PINEAPPLE: Physics-Informed Neuro-Evolution Algorithm for Prognostic Parameter Inference in Lithium-Ion Battery Electrodes

The paper introduces PINEAPPLE, a novel framework combining physics-informed neural networks with evolutionary search to enable rapid, accurate, and robust real-time inference of internal lithium-ion battery electrode parameters solely from voltage-time discharge curves, facilitating non-destructive state-of-health diagnostics for next-generation battery management systems.

The Gist

Imagine you have a high-tech smartphone, but instead of a screen, it has a "black box" battery inside. You can see the battery's voltage (how much "juice" is left), but you can't see what's happening inside the battery cells. Is the lithium moving slowly? Is the material cracking? Is the battery getting sick?

Currently, to find out, you'd have to take the battery apart (destructive testing) or wait until it dies. That's like trying to diagnose a heart attack by waiting for the patient to stop breathing.

This paper introduces PINEAPPLE (Physics-Informed Neuro-Evolution Algorithm for Prognostic Parameter Inference in Lithium-ion battery Electrodes). Think of PINEAPPLE as a super-smart, non-invasive medical scanner for batteries.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Black Box" Mystery

Batteries degrade over time. We know they get worse, but we usually only see the symptoms (the voltage drops faster). We don't know the cause inside the chemical soup.

• Old Way: Use complex math equations (Physics) to guess what's inside. It's accurate but takes forever to calculate—too slow for a real-time app on your phone.
• New Way (Data-Only): Use AI to guess based on past data. It's fast, but it's a "black box" AI. It might guess right today, but if the battery behaves slightly differently tomorrow, the AI gets confused because it doesn't understand why the battery works.

2. The Solution: PINEAPPLE's "Hybrid Brain"

PINEAPPLE combines the best of both worlds: Physics (the rules of how batteries should work) and AI (the speed of learning).

Think of it like teaching a student to solve a puzzle:

• The Physics Part: You give the student the rulebook (The Laws of Physics, specifically how lithium ions diffuse through a sponge-like material).
• The AI Part: You give the student a super-fast brain that can look at a puzzle and instantly guess the solution.

PINEAPPLE trains this AI brain using a method called Meta-Learning. Imagine you are training a chef. Instead of teaching them to cook one specific dish, you teach them the fundamental techniques of cooking so well that they can instantly cook any new dish they've never seen before.

• Result: PINEAPPLE can look at a battery's voltage curve and instantly predict the internal chemical state (like how fast lithium is moving) without needing to run slow, heavy calculations. It's 10 times faster than traditional methods.

3. The "Evolution" Step: Finding the Hidden Truth

Here is the tricky part. You can't see the inside of the battery. You only see the voltage curve on the outside. This is like trying to guess the ingredients of a soup just by tasting the broth. Many different ingredient combinations could taste the same. This is called an "Ill-posed problem."

To solve this, PINEAPPLE uses an Evolutionary Algorithm.

• The Analogy: Imagine a room full of 20 chefs (a population). Each chef guesses a different recipe (a set of internal battery parameters) to explain the soup's taste (the voltage curve).
• Natural Selection: The chefs whose recipes produce a soup that tastes closest to the real voltage curve get to "survive" and pass their recipe on. The others are discarded.
• Mutation: The surviving chefs tweak their recipes slightly (maybe a pinch more salt, a bit less pepper) and try again.
• Result: After a few rounds (generations), the group converges on the most likely recipe that explains the soup.

In the paper, this happens in seconds. It finds the hidden internal parameters (like the diffusion coefficient, which is basically "how fast the lithium moves") that best explain the battery's behavior.

4. What Did They Discover?

The team tested PINEAPPLE on real batteries from a public database (CALCE). They watched the batteries age over hundreds of cycles.

• The Anode (Negative Side): They found that the "speed limit" for lithium moving through the anode steadily and predictably slowed down. This is like a highway getting more traffic jams over time. This matches what scientists already know happens (a layer called SEI grows and blocks the path).
• The Cathode (Positive Side): The story here was more complex. Sometimes the speed slowed down, sometimes the structure got messy. It wasn't a straight line.
• The Big Win: Because PINEAPPLE understands the physics, it didn't just say "the battery is 80% healthy." It said, "The battery is 80% healthy because the anode is getting clogged, but the cathode is still okay."

Why Does This Matter?

This is a game-changer for Battery Management Systems (BMS)—the computer brain inside your electric car or phone.

1. Non-Destructive: You don't have to tear the battery apart to know its health.
2. Real-Time: It's fast enough to run on the computer inside your car while you are driving.
3. Explainable: It tells you why the battery is failing, not just that it is failing. This allows for smarter charging strategies (e.g., "Don't charge this battery fast today because the anode is clogged") to extend its life.

In summary: PINEAPPLE is a smart, physics-aware detective that looks at a battery's "vital signs" (voltage), uses a super-fast AI brain to understand the rules of chemistry, and uses an evolutionary search to figure out exactly what's happening inside the battery's "organs," all in a fraction of a second. It turns a black box into a transparent window.

Technical Summary

1. Problem Statement

Accurate, real-time, and non-destructive estimation of the internal electrochemical states of Lithium-ion (Li-ion) batteries is critical for predicting degradation, optimizing usage, and extending lifespan. However, current approaches face significant limitations:

• Purely Data-Driven Methods: While flexible, they often lack mechanistic interpretability, struggle with generalization across different operating conditions, and frequently exclude critical internal physical parameters (e.g., diffusion coefficients) due to the difficulty of direct measurement.
• Physics-Based Models (e.g., Single Particle Model - SPM): These offer high interpretability but are computationally expensive, making them unsuitable for real-time onboard implementation. Furthermore, they require extensive parameter calibration, which is difficult given inter-battery variability and the "ill-posed" nature of inferring internal states from macro-scale voltage-time (V-t) measurements (where multiple parameter sets can produce similar voltage curves).

The core challenge is to develop a framework that enables rapid, scalable, and interpretable inference of internal battery parameters (like diffusion coefficients) solely from voltage-time discharge curves, without requiring destructive testing or detailed material-specific calibration.

2. Methodology: The PINEAPPLE Framework

The authors propose PINEAPPLE (Physics-Informed Neuro-Evolution Algorithm for Prognostic Parameter Inference in Lithium-ion battery Electrodes), a hybrid framework integrating Physics-Informed Neural Networks (PINNs) with an Evolutionary Search Algorithm.

A. Underlying Physics Model

The framework utilizes the Single Particle Model (SPM), a first-principles model describing lithium-ion diffusion within spherical electrode particles via Fick's Law of Diffusion. The terminal voltage is calculated based on surface concentrations and over-potentials. To address real-world variability and lack of precise material data, the model introduces scaling factors ($\eta$) for key parameters:

• Diffusion coefficients ($D_p, D_n$) for positive (cathode) and negative (anode) electrodes.
• Geometrical composite coefficient ($G_p$).
• Maximum lithium concentration ($c_{max,p}$).

B. Offline Meta-Learning (PINEAPPLE-pretrain)

To overcome the slow convergence and poor generalization of standard PINNs, the authors employ a Baldwinian Neuro-Evolution strategy to train a Lithium-ion Electrochemical PINN (LE-PINN):

• Architecture: A shallow network with a single hidden layer using smooth activation functions (sin, silu, tanh).
• Training Process: An outer evolutionary loop optimizes the distribution of weights for the nonlinear hidden layers. An inner loop performs physics-informed fine-tuning of only the final linear output layer using a Tikhonov regularization (Pseudo-Inverse) approach.
• Advantage: This allows the model to learn a "meta-representation" of the SPM physics. Once pre-trained, the LE-PINN can perform zero-shot predictions for new physical conditions (different diffusion coefficients, etc.) by solving a simple linear system, achieving an order-of-magnitude speed-up (10x) over conventional numerical solvers like PyBaMM while maintaining high accuracy (<0.1% error).

C. Online Inference (PINEAPPLE-inference)

The inference phase estimates the cycle-dependent scaling factors ($\eta$) from real-world V-t curves:

• Optimization: The problem is formulated as an inverse search to minimize the squared error between the measured V-t curve and the LE-PINN predicted V-t curve.
• Algorithm: A Covariance Matrix Adaptation Evolution Strategy (CMA-ES) is used. Being gradient-free and population-based, it effectively handles the ill-posed nature of the problem, avoiding local minima.
• Speed: The entire inference process (50 generations, population size 20) takes less than 5 seconds on a standard GPU.

3. Key Contributions

1. Novel Framework: Introduction of PINEAPPLE, combining meta-learned PINNs with evolutionary search for non-destructive, real-time inference of internal electrochemical parameters.
2. Baldwinian Meta-Learning: Development of a highly generalizable LE-PINN that achieves zero-shot, physics-consistent predictions of electrode concentration dynamics with a 10x speed-up over PyBaMM.
3. Robust Inverse Inference: Formulation of the parameter inference problem as an evolutionary search over physically meaningful scaling factors, eliminating the need for battery-specific calibration.
4. Real-World Validation: Demonstration of the framework's ability to recover consistent, physically plausible degradation trends (e.g., diffusion coefficient evolution) using open-source CALCE dataset (LCO-Graphite batteries) without prior knowledge of electrode materials.

4. Key Results

• Prediction Accuracy: The LE-PINN achieved relative errors of $2.78 \times 10^{-4}$ (positive electrode) and $1.29 \times 10^{-3}$ (negative electrode) on out-of-sample test tasks, comparable to numerical solvers with fine mesh resolutions.
• Computational Efficiency: LE-PINN inference time is ~5.6 ms per case, compared to ~57 ms for PyBaMM (on CPU), representing a 10x acceleration.
• Parameter Inference: Applied to four CALCE battery cells (CX2-34, 36, 37, 38), the framework successfully tracked the evolution of scaling factors over hundreds of cycles.
  • Anode ($\eta_{Dn}$): Showed a consistent, monotonic decline, correlating with the known mechanism of Solid Electrolyte Interphase (SEI) growth hindering Li-ion transport.
  • Cathode ($\eta_{Dp}$): Exhibited a three-phase evolution (stable, roll-off, saturation), suggesting complex, multi-stage degradation mechanisms.
  • Correlation: The inferred parameters showed strong correlation with discharge capacity fade, particularly $\eta_{Dn}$ and maximum concentration $\eta_{cmax,p}$.
• Sensitivity Analysis: The study revealed that the identifiability of parameters is state-dependent. As the battery degrades (lower diffusion coefficients), the voltage curve becomes less sensitive to positive electrode parameters, leading to higher variance in inferred values for later cycles.

5. Significance and Future Outlook

• Physics-Aware BMS: PINEAPPLE offers a pathway to "physics-aware" Battery Management Systems (BMS) that provide not just State of Health (SoH) estimates, but mechanistic insights into why a battery is degrading (e.g., distinguishing between anode transport loss and cathode structural failure).
• Scalability: The framework is model-agnostic. While currently using SPM, it can be scaled to more complex models (P2D, DFN) or different chemistries (e.g., LFP) by retraining the meta-learner.
• Real-Time Applicability: The sub-second inference speed makes it viable for onboard implementation, enabling adaptive control and individual cell-level diagnostics in battery packs.
• Beyond Batteries: The core methodology (PINNs + Evolutionary Optimization) is applicable to other complex physical systems where internal states are unobservable (e.g., fuel cells, structural health monitoring).

Limitations: The current study is a proof-of-concept using LCO-Graphite batteries and the SPM. It does not explicitly model electrolyte dynamics, thermal effects, or SEI resistance, which are absorbed into the "effective" inferred parameters. Future work aims to incorporate these complexities and validate the framework on more diverse, stochastic load profiles.