Universal Battery Performance & Degradation Model

• The UBDM is a unified framework combining convex piecewise-affine functions and physics-informed neural networks to predict battery degradation accurately.
• It leverages normalized battery metrics and hybrid machine learning to integrate seamlessly with optimization tasks and digital twin-based management systems.
• The model’s transferability enables benchmarking, prognostics, and adaptive control across applications from grid optimization to real-time battery health monitoring.

A Universal Battery Performance and Degradation Model (UBDM) is a mathematical and algorithmic framework for predicting, benchmarking, and optimizing the degradation and operational behavior of batteries—typically lithium-ion—across widely varying chemistries, form factors, and use cases. UBDMs are designed to be modular, extensible, and transferable, enabling a unified representation of complex aging phenomena for integration into power system optimization, asset sizing, control, and digital-twin-based battery management systems. Recent implementations span piecewise-affine convex maps, physics-informed machine learning hybrids, input convex neural network surrogates, and hybrid partial differential equation-based architectures, each targeting computational tractability, scalability, and transferability across chemistries and dynamic operating scenarios (Fortenbacher et al., 2017, Zhang et al., 24 Jan 2025, Bills et al., 2020, Lin et al., 2021, Mallick et al., 16 May 2025).

1. Mathematical Foundations of Universal Degradation Mapping

UBDMs originate from the concept of a degradation map, wherein the incremental capacity loss $\Delta Q$ over a small step $\Delta t$ is modeled as a convex piecewise-affine (PWA) function of instantaneous state and control. A canonical formulation is

$$\dot{J}_{\text{deg}} = \max_{i=1,...,K} \left\{ a_{1,i} P_{\text{bat}} + a_{2,i} E + a_{3,i} C_E \right\},$$

where $P_{\text{bat}}$ is battery power, $E$ is stored energy, and $C_E$ the system energy capacity. Normalization yields

$$\frac{\dot{J}_{\text{deg}}}{C_E} = \max_{i=1...K}\, \{\,\alpha_{1,i}\,p + \alpha_{2,i}\,E_n + \alpha_{3,i}\,\}$$

with $p = P_{\text{bat}}/C_E$, $E_n = E/C_E$. Each affine region is defined over a convex hull in the $(p, E_n)$ plane. This structure, combined with convexification (e.g., Delaunay triangulation of the raw, potentially nonconvex map), guarantees tractability for integration into convex optimal control and simulation (Fortenbacher et al., 2017).

Hybrid and machine-learning-driven UBDMs augment or replace the degradation map with neural or physics-informed surrogates, such as universal-ODEs (U-ODEs) (Bills et al., 2020), partial input-convex neural networks (PICNNs) (Mallick et al., 16 May 2025), or physics-informed neural PDE frameworks (Zhang et al., 24 Jan 2025). These architectures maintain a separation between physically interpretable variables and nonparametric corrections, preserving generalizability and interpretability across scenarios.

2. Scaling Laws and Chemisry-Agnostic Normalization

A UBDM is designed to scale seamlessly across pack configurations, chemistries, and power/energy ratings. The cell-level degradation law is normalized using the following prescription:

• Energy capacity: $$C_E = N_{\text{par}} \, C_Q^c \, N_{\text{ser}} \, \bar{V}_{\text{oc}}^c$$
• Power: $$P_{\text{bat}} \approx I_{\text{bat}} N_{\text{ser}} \, \bar{V}_{\text{oc}}^c$$
• Normalized variables: $E_n = E/C_E$, $p = P_{\text{bat}}/C_E$

After normalization, the universal degradation map maintains identical shape regardless of $N_{\text{ser}}$, $N_{\text{par}}$, or individual cell properties. For machine-learning-based models, chemistry and configuration-agnostic features—such as 2D histograms of (current, voltage) dwell times—create an input space that is robust to protocol and hardware variations, supporting transfer learning across platforms (Zhang et al., 24 Jan 2025).

3. Machine-Learning–Hybrid and Data-Driven Extensions

UBDMs have evolved to incorporate hybrid architectures that combine physics-based backbone models, such as the pseudo-two-dimensional (P2D) or single-particle models, with machine-learned residual or surrogate networks:

• Hybrid P2D+ML Residuals: The state and parameter evolution equations for solid and electrolyte concentrations, potentials, and temperature are augmented with neural-residual functions $R_u$ and $R_\theta$, addressing systematic model bias and unmodeled effects. Uncertainty calibration is performed using negative log-likelihood losses, and online updates are realized with ensemble Kalman filters (Lin et al., 2021).
• Universal ODEs (U-ODEs): In Cellfit, battery state vectors evolve according to

$$\frac{du}{dt} = f_{\text{electrochem}} + f_{\text{thermal}} + f_{\text{degr}}$$

with mechanistic charge and resistance aging models supplemented by neural-net residuals that learn unmodeled or secondary physics, trained on time-series voltage and temperature trajectories (Bills et al., 2020).

• Partially Input-Convex Neural Networks: The ICNN-based UBDM learns $D(n,T,\text{SOC},r)$ as the battery capacity loss rate, ensuring convexity with respect to the charging rate $r$ while allowing for arbitrary nonconvex dependencies on cycle number, temperature, and SOC. This facilitates embedding in convex optimization routines for smart grid and V2G problems (Mallick et al., 16 May 2025).
• Physics-Informed Neural PDE Frameworks: UBDMs may adopt DeepHPM/physics-informed neural PDE solvers, where the solution and the underlying unknown right-hand-side PDE function are jointly learned from empirical (current, voltage) dwell-time distributions, supporting scenario-aware transfer learning and advanced feature engineering (Zhang et al., 24 Jan 2025).

4. Integration into Forecasting, Optimization, and Control

A central advantage of the convex PWA or convexified ML UBDMs is the ability to integrate directly into convex optimization problems, including:

• Optimal Power Flow (OPF): UBDM constraints are imposed in epigraph form, introducing auxiliary variables for instantaneous degradation rates, preserving problem convexity (LP or SOCP) (Fortenbacher et al., 2017).
• Asset Sizing and Placement: Embedding UBDM maps into mixed-integer LPs enables simultaneous optimization of placement, sizing, and operational scheduling of batteries with explicit tradeoffs between degradation cost and grid-service revenue (Fortenbacher et al., 2017).
• Multi-Objective V2G Optimization: The ICNN-based UBDM is employed as a real-time surrogate inside a multi-objective framework, trading off financial gains against lifetime degradation, ensuring computational tractability by maintaining convexity in the control variable (charging/discharging rate) (Mallick et al., 16 May 2025).
• Digital Twin Integration and BMS Functions: Advanced BMS implementations use scenario-adaptive, transfer-learned UBDMs as digital twins for prognostics, online health classification, knee-onset detection/prediction, and second-life repurposing (Zhang et al., 24 Jan 2025). Fine-tuning involves freezing physics layers while retraining only surrogate or interface layers on minimal new data, delivering highly adaptive yet computationally efficient models.

5. Benchmark Parameters and Chemistry-Specific Insights

UBDMs facilitate benchmarking across diverse Li-ion chemistries by providing tables of convex hull parameters for normalized degradation maps:

• LFP (LiFePO₄): 16-plane PWA, NRMSE = 3.33%; strong current dependence.
• NMC/LMO (LiMnNiCo/LiMn₂O₄): 12-plane PWA, NRMSE = 1.07%; pronounced coupling of capacity fade to current rate.
• LCO (LiCoO₂): 14-plane PWA, NRMSE = 1.06%; higher voltage/SoC dependence.

Comparative analysis reveals that LFP and NMC/LMO are more sensitive to current-induced aging, while LCO exhibits heightened degradation at high state of energy. Convex hull approximation errors remain below 3% for all chemistries, validating robustness for benchmarking and comparative studies (Fortenbacher et al., 2017).

6. Training, Validation, and Transfer Learning

UBDM development across data-driven and hybrid platforms employs rigorous multi-stage training protocols:

• Physics-based and hybrid models: Parameters are initialized to physically plausible values, then jointly trained on cycle data using composite loss functions calibrated for data and physical constraint satisfaction. Validation employs cross-validation (e.g., 5-fold by cell profile), and out-of-sample performance is evaluated on independent protocols or external cells (Lin et al., 2021, Zhang et al., 24 Jan 2025).
• Data-driven surrogates: Models such as ICNNs and DeepHPM-based PINNs are trained on large cycling datasets with periodic hold-out validation (e.g., NASA randomized cycling for ICNNs), ensuring performance generalizes across unseen cells and protocols (Mallick et al., 16 May 2025).
• Transfer learning: Scenario-aware pipelines enable global UBDMs to be fine-tuned to local dynamic cycling data by re-training only feature mapping or surrogate solution layers, minimizing the required amount of new labeled data. On-target adaptation recovers nearly all source accuracy and enables robust performance in dynamic load scenarios (Zhang et al., 24 Jan 2025).

7. Applications, Limitations, and Outlook

UBDMs have been adopted for diverse applications including grid-level OPF dispatch with aging-aware constraints, eVTOL mission-cycle thermal and state-of-health forecasting, adaptive V2G revenue optimization, and deployment as real-time BMS digital twins for online diagnostics and predictive control. Notable limitations include:

• Extrapolation outside the training/validation envelope may degrade performance, particularly for machine-learning-augmented UBDMs (Bills et al., 2020).
• Some feature sets require precise or periodic recalibration for novel chemistries, unusual cycling, or new ambient conditions.
• Current frameworks typically rely on capacity or impedance tests for ground-truth labeling; full in situ, non-invasive health indicators are an area of ongoing research (Zhang et al., 24 Jan 2025).

This suggests future UBDM advances will likely integrate multiscale coupling (cell, pack, system), probabilistic uncertainty quantification, and automated adaptation to new battery technologies and missions, converging toward extensible, plug-and-play digital models spanning the full electrochemical storage ecosystem (Fortenbacher et al., 2017, Zhang et al., 24 Jan 2025, Lin et al., 2021, Bills et al., 2020, Mallick et al., 16 May 2025).