-limiting step, and the appropriate choice depends on cell chemistry and operating conditions.
Lithium plating
When the anode potential drops below 0 V vs. Li/Li⁺, metallic lithium deposits on the graphite surface instead of intercalating. Plated lithium follows one of three paths: it strips back during discharge (reversible), it reacts with electrolyte to form additional SEI (partially reversible), or it loses electrical contact and becomes dead lithium (irreversible). Dead lithium is permanent LLI.
Lithium plating is the central concern in fast-charge protocol design. The engineering question is straightforward: what is the maximum C-rate that keeps the anode potential above 0 V across a given temperature range? PyBaMM's plating submodels (reversible, partially reversible, and irreversible) track plated lithium, stripping, and dead lithium as separate quantities, making the anode potential constraint directly observable in simulation.
Electrode mechanical effects
Repeated lithiation and delithiation cause volume changes in electrode particles (roughly 10% for graphite, far more for silicon-containing anodes). Non-uniform lithium concentration gradients within particles generate diffusion-induced stress. When stress exceeds the fracture toughness, cracks propagate.
The critical coupling is “SEI on cracks”: fresh particle surface exposed by cracking is immediately available for electrolyte reduction, creating new SEI and accelerating LLI. In PyBaMM, crack length is tracked as a state variable, and the SEI-on-cracks option couples cracking with SEI growth rate on the newly exposed surface area.
Electrolyte depletion
Every SEI-forming side reaction consumes electrolyte solvent. At high temperature or high voltage, oxidation reactions at the cathode contribute additional solvent consumption. Over thousands of cycles, the cumulative loss reduces ionic conductivity and can cause localized dry-out in the porous electrode structure. Electrolyte depletion is often the life-limiting factor in cells operated at elevated temperature or high upper cutoff voltage.
Coupled mechanisms
Real cells do not degrade by one mechanism at a time. SEI growth consumes lithium and electrolyte. Particle cracking exposes fresh surface that accelerates SEI growth. Lithium plating produces dead lithium and triggers additional SEI formation on plated deposits. All four mechanisms run in parallel, and their interactions determine the shape of the aging trajectory.
PyBaMM's coupled degradation framework runs SEI growth, lithium plating, particle cracking, and LAM simultaneously within a single simulation. The coupling is physical, not post-hoc: crack-driven surface area increase feeds back into the SEI growth rate within each time step.
When to use. Tier 3 models are the right choice when you need to predict degradation under novel duty cycles, optimize charging protocols to avoid plating, or make electrode design decisions (particle size, porosity, coating thickness) informed by mechanism-level tradeoffs. The cost is higher parameterization effort, typically requiring half-cell data, EIS, and sometimes post-mortem analysis to constrain model parameters.
In Ionworks. Ionworks provides parameterization workflows for all Tier 3 submodels: SEI kinetic parameters, plating exchange current density, cracking stress thresholds, and LAM rates for both electrodes. These parameters are fit to experimental data with train/test validation, and the resulting parameterized models are stored alongside the cell data they were derived from.
Tier comparison: tradeoff matrix
| Dimension | Tier 1 | Tier 2 | Tier 3 |
|---|---|---|---|
| Implementation difficulty | Low | Medium | High |
| Data requirement | Low (cycling data only) | Medium (OCV, rate capability, half-cell) | High (teardown, half-cell, EIS) |
| Accuracy (interpolation) | High | High | Highest |
| Accuracy (extrapolation) | Low | Medium | Highest (physics extrapolates) |
| Predictability for new duty cycles | Poor | Medium | Maximum |
| Physical interpretability | None | Partial (identifies what) | Full (explains why) |
| Design guidance | None | Partial (identifies limiting electrode) | Maximum (target specific mechanisms) |
The right tier depends on the question. Warranty estimation for a well-characterized cell at known operating conditions? Tier 1 is likely sufficient. Diagnosing why a field failure occurred earlier than expected? Tier 2 points to the limiting mode. Designing a fast-charge protocol for a new cell format at 0 °C? Tier 3 is where the physics lives.
How Ionworks fits in
How Ionworks supports all three tiers
01
Structured data ingestion
Cycling data from Maccor, Neware, Arbin, BioLogic, and Novonix is parsed and organized around cell instances and cycle metrics. Every cell has a traceable history: formation protocol, cycling conditions, and computed metrics like discharge capacity, DCIR, and coulombic efficiency per cycle. Clean, structured data is the foundation for any tier of modeling.
02
Parameterized degradation models
Fit SEI growth rate constants, plating exchange current densities, and LAM parameters to experimental aging data. Ionworks parameterization uses train/test splits to report predictive accuracy, not just goodness-of-fit. Parameterized models are versioned and linked to the experimental datasets they were trained on.
03
Degradation sweeps
Vary C-rate, temperature, and SoC window systematically to predict cycle life across a matrix of operating conditions. A single parameterized model can generate hundreds of simulated aging trajectories, replacing months of physical testing with simulation runs that complete in minutes.
04
Protocol optimization
Find the fastest charge rate that avoids lithium plating across a temperature range. Ionworks uses PyBaMM's anode-potential monitoring to identify the plating onset boundary, then searches the protocol space (CC-CV profiles, multi-step profiles, temperature-dependent limits) for the optimal balance between charge time and cycle life.
All four capabilities connect in a single workflow: ingest data, parameterize models, sweep conditions, optimize protocols. The Python SDK provides full programmatic access, so teams can integrate Ionworks into existing analysis pipelines or automate parameterization across a library of cell chemistries.
Frequently asked questions
Move beyond empirical curve fits
Book a demo to see how Ionworks connects your cycling data to PyBaMM's physics-based degradation models.