A Scientific Machine Learning Approach for

Predicting and Forecasting Battery Degradation in

Electric Vehicles

Sharv Murgai Hrishikesh Bhagwat

Monta Vista High School Indian Institute of Technology Madras
arXiv:2410.14347v1 [cs.LG] 18 Oct 2024

Cupertino, CA [email protected]
[email protected]

Raj Abhijit Dandekar Rajat Dandekar Sreedath Panat

Vizuara AI Labs Vizuara AI Labs Vizuara AI Labs
[email protected] [email protected] [email protected]

Abstract

Carbon emissions are rising at an alarming rate, posing a significant threat to

global efforts to mitigate climate change. Electric vehicles (EVs) have emerged
as a promising solution to reduce emissions in the transportation sector, but their
reliance on lithium-ion batteries introduces the critical challenge of battery degra-
dation. Accurate prediction and forecasting of battery degradation over both short
and long time spans are essential for optimizing performance, extending battery
life, and ensuring effective long-term energy management. This directly influ-
ences the reliability, safety, and sustainability of EVs, supporting their widespread
adoption and aligning with key United Nations Sustainable Development Goals
(SDGs). In this paper, we present a novel approach to the prediction and long-term
forecasting of battery degradation using a Scientific Machine Learning (SciML)
framework that leverages Neural Ordinary Differential Equations (NeuralODEs)
and Universal Differential Equations (UDEs). Unlike traditional machine learning
models, which often function as black-box systems, SciML integrates domain
knowledge with neural networks, offering more interpretable and scientifically
grounded solutions for both predicting short-term battery health and forecasting
degradation over extended periods. This hybrid approach captures both known and
unknown degradation dynamics, improving predictive accuracy while reducing data
requirements. We incorporate ground-truth data to inform our models, ensuring that
both the predictions and forecasts reflect practical conditions. The model achieved
a mean squared error of 9.90 with the UDE and 11.55 with the NeuralODE, in
experimental data, a loss of 1.6986 with the UDE, and a mean squared error of 2.49
in the NeuralODE, demonstrating the enhanced precision of our approach. This
integration of data-driven insights with SciML’s strengths in interpretability and
scalability allows for robust battery management. By enhancing battery longevity
and minimizing waste, our approach contributes to the sustainability of energy
systems and accelerates the global transition toward cleaner, more responsible
energy solutions, aligning with the UN’s SDG agenda.

Preprint. Under review.

 1 Introduction
1.1 Background and Motivation

As of 2023, global carbon emissions have surged from 25.5 billion tonnes in 2000 to 37.55 billion
tonnes [1]. Electric vehicles (EVs), which made up 18% of total vehicle sales in 2023 [2], are central
to reducing emissions, representing a market worth $500.48 billion [3]. Lithium-ion batteries power
these EVs but face degradation over time, leading to reduced capacity and efficiency. This decline
affects EV reliability and raises sustainability concerns. Addressing battery degradation is crucial
for improving EV performance, extending battery lifespan, and supporting the broader adoption of
sustainable transport solutions.

1.2 Problem Statement

Battery degradation refers to the irreversible decline in a battery’s ability to store or deliver energy,
characterized by reduced charge capacity and increased internal resistance. This is primarily driven
by two forms of degradation: calendar degradation, which occurs over time, and cycle degradation,
caused by repeated charging and discharging. Together, they impact the battery’s long-term per-
formance. The state of health (SoH) quantifies this degradation as a percentage comparing current
capacity to the original capacity when new. Monitoring SoH is crucial for extending battery life,
especially in electric vehicles (EVs), allowing users to manage batteries more effectively, reduce
replacements, conserve materials like lithium and cobalt, and support recycling and second-life
applications.
However, predicting SoH remains challenging. Current models depend on empirical constants, which
vary by battery chemistry, usage, and conditions, making universal models difficult. Validating
these models requires long-term testing, often taking years. Additionally, the scarcity of long-term
experimental data further complicates predictions, as many datasets are incomplete or short-term.
To address this, we are utilizing high-quality experimental data from Sandia National Laboratories
(SNL) to validate and enhance our model, aiming for more accurate and reliable SoH forecasting
across various battery types.

1.3 Contributions

To accurately predict the SoH of EV batteries, we utilize a SciML [7] framework that integrates
physical models with deep learning techniques, offering more transparency than traditional black-box
methods. In this paper, we introduce a UDE [13] and a NeuralODE [14] framework built to model
and forecast battery SoH with upmost precision. The key objectives of this research are:
• Develop a UDE and NeuralODE framework that integrates ordinary differential equations (ODEs)
with neural networks to model and predict battery health over time.
• Use synthetic ground-truth data and real-world experimental data from Sandia National Laboratories
(SNL) [6] to forecast battery degradation.
• Demonstrate the practical application of UDEs and NeuralODEs for long-term battery performance,
focusing on efficiency and sustainability.
• Show how this approach supports global sustainability goals, aligning with the United Nations
Sustainable Development Goals (SDGs) such as SDG 7, SDG 11, SDG 12, and SDG 13.

2 Literature Review
Wang et al. [19] introduce a physics-informed neural network (PINN) to estimate the state-of-health
(SoH) of lithium-ion batteries by combining empirical models with experimental data. While effective
for cycle degradation, it overlooks calendar degradation and lacks Neural ODEs or UDEs, limiting its
ability to generalize across battery chemistries. Hofmann et al. [20] also develop a PINN, using data
from labs, simulations, and vehicles to predict SoH, but similarly focus on cycle degradation and fail
to incorporate physical laws with Neural ODEs or UDEs.
Ye et al. [21] propose a Physics-Informed Neural Network (PIFNN) that integrates battery properties
to improve SoH accuracy, but its reliance on specific datasets reduces generalizability. The absence of

2
 Figure 1: A schematic diagram showing the internal structure of a lithium ion battery

Neural ODEs or UDEs weakens its ability to handle complex degradation processes, particularly with
noisy data. Anonymous [4] employ a Scientific Machine Learning (SciML) framework with Universal
Differential Equations (UDEs) to predict battery degradation, but lack of real-world validation remains
a limitation.

3 System Model and Methodology

3.1 The chemistry behind battery degradation

To model lithium-ion battery degradation effectively, it’s important to understand the chemical
processes involved. As shown in Figure 1, lithium-ion batteries consist of copper or aluminum current
collectors, a graphite anode, and a 2,2,5,5-tetramethyloxolane (TMO) cathode. The two main forms
of degradation are calendar degradation, which occurs over time, and cycle degradation, caused by
repeated charge-discharge cycles. Calendar degradation involves the growth of the solid electrolyte
interphase (SEI) layer on the anode, consuming lithium ions and reducing capacity, even when the
battery is not in use. Cycle degradation occurs during charging and discharging, as lithium ions move
between the anode and cathode, damaging the SEI layer and causing mechanical stress, cracks, and
structural damage. This increases internal resistance and reduces capacity. High temperatures and
voltage fluctuations accelerate these processes, while external factors like graphite exfoliation, SEI
decomposition, and binder degradation further contribute to battery deterioration.

3.2 Modeling Battery Degradation: ODEs

In this research, we have used the typical lithium-ion Nickel-Manganese-Cobalt (NMC) battery that
is used in EVs, particularly a Nissan Leaf to accurately predict and forecast SoH.

3.2.1 Calendar Degradation

Calendar degradation occurs over time, even when the battery is not in use. It depends on the state of
charge (SoC) [24] of a battery, which represents the ratio of the current available charge (or energy)
within a battery to its total charge capacity when fully charged, the time in days (t), the activation
energy (Ea ), which is the minimum amount of energy required to initiate a chemical reaction by
enabling reactants to overcome the energy barrier and transition into products. Additionally, it is
dependent on the internal temperature of the battery. The ODE governing calendar degradation is an
Arrhenius formulation that quantifies the impact of temperature. In addition to these parameters, the
equation uses a pre-exponential factor f , that is a piece-wise function of SoC as described in (1), and
the values are outlined in Table 1. In this paper, we have used a base calendar degradation model
from the Wang model [11].

2

−1.04 · SoC (t) + 89.72 · SoC(t) + 1224.6 if SoC(t) ≤ 50

f (SoC) = 10.35 · SoC 2 (t) − 1083.6 · SoC(t) + 31447 if 50 ≤ SoC(t) ≤ 70 (1)

2.64 · SoC 2 (t) − 409.55 · SoC(t) + 22035 if 70 ≤ SoC(t)

Z t
f (SoC) 1 −Ea
qcal (t) = · √ · exp R·Tb dt (2)
0 2 t

3
 SoC (%) 0 10 20 30 40 50
f 1224.6 2017.8 2603.0 2980.2 3149.4 3110.6
SoC (%) 60 70 80 90 100 -
f 3691.0 6310.0 6167.0 6559.5 7480.0 -
Table 1: Values Corresponding to SoC (%) and the pre-exponential factor f

Parameter Description Value

f (SoC) Pre-exponential factor Describes in equation 1
t Time (days) Battery life so far (duration of usage)
Ea Activation energy 24500.0 J · mol−1
Tb Battery temperature 25° C (298.0K)
R Gas constant 8.314 J · mol−1 · K−1
Table 2: Parameters and descriptions for calendar degradation

Using the values of the pre-exponential function in Table 1, we can get the final ODE modeling
calendar degradation as demonstrated in (2).
where each of the parameters and variables are described in Table 2.

3.2.2 Cycle Degradation

Cycle degradation occurs over time due to charge and discharge cycles, and is dependent on discharge
current (Ib ), battery temperature (Tb ), and the battery capacity (Q). In this paper, we have used the
Wang model [11] to accurately model the cycle degradation of the specific NMC lithium-ion battery
used in Nissan Leafs. The parameters and variables are described in Table 3. In order to calculate the
discharge current of the battery, we can use (4) with the parameters as shown in Table 4 [10].

Z
Ib
 
d·Tb +e
·Ib
a · Tb2 + b · Tb + c · exp


qcycl = Q
·dt (3)
Q · 3600
1 Odo · η
Ib = · (4)
Vnom ∆tdriving

In the absence of real-world testing conditions and to simplify the model in the early stages of
our research, we made several assumptions regarding average driving time and mileage. These
assumptions were based on data from multiple studies conducted in China [26], which indicate that
the typical annual driving distance is approximately 20,000 kilometers. Using (3) and (4), we can
calculate cycle degradation.

Parameter Description Value

2
a Empirical coefficient in 1/AhK 8.61 × 10−6
b Empirical coefficient in 1/AhK −5.13 × 10−3
c Empirical coefficient in 1/Ah 7.63 × 10−1
d Empirical coefficient in 1/K(C-rate) −6.7 × 10−3
Tb Battery temperature 25° C (298.0K)
e Empirical coefficient in 1/(C-rate) 2.35
Ib Discharge current in Amperes Eq 4
Q Battery capacity in Amp-Hours (Ah) 176.4 Ah
Table 3: Parameters and their descriptions for cycle degradation

4
 Parameter Description Value
Vnom Battery nominal voltage 350.4 V
Odo Odometer readings Average of 60kms/day
η Driving specific energy consumed 180 Wh/km
∆tdriving Driving time Average of 2hrs
Table 4: Parameters and their descriptions for discharge current

3.2.3 Cumulative Degradation

Based on (1), (2), (3) and (4), we can calculate the cumulative degradation, capacity and the SoH of
the battery using (5), (6) and (7):
qtotal = qcycl + qcal (5)
SoH = 100% − qtotal (6)
Q = Qnom · (100% − qtotal ), (7)

where Qnom is the maximum capacity of the battery when it was not used when t = 0, around 176.4
Ah.

Figure 2: (a) Simulation of SoH over a timespan of 10 years; (b) and with gaussian noise; (c)
Experimental data

After running simulations under these conditions using the DifferentialEquations.jl [27] package in
Julia, as outlined in previous sections, the results indicate a total degradation of approximately 20%,
as illustrated in Figure 2(a), which aligns with real-world expectations. At an optimal temperature of
25°C, cycle degradation contributes only about 1.5% to the overall degradation, with the majority
attributed to calendar effects.

3.3 Prediction of SoH: synthetic ground-truth data

In the initial stage of the project, we lacked experimental ground-truth data, so we generated synthetic
data using the ground-truth equations (2), (3), (4), and (5). Gaussian noise with an amplitude of 0.20
was added to the cumulative degradation (Figure 2(a)) to simulate real-world conditions, resulting
in the data shown in Figure 2(b). This noisy dataset serves as a benchmark for predictions made
by the UDE and NeuralODE models. We now explore SciML frameworks, including UDEs and
NeuralODEs, to forecast battery degradation over time.

• UDE: A UDE combines traditional mathematical models with neural networks, blending physics-
based knowledge with data-driven methods. This hybrid approach allows UDEs to capture complex
system dynamics, effectively modeling systems with both known physical laws and data-dependent
behaviors.
For the UDE, we chose to replace the √1t term from (2), and the a · Tb2 + b · Tb + c ·


 
d·Tb +e
b Ib ·I
exp Q · Q·3600 , from (3), or the cycle degradation expression with neural networks.
The resultant equation is described in (8), where NN1 (t) is a neural network that approxi-
mates √1t and NN2 (Tb , Ib , Q) is a neural network that approximates the a · Tb2 + b · Tb + c ·


 
d·Tb +e
·Ib Ib
exp Q
· Q·3600 term.

5
 dqtotal f (SoC)
 
−Ea
= · NN1 (t) · exp R·Tb
+NN2 (Tb , Ib , Q) (8)
dt 2

The √1t term was replaced with a neural network to capture the time dependency of calendar
degradation more accurately, which is crucial for understanding long-term battery performance. This
allows the UDE to learn the time-dependent behavior, enhancing the SciML framework. Similarly,
the cycle degradation term from equation (3) was replaced with a neural network, reducing reliance
on empirical constants and enabling the model to learn complex patterns directly from data, making
the degradation process more flexible and precise. Training the UDE on synthetic ground-truth data
was done in two stages. Each neural network had one input neuron and three hidden layers (10,
5, and 5 neurons) with ReLU activation. In the second stage, batch normalization was applied to
each hidden layer for faster and more stable training, and a dropout layer (10%) was used to reduce
overfitting. These steps improved the overall performance of the UDE.

• NeuralODE: A NeuralODE uses neural networks to learn continuous-time dynamics by solving

ODEs. Unlike discrete layers, it models changes over time, making it ideal for time-series prediction
and system dynamics. The neural network defines the system’s derivative, which is integrated over
time by an ODE solver for flexible, efficient modeling.

dqtotal
= f (h(t), t, θ) (9)
dt

where qtotal represents the total state of the system, dqtotal

dt is the time derivative of the system state,
h(t) is the hidden state of the system at time t, and f (h(t), t, θ) is a neural network parameterized
by θ, which captures the system dynamics. The NeuralODE for predicting battery degradation,
implemented in Julia’s Flux.jl [28], features an input layer, three hidden layers (32, 64, and 32
neurons with ReLU activation), and an output layer. Key inputs include time, temperature, SoC,
current, capacity, voltage, and cycle count. Batch normalization ensures stability, while dropout (10-
20%) mitigates overfitting. The output layer predicts the time derivative of total degradation dqtotal
dt ,
which is integrated using an ODE solver to compute cumulative degradation. This architecture,
combined with DiffEqFlux.jl [29], effectively models battery degradation dynamics.

3.4 Prediction of SoH: Using real-world experimental data

3.4.1 Dataset

The dataset from Sandia National Labs [5] is a key resource for studying lithium-ion NMC battery
degradation. It includes experimental data on cycle count, capacity, voltage, current, and temperatures.
NMC cells (3Ah capacity) were cycled to end of life (80% SOH) at various temperatures (15 °C, 25
°C, 35 °C), depths of discharge (0–100%, 20–80%, 40–60%), and discharge rates (0.5C, 1C, 2C, 3C).
After basic preprocessing to remove outliers and missing values, we generated a plot (Figure 2(c))
showing capacity degradation over time under these conditions.

4 Experiments and Results

4.1 Training

For experiments on both the NeuralODE and UDE, hyperparameter tuning was performed, testing
various timespans, optimizers, and learning rates. Optimizers like Adam [30], Nesterov [31], Ad-
aBelief [32], AdaGrad [33], Sophia [34], and RMSProp [35] were evaluated with learning rates of
0.1, 0.01, 0.001, and 0.0001, over 10,000 to 80,000 iterations, in increments of 10,000. The UDE and
NeuralODE was trained on timespans from 1 to 10 years, and the MSE was recorded for each setup
to determine the best configuration. To evaluate the preformances of the models, MSE was used.

6
 TimeSpan (days) Optimizer StepSize Iterations Loss
(1.0, 365.0) Nesterov 0.01 80000 4.172145
(1.0, 730.0) RMSProp 0.01 80000 4.874884
(1.0, 1095.0) ADAM 0.01 60000 6.640595
(1.0, 1460.0) ADAM 0.01 50000 55.925221
(1.0, 1825.0) Sophia 0.01 80000 27.453747
(1.0, 2190.0) AdaBelief 0.01 60000 9.901479
(1.0, 2555.0) ADAM 0.01 60000 24.132372
(1.0, 2920.0) AdaGrad 0.01 40000 27.752057
(1.0, 3285.0) ADAM 0.01 70000 38.811691
(1.0, 3650.0) ADAM 0.01 50000 32.714843
Table 6: Optimization Results for UDE training with batch normalization and dropout layers (Experi-
ment 2)

1
Figure 3: (a) UDE prediction over a 6-year timespan using ADAM; (b) Comparison of √
t
term to
the UDE term

4.1.1 Synthetic ground-truth data: UDE

TimeSpan* (days) Optimizer StepSize Iterations Loss

(1.0, 365.0) RMSProp 0.0001 10000 12.321524
(1.0, 730.0) ADAM 0.1 50000 27.327161
(1.0, 1095.0) Sophia 0.01 80000 46.017585
(1.0, 1460.0) AdaBelief 0.01 50000 65.239215
(1.0, 1825.0) Nesterov 0.01 70000 81.351090
(1.0, 2190.0) ADAM 0.01 80000 101.366405
(1.0, 2555.0) ADAM 0.01 30000 112.077484
(1.0, 2920.0) AdaGrad 0.01 70000 134.881377
(1.0, 3285.0) Sophia 0.001 50000 151.313181
(1.0, 3650.0) Nesterov 0.001 70000 174.597625
Table 5: Optimization Results for UDE training (Experiment 1). *TimeSpan is from day 1 to specified
end day

In the first experiment (Table 5), without dropout or batch normalization, high losses occurred for
timespans over 5 years and lacked accuracy for shorter spans. A second experiment introduced batch
normalization and dropout, reducing overfitting and improving MSE across all timespans (Table 6).
The UDE achieved an MSE of 9.90 over 6 years, closely matching ground-truth data (Figure 3(a)).
Finally, we compared the UDE estimation with the term (Figure 3(b)), confirming that the UDE
accurately captures the time-dependent degradation of the battery.

4.1.2 Synthetic ground-truth data: NeuralODE

In order to evaluate the performance of the NeuralODE, we also use MSE across all the iterations.
Only one experiment was conducted, which gave a MSE of 11.55 over a 7-year timespan, as
demonstrated in Figure 4. The results of the experiment are shown in Table 7. Batch normalization
and dropout layers were included from the start. There were lower losses over smaller timepsans.

7
 Figure 4: NeuralODE prediction over 7-year timespan using ADAM.

TimeSpan (days) Optimizer StepSize Iterations Loss

(1.0, 365.0) Nesterov 0.01 80000 5.23
(1.0, 730.0) AdaBelief 0.001 70000 8.79
(1.0, 1095.0) Adam 0.001 60000 6.51
(1.0, 1460.0) RMSProp 0.01 50000 12.77
(1.0, 1825.0) Sophia 0.01 80000 19.87
(1.0, 2190.0) RMSProp 0.001 60000 22.32
(1.0, 2555.0) Adam 0.001 70000 27.98
(1.0, 2555.0) AdaBelief 0.01 60000 11.55
(1.0, 2920.0) AdaGrad 0.01 40000 30.12
(1.0, 3285.0) RMSProp 0.001 80000 35.78
(1.0, 3650.0) ADAM 0.01 50000 27.71
Table 7: Optimization Results for NeuralODE training on synthetic data with batch normalization
and dropout layers

However, they are not enough to perform adequate forecasting, and nondeterministic to make
conclusions.
4.1.3 Experimental data: UDE
Using the pre-processed dataset, we performed an 80-20 train-test split, with the training set com-
prising 45,769 records. Adam provided the best performance with a loss of 1.6980, while other
optimizer results are shown in Table 8. The UDE model accurately captured time-dependent and cycle
degradation dynamics, as shown in Figure 5(a). This prediction improves battery health monitoring,
enabling earlier decisions on repurposing, recycling, or extending battery life, reducing the need for
frequent replacements. By minimizing resource extraction and waste, it supports global sustainability
efforts, aligning with the UN’s 9th, 11th, 12th, and 13th SDGs.

4.1.4 Experimental data: NeuralODE

Using the preprocessed dataset, we conducted an 80-20 train-test split to predict battery degradation
over the final 20% of the data. The training set comprised 45,769 records. After testing multiple

Figure 5: (a) UDE prediction over experimental data using ADAM; (b) NeuralODE prediction over
experimental data using ADAM and DOPRI5 solver.

8
 Optimizer (UDE) Loss Optimizer (NeuralODE) Loss
Nesterov 7.1243 Nesterov 20.10
RMSProp 23.6789 RMSProp 14.219
Adam 1.6980 Adam 2.1
AdaGrad 14.5231 AdaGrad 10.14
AdaBelief 11.9874 AdaBelief 9.10
Sophia 6.2456 Sophia 8.20
Table 8: Optimizer vs Loss Comparison for UDE Table 9: Optimizer vs Loss Comparison for Neu-
training on Experimental Data ralODE training on Experimental Data

Figure 6: (a) UDE forecast on 20% of experimental data over full timespan; (b) Forecasting of the
UDE over 10 years on synthetic ground-truth data

optimizers and the DOPRI solver, Adam demonstrated the best performance with training MSE loss
of 2.49 as shown in Figure 5(b), while the results of other optimizers are summarized in Table 9.

4.2 Battery Degradation Forecasting

The UDE, trained on synthetic data over 6 years, achieved an MSE of 9.90, while the NeuralODE
reached 11.55 (Figure 3(a)). Forecasting over 10 years is critical for sustainable EV battery man-
agement. On experimental data, the UDE achieved an MSE of 1.6980, closely matching real-world
SoH values (Figure 6(a)), while the NeuralODE had an MSE of 2.1 (Figure 7). These long-term
predictions enhance battery life, support timely recycling, and reduce resource demand, aligning with
the UN’s 9th, 11th, 12th, and 13th sustainability goals. The UDE’s forecast of 80% SoH after 10
years (Figure 6(b)) highlights its capability for sustainable battery management.

5 Conclusion
In this work, we explained the chemical processes behind calendar and cycle degradation, including
SEI growth, particle cracking, and anode dissolution. We employed Scientific Machine Learning
(SciML) to predict and forecast battery degradation using NeuralODE and UDE models, achieving

Figure 7: NeuralODE forecast on 20% of experimental data over full timespan

9
MSE losses of 9.90 and 11.55 on synthetic data, and 1.6980 and 2.10 on real-world data, respectively.
SciML integrates machine learning with physical models, offering more accurate and interpretable
predictions by capturing complex battery dynamics. Unlike traditional black-box methods, which of-
ten overlook the physical processes driving battery degradation, SciML leverages domain knowledge
to improve generalization, making it an essential tool for battery health forecasting and advancing
sustainable energy storage solutions.

References
[1] Statista. "Global CO2 Emissions by Year 1940-2023." Statista, 2023. Available at: https://www.
statista.com/statistics/276629/global-co2-emissions/.
[2] International Energy Agency (IEA). "Electric Vehicles." IEA, 2024. Available at: https://www.iea.
org/energy-system/transport/electric-vehicles.
[3] Statista. "Electric Vehicles - Worldwide | Statista Market Forecast." Statista, 2024. Available at: https:
//www.statista.com/outlook/mmo/electric-vehicles/worldwide.
[4] Sharv M., et al, "Modeling and Forecasting Battery Degradation Using Scientific Machine Learning for
Sustainability." Anonymous, Supplementary Material, 2024.
[5] Preger, Y., et al. "Degradation of Commercial Lithium-Ion Cells as a Function of Chemistry and Cycling
Conditions." Journal of the Electrochemical Society, vol. 167, no. 12, IOP Publishing, Sept. 2020, p.
120532. https://doi.org/10.1149/1945-7111/abae37.
[6] BatteryArchive.org. "Battery Archive." BatteryArchive.org, 2024. Available at: https://www.
batteryarchive.org/snl_study.html.
[7] SciML. "SciML: Open Source Software for Scientific Machine Learning." SciML, 2022. Available at:
https://sciml.ai/.
[8] Wikipedia Contributors. "State of Health." Wikipedia, Wikimedia Foundation, 31 July 2024. https:
//en.wikipedia.org/wiki/State_of_health.
[9] Edge, J. S., et al. "Lithium Ion Battery Degradation: What You Need to Know." Physical Chemistry
Chemical Physics, vol. 23, no. 14, Royal Society of Chemistry, Jan. 2021, pp. 8200–21. https://doi.
org/10.1039/d1cp00359c.
[10] Calearo, M., and L. Engelhardt. "A Simplified Electric Vehicle Battery Degradation Model Vali-
dated with the Nissan LEAF E-plus 62-KWh." 2024, p. 2023. https://backend.orbit.dtu.dk/
ws/portalfiles/portal/339472552/EVTeC2023_-_simplified_degradation_model_full_
paper_-_Finale.pdf.
[11] Wang, J., et al. "Degradation of Lithium Ion Batteries Employing Graphite Negatives and
Nickel–Cobalt–Manganese Oxide + Spinel Manganese Oxide Positives: Part 1, Aging Mechanisms
and Life Estimation." Journal of Power Sources, vol. 269, Elsevier BV, Dec. 2014, pp. 937–48.
https://doi.org/10.1016/j.jpowsour.2014.07.030.
[12] "Sandia National Laboratories: Exceptional Service in the National Interest." Sandia National Laboratories,
2024. https://www.sandia.gov/.
[13] Rackauckas, C., et al. "Universal Differential Equations for Scientific Machine Learning." ArXiv.org, 2020.
https://arxiv.org/abs/2001.04385.
[14] Chen, R. T. Q., et al. "Neural Ordinary Differential Equations." ArXiv.org, 2018. https://arxiv.org/
abs/1806.07366.
[15] "Goal 7 | Department of Economic and Social Affairs." United Nations, 2023. https://sdgs.un.org/
goals/goal7.
[16] "Goal 11 | Department of Economic and Social Affairs." United Nations, 2023. https://sdgs.un.org/
goals/goal11.
[17] "Goal 12 | Department of Economic and Social Affairs." United Nations, 2023. https://sdgs.un.org/
goals/goal12.
[18] "Goal 13 | Department of Economic and Social Affairs." United Nations, 2023. https://sdgs.un.org/
goals/goal13.

10
[19] Wang, F., et al. "Physics-Informed Neural Network for Lithium-Ion Battery Degradation Stable Modeling
and Prognosis." Nature Communications, vol. 15, no. 1, Nature Portfolio, May 2024. https://doi.org/
10.1038/s41467-024-48779-z.

[20] Hofmann, T., et al. "Physics-Informed Neural Networks for State of Health Estimation in Lithium-Ion
Batteries." Journal of the Electrochemical Society, vol. 170, no. 9, Institute of Physics, Sept. 2023, pp.
090524–24. https://doi.org/10.1149/1945-7111/acf0ef.

[21] Ye, J., et al. "A Method for Estimating the State of Health of Lithium-Ion Batteries Based on Physics-
Informed Neural Network." Energy, vol. 294, Elsevier BV, May 2024, pp. 130828–28. https://doi.
org/10.1016/j.energy.2024.130828.

[22] Navidi, S., et al. "Physics-Informed Machine Learning for Battery Degradation Diagnostics: A Comparison
of State-of-The-Art Methods." Energy Storage Materials, vol. 68, Elsevier BV, Apr. 2024, pp. 103343–43.
https://doi.org/10.1016/j.ensm.2024.103343.

[23] Raissi, M., et al. "Physics-Informed Neural Networks: A Deep Learning Framework for Solving Forward
and Inverse Problems Involving Nonlinear Partial Differential Equations." Journal of Computational
Physics, vol. 378, Elsevier BV, Feb. 2019, pp. 686–707. https://doi.org/10.1016/j.jcp.2018.10.
045.
[24] Waag, W., and D. U. Sauer. "SECONDARY BATTERIES – LEAD–ACID SYSTEMS | State-of-
Charge/Health." Elsevier EBooks, Elsevier BV, Jan. 2009, pp. 793–804. https://doi.org/10.1016/
b978-044452745-5.00149-0.
[25] Ye, J., et al. "A Method for Estimating the State of Health of Lithium-Ion Batteries Based on Physics-
Informed Neural Network." Energy, vol. 294, Elsevier BV, May 2024, pp. 130828–28. https://doi.
org/10.1016/j.energy.2024.130828.

[26] Ou, S., et al. "Intensity and Daily Pattern of Passenger Vehicle Use by Region and Class in China:
Estimation and Implications for Energy Use and Electrification." Mitigation and Adaptation Strategies
for Global Change, vol. 25, no. 3, Springer Science+Business Media, Oct. 2019, pp. 307–27. https:
//doi.org/10.1007/s11027-019-09887-0.

[27] Rackauckas, C., and Q. Nie. "DifferentialEquations.jl – a Performant and Feature-Rich Ecosystem for
Solving Differential Equations in Julia." Journal of Open Research Software, vol. 5, no. 1, Ubiquity Press,
May 2017, pp. 15–15. https://doi.org/10.5334/jors.151.

[28] "Welcome · Flux." Fluxml.ai, 2024. https://fluxml.ai/Flux.jl/stable/.

[29] "DiffEqFlux.jl: High Level Scientific Machine Learning (SciML) Pre-Built Architectures · DiffEqFlux.jl."
Sciml.ai, 2020. https://docs.sciml.ai/DiffEqFlux/stable/.
[30] Kingma, D. P., and J. Ba. "Adam: A Method for Stochastic Optimization." ArXiv.org, 2014. https:
//arxiv.org/abs/1412.6980.

[31] Botev, A., et al. "Nesterov’s Accelerated Gradient and Momentum as Approximations to Regularised
Update Descent." ArXiv.org, 2016. https://arxiv.org/abs/1607.01981.

[32] Zhuang, J., et al. "AdaBelief Optimizer: Adapting Stepsizes by the Belief in Observed Gradients."
ArXiv.org, 2020. https://arxiv.org/abs/2010.07468.

[33] Duchi, J., et al. "Adaptive Subgradient Methods for Online Learning and Stochastic Optimization."
Journal of Machine Learning Research, vol. 12, 2011, pp. 2121–59. http://www.jmlr.org/papers/
volume12/duchi11a/duchi11a.pdf.

[34] Liu, H., et al. "Sophia: A Scalable Stochastic Second-Order Optimizer for Language Model Pre-Training."
ArXiv.org, 2023. https://arxiv.org/abs/2305.14342.

[35] Hinton, G., et al. "Neural Networks for Machine Learning Lecture 6a Overview of Mini-Batch Gradient
Descent." http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf.
[36] Wikipedia Contributors. "Mean Squared Error." Wikipedia, Wikimedia Foundation, 11 June 2024. https:
//en.wikipedia.org/wiki/Mean_squared_error.

11