FlowBatt: Flow Matching for Probabilistic Battery Degradation Prediction
Authors: Hamidreza Eivazi, Eslam Sharaawy, Stefan Wittek, André Hebenbrock, Raphael Ginster, Steffen Blömeke
Organizations: Division Data-Driven Modeling of Mechanical Systems, Institute of Applied Mechanics, Technische Universität Braunschweig, Germany · Institute for Software and Systems Engineering, Clausthal University of Technology, Germany · Institute of Chemical and Electrochemical Process Engineering, Clausthal University of Technology, Germany · Institute of Automotive Management and Industrial Production, Technische Universität Braunschweig, Germany · Institute of Machine Tools and Production Technology, Technische Universität Braunschweig, Germany
Battery degradation remains a central challenge in the development and broad application of sustainable energy technologies. Accurate degradation prediction is challenging, as battery aging emerges from complex and heterogeneous interactions among cycling behavior, operating conditions, and cell chemistry. Existing machine-learning approaches typically focus on deterministic end-of-life predictions or modeling degradation curves within fixed chemistry settings, leaving probabilistic state of health (SOH) trajectory modeling across heterogeneous chemistries underexplored. We introduce FlowBatt, a use-inspired conditional generative framework that adapts flow matching with a diffusion transformer (DiT) backbone to battery degradation trajectory modeling. FlowBatt models full SOH trajectories as probabilistic degradation processes conditioned on early-cycle capacity data. To strengthen this conditioning pathway, we use explainable AI analysis to diagnose and improve the encoder that maps early-cycle capacity matrices into conditioning vectors. We evaluate FlowBatt on five public benchmark settings spanning diverse chemistries and aging conditions, comparing it with supervised and diffusion-based trajectory models as well as established baselines for remaining useful life (RUL) prediction under shared data splits. FlowBatt achieves the best trajectory-wise SOH prediction performance on all five datasets, with errors below 2.5%, and delivers competitive RUL prediction performance. By generating multiple plausible trajectories, FlowBatt provides empirical prediction intervals for battery degradation modeling. These results indicate that flow-matching-based trajectory modeling is a promising framework for probabilistic battery health prediction.
Figures & tables
Figure 1: Schematic view of the model architecture. Adapted and modified from [ 13 ] . Modifications include context-specific changes.
Figure 2: Comparison of denoising trajectories for a test sample from the MATR-1 dataset using flow matching (top) and the diffusion process (bottom). The pink line shows the generated sample and the cyan line shows the reference for a given capacity matrix. The displayed FlowBatt states correspond to requested ODE output times for visualization.
Method
MATR-1
MATR-2
HUST
MIX-20
MIX-100
BatLiNet
1.86 ± 0.06
3.62 ± 0.24
2.87 ± 0.23
2.04 ± 0.06
2.34 ± 0.22
RF (BatteryML)
2.01 ± 0.04
2.50 ± 0.01
3.57 ± 0.03
3.42 ± 0.02
5.45 ± 0.04
MLP (BatteryML)
1.97 ± 0.11
3.06 ± 0.17
4.11 ± 0.13
2.23 ± 0.09
3.13 ± 0.14
Transformer (ours)
1.09 ± 0.07
2.25 ± 0.22
3.18 ± 0.16
1.42 ± 0.10
1.76 ± 0.06
DiffBatt (ours)
0.92 ± 0.04
1.85 ± 0.03
2.44 ± 0.14
1.15 ± 0.08
1.71 ± 0.09
FlowBatt (ours)
0.87 ± 0.02
1.77 ± 0.04
2.10 ± 0.12
0.88 ± 0.03
1.41 ± 0.02
Table 1: Trajectory-wise SOH prediction results reported as mean RMSE (%) over the predicted degradation trajectories. The ± values denote standard deviations of RMSE across ten generative draws from one trained checkpoint for FlowBatt and DiffBatt, and across ten training initializations for the other models. Best results are shown in bold and second-best results are underlined. This table evaluates full-trajectory SOH prediction. RMSE is computed over the predicted degradation trajectories up to EOL and therefore measures a different task from the cycle-wise SOH setting reported in section A.5.3 .
Samples
Metric
MATR-1
MATR-2
HUST
MIX-20
MIX-100
10
SOH coverage
86.0%
67.0%
65.0%
83.6%
81.7%
SOH width
0.30%
0.42%
1.34%
0.52%
0.64%
50
SOH coverage
87.4%
70.0%
71.0%
86.4%
85.4%
SOH width
0.36%
0.50%
1.62%
0.63%
0.77%
100
SOH coverage
88.0%
70.0%
73.0%
87.0%
86.0%
SOH width
0.37%
0.51%
1.65%
0.65%
0.80%
Table 2: Empirical SOH prediction-interval coverage of FlowBatt for a nominal 90% interval. Coverage is the fraction of reference SOH values contained within the generated prediction interval. SOH width denotes the average interval width in percentage points. Best coverage values are shown in bold.
Figure 3: Example SOH prediction intervals obtained from FlowBatt across the five benchmark datasets. The cyan curve denotes the reference SOH trajectory, the magenta dashed curve denotes the median prediction, and the shaded region denotes the empirical 90% prediction interval. The displayed batteries are chosen as the test sample with median predictive uncertainty.
Method
MATR-1
MATR-2
HUST
MIX-100
MIX-20
RMSE
MAPE (%)
RMSE
MAPE (%)
RMSE
MAPE (%)
RMSE
MAPE (%)
RMSE
MAPE (%)
Training Mean
399
28
511
36
420
18
573
59
593
102
“Variance” Model [ 43 ]
138
15
196
12
398
17
521
39
601
95
“Discharge” Model [ 43 ]
86
8
173
11
322
14
1743
47
>2000
>100
“Full” Model [ 43 ]
100
11
214
12
335
14
331
22
441
53
Ridge Regression [ 3 ]
125
13
188
11
1047
36
395
30
806
150
Table 3: Comparison with baseline methods for the RUL prediction task. Best results based on RMSE are shown in bold, and second-best results are underlined. For the models introduced in this study, we report the mean and standard deviation of each error metric across ten generative draws from one trained checkpoint for FlowBatt and DiffBatt, and across ten training initializations for Transformer.
Figure 4: Results obtained from FlowBatt for RUL prediction. Colorbar shows the standard deviation σRUL of the predicted RUL from ten generated SOH trajectories per sample.
w
Metric
RF
0.5
1.0
2.0
3.0
4.0
FID ( ↓ )
NA
0.129
0.001
0.001
0.002
0.003
Precision ( ↑ )
NA
0.97
1.00
1.00
1.00
1.00
Recall ( ↑ )
NA
0.95
0.98
0.59
0.56
0.54
RUL RMSE ( ↓ )
136
119
101
102
104
107
RUL MAPE (%) ( ↓ )
12
11
10
10
10
10
Table 4: Results for the MATR-1 dataset obtained using random forests trained on synthetic data generated with different diffusion guidance w , against a random forest trained without synthetic data following [ 51 ] . Best values per RUL error metrics, determined before rounding, are shown in bold.
Appendix figures & tables20 assets
Supplementary material from the paper’s appendix.
Appendix
Benchmark
Source data
Chemistry
Early cycles
Train/test
MATR-1
MATR-1
LFP/graphite
100
41/42
MATR-2
MATR-2
LFP/graphite
100
41/40
HUST
HUST
LFP/graphite
100
55/22
MIX-20
Combined
LCO, LFP, NCA, NMC, NMC–LCO
20
256/147
MIX-100
Combined
LCO, LFP, NMC, NMC–LCO
100
205/137
Appendix
Table 5: Benchmark composition, conditioning horizon, and final train/test battery counts.
Figure 5: Train (up) and test (bottom) samples for each cell chemistry. The data is scaled using the SOH of the first cycle.
Category
Major hyperparameters
Training
Epochs: 10000
Initial learning rate: 1×10−3
Optimizer: AdamW
Learning rate scheduler: Cosine schedule with 100 warmup steps
Model (DiT)
Sequence length: 256
Capacity matrix shape: (1,100,100) , or (1,20,100) for MIX-20
Appendix
Table 6: Training and model hyperparameters.
Seed
0
1
2
3
4
5
6
7
8
9
Benchmark CNN [ 51 ]
76
67
64
74
60
82
65
79
367
78
CNN (Replicated)
71
77
83
306
79
60
78
62
88
64
XAI-CNN (Ours)
57
54
54
72
66
64
65
60
59
61
Appendix
Table 7: RUL prediction results obtained from CNN models with different initialization seeds for MATR-1 dataset.
Figure 6: (a) One sample of capacity matrix from MATR-1 train dataset. (b) All test samples of MATR-1 dataset.
Figure 7: Grad-CAM heatmaps for CNN model proposed in [ 3 , 51 ] .
Figure 8: Grad-CAM heatmaps for XAI-CNN model.
Category
Major hyperparameters
Evaluation
Cycle-wise SOH
Trajectory-wise SOH
MLP
Input representation: capacity matrix
Input settings: final early-life curve minus baseline curve
Table 8: Training and model settings for the BatteryML baselines used in the SOH experiments.
Method
MATR-1
MATR-2
HUST
MIX-20
MIX-100
RF (final-cycle curve)
2.01 ± 0.04
2.50 ± 0.01
3.57 ± 0.03
3.42 ± 0.02
5.45 ± 0.04
MLP (final-cycle curve)
1.97 ± 0.11
3.06 ± 0.17
4.11 ± 0.13
2.23 ± 0.09
3.13 ± 0.14
RF (full matrix)
2.26 ± 0.02
3.05 ± 0.06
4.20 ± 0.03
1.50 ± 0.01
2.51 ± 0.02
MLP (full matrix)
2.75 ± 0.08
4.04 ± 0.35
4.16 ± 0.23
2.87 ± 0.20
2.91 ± 0.14
Appendix
Table 9: Trajectory-wise SOH RMSE (%) for RF and MLP using the final early-life curve or the complete capacity matrix. Values are mean ± standard deviation over ten training runs. Final-cycle results are reproduced from table 1 .
Epochs: 1000; train batch size: 128; test batch size: 1; train support size: 2; test support size: 32; gradient accumulation steps: 1
Target setup
Original BatLiNet uses RUL labels with log-scale and Z-score transformation. For SOH experiments, only the target label was adapted while keeping the remaining setup unchanged.
Appendix
Table 10: Training and model settings for the BatLiNet baseline used in the SOH experiments.
Model
Published benchmark
Real only
Gaussian jitter
FlowBatt synthetic data
RF
140
136 ± 3
110 ± 1
101 ± 2
MLP
162 ± 7
88 ± 7
85 ± 4
79 ± 4
Appendix
Table 11: RUL RMSE (cycles) for real-only training and data augmentation on MATR-1 . Published RF and MLP benchmark results from Zhang et al. [52] are included for reference. Values include the reported standard deviations where available. The lowest reported RMSE for each downstream model is shown in bold.
Figure 9: MATR-1 synthetic data generated by FlowBatt with different guidance strengths.
Figure 10: SOH predictions obtained from FlowBatt against reference for all the test samples of MIX-100 dataset. The pink dashed line shows the prediction and the cyan solid line shows the reference. The gray area shows the prediction uncertainty computed from ten generated samples.
Figure 11: SOH predictions obtained from DiffBatt against reference for all the test samples of MIX-100 dataset. The pink dashed line shows the prediction and the cyan solid line shows the reference. The gray area shows the prediction uncertainty computed from ten generated samples.
Figure 12: SOH predictions obtained from Transformer model against reference for all the test samples of MIX-100 dataset. The pink dashed line shows the prediction and the cyan solid line shows the reference. The gray area shows the variability across predictions from ten independently trained Transformer models.
Method
MATR-1
MATR-2
HUST
MIX-20
MIX-100
BatLiNet
0.33 ± 0.01
0.18 ± 0.01
0.05 ± 0.03
0.33 ± 0.01
0.27 ± 0.02
RF (BatteryML)
0.88 ± 0.01
1.52 ± 0.03
1.18 ± 0.03
2.19 ± 0.05
2.04 ± 0.07
MLP (BatteryML)
1.01 ± 0.13
1.45 ± 0.30
1.58 ± 0.13
3.59 ± 1.31
3.72 ± 1.49
FlowBatt (ours)
0.07 ± 0.01
0.23 ± 0.01
0.06 ± 0.01
0.89 ± 0.06
0.15 ± 0.01
Appendix
Table 12: Cycle-wise SOH prediction results at cycle 100, reported as RMSE (%) across cells. Best results are shown in bold and second-best results are underlined. This table follows the BatteryML-style cycle-wise SOH protocol, where SOH is predicted at a predefined cycle index and RMSE is computed across cells. These results are therefore not directly comparable to the trajectory-wise SOH prediction results in table 1 .
Method
Cycle
MATR-1
MATR-2
HUST
MIX-20
MIX-100
FlowBatt (ours)
200
0.36 ± 0.03
0.73 ± 0.03
0.18 ± 0.03
0.93 ± 0.02
1.13 ± 0.08
300
0.69 ± 0.03
0.94 ± 0.04
0.21 ± 0.03
1.12 ± 0.04
0.82 ± 0.03
RF (BatteryML)
200
1.25 ± 0.35
1.57 ± 0.05
1.17 ± 0.02
2.30 ± 0.00
3.54 ± 0.87
300
1.48 ± 0.04
1.65 ± 0.03
1.18 ± 0.03
2.25 ± 0.02
1.64 ± 0.04
MLP (BatteryML)
200
1.31 ± 0.11
1.25 ± 0.35
1.52 ± 0.14
3.98 ± 0.88
1.64 ± 0.02
300
1.97 ± 0.28
1.19 ± 0.18
1.51 ± 0.14
3.66 ± 0.55
3.47 ± 0.57
Appendix
Table 13: Extended cycle-wise SOH prediction results at cycles 200 and 300, reported as RMSE (%) across cells.
Method
Metric
MATR-1
MATR-2
HUST
MIX-20
MIX-100
FlowBatt
SOH coverage
86.0%
67.0%
65.0%
83.6%
81.7%
SOH width
0.30%
0.42%
1.34%
0.52%
0.64%
DiffBatt
SOH coverage
76.6%
64.8%
56.5%
79.8%
79.1%
SOH width
0.25%
0.36%
2.52%
0.97%
1.37%
Quantile RF
SOH coverage
93.0%
96.0%
67.3%
91.9%
89.4%
SOH width
3.06%
5.79%
6.10%
3.82%
4.07%
Appendix
Table 14: Comparison of nominal 90% SOH prediction intervals. Coverage is reported in percent and mean width in SOH percentage points. FlowBatt and DiffBatt use ten generated trajectories per test battery. Quantile RF and Quantile MLP results are averaged over ten training seeds.
Figure 13: SOH predictions obtained from Quantile RF against the reference for all test samples of the MIX-100 dataset. The pink dashed line shows the median prediction and the cyan solid line shows the reference. The gray area shows the nominal 90% prediction interval between the 5th and 95th percentiles.
Figure 14: SOH predictions obtained from Quantile MLP against the reference for all test samples of the MIX-100 dataset. The pink dashed line shows the median prediction and the cyan solid line shows the reference. The gray area shows the nominal 90% prediction interval between the 5th and 95th percentiles.
Probabilistic time series forecasting estimates distributions of future values or trajectories to quantify uncertainty and support risk-aware decisions.