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

Source: arXiv - 2602.18042v1

Overview

The paper presents PINEAPPLE, a hybrid framework that couples physics‑informed neural networks (PINNs) with an evolutionary search algorithm to infer hidden electrochemical parameters of lithium‑ion battery electrodes from simple voltage‑time discharge data. By embedding the governing physics of ion transport directly into the learning process, the authors achieve ultra‑fast, non‑destructive state estimation that could reshape real‑time battery management.

Key Contributions

• Physics‑informed neuro‑evolution: Introduces a meta‑learned PINN that is jointly optimized with an evolutionary algorithm for rapid parameter inference.
• Zero‑shot accuracy: Demonstrates sub‑0.1 % prediction error on electrode voltage curves without any task‑specific fine‑tuning.
• Order‑of‑magnitude speed‑up: Inference runs ~10× faster than traditional numerical solvers used for the same physics‑based models.
• Robust cycle‑level tracking: Recovers the evolution of key internal parameters (e.g., Li‑ion diffusion coefficients) across many charge‑discharge cycles from the CALCE open‑source battery dataset.
• Interpretability & Generality: Shows consistent parameter trends across different cells without hand‑crafted degradation heuristics, highlighting the regularizing power of physics constraints.

Methodology

1. Physics‑Based Model: The authors start from the Doyle‑Fuller‑Newman (DFN) electrochemical model that describes ion diffusion, reaction kinetics, and charge transport in the electrode.
2. PINN Construction: A neural network is trained to satisfy both the measured voltage‑time data (data loss) and the DFN governing equations (physics loss). This dual‑objective loss forces the network to respect known battery physics.
3. Neuro‑Evolution Search: Instead of gradient‑only training, an evolutionary algorithm (population‑based mutation & selection) explores the space of network weights and hyper‑parameters. Evolutionary search is well‑suited for the highly non‑convex loss landscape introduced by the physics constraints.
4. Meta‑Learning: The PINN is meta‑trained across many discharge curves so that, at inference time, it can predict parameters for a new curve in a single forward pass (“zero‑shot”).
5. Parameter Extraction: Once the PINN reproduces the voltage curve, the underlying physical parameters (diffusion coefficients, reaction rates, etc.) are read out from the network’s internal representation using analytical relationships derived from the DFN model.

Results & Findings

• Prediction Accuracy: Across 30+ discharge tests, the voltage predictions deviate by less than 0.1 % from the ground‑truth DFN solver.
• Inference Speed: Parameter inference completes in milliseconds on a single GPU, compared to seconds‑to‑minutes for conventional inverse‑problem solvers.
• Parameter Trends: The inferred diffusion coefficients consistently decrease with cycle count, matching known degradation patterns, and this trend holds for all cells in the CALCE repository without any cell‑specific tuning.
• Robustness: The evolutionary component prevents the model from getting stuck in local minima, yielding stable results even when the discharge data are noisy or partially missing.

Practical Implications

• Real‑time Battery Management Systems (BMS): BMS can now obtain cell‑level state‑of‑health (SoH) metrics on‑the‑fly, enabling smarter load balancing, predictive maintenance, and extended pack life.
• Manufacturing Quality Control: Rapid, non‑destructive probing of electrode parameters can flag out‑of‑spec cells early in the production line, reducing scrap rates.
• Design of Next‑Gen Chemistries: Researchers can quickly evaluate how new electrode materials affect diffusion and reaction kinetics without costly electrochemical impedance spectroscopy.
• Scalable Fleet Monitoring: Because inference is lightweight, cloud‑based services can ingest voltage streams from thousands of electric‑vehicle batteries and continuously update health dashboards.

Limitations & Future Work

• Model Scope: The current implementation assumes the DFN model’s physics are sufficient; exotic degradation mechanisms (e.g., SEI growth, mechanical fracture) are not explicitly modeled.
• Data Requirements: Accurate inference relies on high‑resolution voltage‑time curves; extremely sparse or heavily filtered data may degrade performance.
• Hardware Dependency: While inference is fast on GPUs, embedded BMS hardware may need optimized inference kernels or model compression.
• Future Directions: Extending PINEAPPLE to incorporate additional physics (thermal effects, aging chemistries), exploring hybrid gradient‑evolution training for even faster convergence, and validating on commercial‑grade battery packs under real‑world driving cycles.

Authors

• Karkulali Pugalenthi
• Jian Cheng Wong
• Qizheng Yang
• Pao-Hsiung Chiu
• My Ha Dao
• Nagarajan Raghavan
• Chinchun Ooi

Paper Information

• arXiv ID: 2602.18042v1
• Categories: cs.CE, cs.NE, physics.comp-ph
• Published: February 20, 2026
• PDF: Download PDF