Physics-Informed Shuffle Reservoir Spiking Neural Network for Battery SOH Estimation

This repository contains a physics-informed shuffle reservoir spiking neural network (PI-SRSNN) framework for lithium-ion battery state-of-health (SOH) estimation. The project combines event-driven spiking reservoir computing, grouped channel-shuffled recurrent connectivity, physics-guided battery degradation features, and lightweight ridge-regression readouts to estimate relative capacity across heterogeneous battery datasets.

The goal is to keep SOH prediction accurate while reducing training complexity compared with fully trainable recurrent deep networks such as LSTM, GRU, and CNN-LSTM models.

Highlights

• Physics-informed SOH estimation using impedance growth, square-root-time aging, Arrhenius effective age, Nernst/OCV consistency, and physically plausible SOH clipping.
• Shuffle reservoir spiking core with leaky integrate-and-fire neurons, grouped block-diagonal recurrent connectivity, and deterministic channel shuffling.
• Lightweight training because the recurrent reservoir is fixed and only the readout is trained.
• Cross-dataset evaluation on NASA PCoE, CALCE cylindrical SP cells, and LG MJ1 dynamic degradation data.
• Sub-percent prediction error with an overall row-level RMSE of 0.370% and a balanced-dataset RMSE of 0.345% across 1,363 held-out samples from 29 cells/groups.

Reported Results

The latest saved evaluation artifacts report the following held-out performance:

Dataset	Evaluation protocol	Test samples/groups	RMSE (%)	MAE (%)
NASA PCoE	Leave-one-battery-out	636 rows / 4 batteries	0.536	0.443
CALCE SP cells	Hold-out battery	14 rows / 2 cells	0.468	0.216
LG MJ1	70/30 cell-level split	713 rows / 23 cells	0.031	0.025
Overall row-level	Pooled NASA + CALCE + LG	1,363 rows / 29 groups	0.370	0.222
Balanced dataset	Mean of dataset-level metrics	NASA + CALCE + LG equally weighted	0.345	0.228

These values are stored in:

• grandpipeline_summary.csv
• grandpipeline_crossdataset_summary.csv
• final_tune_best_config.json
• nasa_holdout_summary.csv
• calce_holdout_SP20_1_summary.csv
• calce_holdout_SP20_3_summary.csv
• lg_train70_summary.csv

Method Overview

The proposed PI-SRSNN pipeline has four main components.

1. Input Representation

Each battery record is converted into:

• a fixed-length temporal sequence of voltage, current, temperature or power;
• summary features such as duration, throughput, voltage/current statistics, temperature behavior, impedance indicators, and drift from earlier operating states.

2. Shuffle Reservoir Spiking Core

The temporal sequence is processed by a leaky integrate-and-fire reservoir. Instead of a fully dense recurrent matrix, the model uses grouped block-diagonal recurrence. A fixed channel-shuffle permutation mixes spike activity between groups at every recurrent update. This keeps the reservoir expressive while lowering recurrent synaptic cost.

3. Physics-Informed Guidance

Physics-guided features are concatenated with the reservoir representation. The implemented priors include:

• impedance/ECM SOH prior for NASA;
• square-root-time degradation prior;
• Arrhenius-weighted effective age for temperature-sensitive CALCE data;
• Nernst/OCV-consistent voltage-capacity estimate;
• final SOH range clipping and Coulomb-consistency style post-processing.

4. Lightweight Readout

The readout is a ridge-regression ensemble trained on standardized reservoir, summary, and physics features. Since the recurrent spiking reservoir is fixed, training is much faster and simpler than backpropagation through time.

Repository Structure

.
├── nasa_rsnn_pipeline.py          # Physics-informed shuffle RSNN pipeline for NASA PCoE
├── calce_rsnn_pipeline.py         # Physics-informed shuffle RSNN pipeline for CALCE SP cells
├── lg_rsnn_pipeline.py            # Physics-informed shuffle RSNN pipeline for LG MJ1
├── physics_layer.py               # Battery physics priors and output constraints
├── physics_readout.py             # Optional physics-loss linear readout
├── grandpipeline.py               # Cross-dataset aggregation and final metrics
├── apply_final_tune.py            # Regenerates tuned summaries and final grand metrics
├── final_tune_best_config.json    # Best saved tuning configuration
├── *_summary.csv                  # Saved evaluation summaries
├── *_predictions.csv              # Saved prediction outputs
└── requirements.txt               # Python dependencies

Raw datasets are intentionally not committed because they are large and should be downloaded from their original sources.

Datasets

This project uses three public battery datasets:

1. NASA PCoE Li-ion battery dataset

   • Cells: B0005, B0006, B0007, B0018
   • Protocol: charge/discharge/impedance cycling
   • Evaluation: leave-one-battery-out

2. CALCE cylindrical SP cell dataset

   • Cells: SP20-1 and SP20-3
   • Protocol: low-current OCV, incremental OCV, and initial-capacity tests at multiple temperatures
   • Evaluation: hold-out cell

3. LG MJ1 dynamic degradation dataset

   • Dynamic and path-dependent drive-cycle profiles
   • Evaluation: 70/30 split at cell level

Installation

Create a Python environment and install the dependencies:

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt

On Linux/macOS:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Reproducing the Saved Metrics

After placing the raw datasets in the expected local paths, run:

python apply_final_tune.py

or run the cross-dataset aggregation from existing artifacts:

python grandpipeline.py

The main output files are:

• grandpipeline_predictions.csv
• grandpipeline_summary.csv
• grandpipeline_crossdataset_summary.csv
• grandpipeline_tuning_summary.csv

Key Result

The best saved configuration is:

{
  "nasa_ridge": 50.0,
  "calce_ridge": 10.0,
  "n_seeds": 10.0,
  "aggregator": "mean",
  "coulomb_tol": 0.08,
  "overall_row_rmse_percent": 0.36984682395168267,
  "balanced_dataset_rmse_percent": 0.34509391900738645
}

Notes on Computation and Energy

The model reduces training cost because the reservoir weights are fixed and only a ridge readout is trained. The grouped recurrent matrix reduces recurrent synaptic count by a factor related to the number of groups, while the channel shuffle preserves cross-group information flow without adding trainable parameters. This makes the architecture attractive for embedded battery-management applications where memory, computation, and switching activity matter.