DIDBlock: Decoupling Degradation Factors

• DIDBlock is a neural module that decouples statistically, physically, and semantically distinct degradation ingredients across diverse domains.
• It employs parallel subnets and physics-informed loss constraints to isolate thermodynamic, kinetic, and artifact factors for targeted prediction and restoration.
• Empirical evaluations demonstrate enhanced capacity prediction, higher PSNR in multi-degradation image restoration, and robust temporally consistent video recovery.

The Degradation Ingredient Decoupling Block (DIDBlock) is a neural module concept first appearing in multi-degradation learning for batteries, images, and videos, with explicit design for the statistical, physical, and semantic decoupling of disparate degradation factors (“ingredients”) from observed system responses. DIDBlock generalizes the separation of entangled mechanisms—chemical (battery), artifact (vision), or otherwise—into orthogonalized latent representations that enable interpretable, adaptive, and physiochemically or semantically plausible recovery or prediction.

1. Concept and Motivation

DIDBlock was conceived to address the challenge that system-level degradations—whether capacity fade in batteries or coexisting artifacts like haze, flare, and noise in images and video—arise from multiple overlapping factors with distinct causes, temporal evolution, and impacts. In classical ML pipelines, such ingredient mixtures are either modeled holistically or tackled by task-specific branches without enforcing explicit disentanglement, which can obscure underlying mechanisms and limit adaptation or interpretability.

DIDBlock introduces explicit architectural and loss constraints to separate (“decouple”) the statistically or physically distinct components, enabling i) targeted prediction or correction aligned with real-world processes (e.g., thermodynamics/kinetics in batteries, flare/haze in UDC video), and ii) the formation of invariant/orthogonal representations that facilitate transfer learning, adaptation, or restoration (Tao et al., 1 Jun 2024, Gao et al., 7 Nov 2025, Liu et al., 8 Mar 2024).

2. DIDBlock in Physics-Informed Battery Degradation Modeling

In the context of battery health prediction (Tao et al., 1 Jun 2024), DIDBlock is the central module within a physics-informed ML pipeline for non-destructive and temperature-adaptable capacity trajectory prediction.

Architecture: At each cycle $t$, inputs comprise prior-cycling (IMV) features—the cut-off voltages at nine predefined state-of-charge (SOC) levels ($U_1,\ldots,U_9$), intra- and inter-step charge features (capacity, resistance, voltage gradients, and transients), and environmental temperature $T_t$. The module implements:

• Featurization Embedding Layer: Fully connected transform $f_1:\mathbb{R}^N\to\mathbb{R}^{32}$ with LeakyReLU. Inputs $N\sim42$.
• Physics-Constraint Layer: Arrhenius scaling via a learned activation energy $E_a$ modulates the kinetic subnet:

$$\hat\eta_\mathrm{kin} \leftarrow \hat\eta_\mathrm{raw} \exp\left[-\frac{E_a}{k_BT}\right]$$

• Parallel Ingredient Subnets:
  • $$f_\mathrm{thermo}(\cdot):\mathbb{R}^{32}\to\mathbb{R}$$ predicts macroscopic thermodynamic loss $AE_t$,
  • $$f_\mathrm{kinetic}(\cdot):\mathbb{R}^{32}\to\mathbb{R}$$ predicts kinetic loss $\eta_t$,
  • Combined reconstruction: $\Delta U_t = AE_t + \eta_t$.
• Chain-of-Degradation Integration: Updates state $S_{t+1} = S_t + [\Delta U_t, AE_t, \eta_t]$ for next cycle, providing sequential composition of degradation effects.

This block is invoked at each time step to generate longitudinally resolved ingredient trajectories, serving a downstream trajectory model that maps $\{AE_t,\eta_t\}_{t=1}^C$ to the observable capacity $Q(t)$ over the full cell life.

Featurization Taxonomy: Raw charging data are distilled into material-agnostic markers for IMV, intra-step, and inter-step dynamics, ensuring the decomposition is grounded in physically meaningful indicators, not surface statistics.

3. DIDBlock in Vision: Multi-Degradation Image and Video Restoration

3.1. Image Restoration (Gao et al., 7 Nov 2025)

In the IMDNet architecture, DIDBlock operates at each encoder stage to decouple mixed artifacts (rain, haze, noise, etc.) from semantic content:

• Spatial Encoding: Inputs $E_{l-1}$ (feature from prior depth) and $I_l$ (degraded image) are concatenated, reduced via $1\times1$ convolution, and passed through a NAFBlock (depth-wise conv + simple gate).
• Frequency Decomposition: Learnable dynamic filtering splits output into high- and low-frequency branches $(F_H, F_L)$.
• Statistical Coefficient Extraction: For $E_l, F_H, F_L$, compute global average and stddev vectors, pass through channel-wise MLP + simple gate. These are summed and averaged (over six stats) to generate a degradation embedding $DI_l$.
• Clean/Degradation Recoupling:
  • Degradation embedding $DI_l$ modulates $E_l$ ($DI_l = (\mathrm{sum~stats}/6) \otimes E_l$).
  • Clean feature $CF_l = E_l - DI_l$ is formed as the residual, passed via skip-connection.

Training: Orthogonality (cosine similarity) loss, $$L_d(CF_i, DI_i)= \frac{CF_i \cdot DI_i}{\|CF_i\|_2\,\|DI_i\|_2}$$, together with pixel-, edge-, and frequency-domain losses, enforces statistical independence between clean and degradation pathways.

Empirical Gains: Encoder swap to DIDBlock yields $+0.63$ dB PSNR in multi-degradation; full IMDNet pipeline achieves $+4.18$ dB vs. NAFNet baseline; DI embeddings demonstrate superior clustering by degradation type in t-SNE.

3.2. Video Restoration (UDC) (Liu et al., 8 Mar 2024)

In UDC scenarios, DIDBlock (recalling Decoupling Attention Module/DAM) operates along the temporal axis to disentangle lighting artifacts:

• Soft-Mask Generation: Frame $I_t$ is split into “flare” ($M^{\mathrm{flare}}_t$) and “haze” ($M^{\mathrm{haze}}_t = 1 - M^{\mathrm{flare}}_t$) masks, using intensity thresholds.
• Path-Specific Feature Flows: Separate convolutional subnets process long-term features for flare and short-term features for haze, each masked and temporally aligned via optical flow (SPyNet).
• Attention-Gated Refinement: Intermediate restoration supervises partial outputs; attention map $A_t$ further modulates each ingredient pathway.
• Hierarchical Recurrence: Multi-scale (coarse-to-fine) U-Net backbone; DIDBlock recurses at scales $2\times, 4\times, 8\times$.
• Losses: Charbonnier penalties on both intermediate and final images.

This structure addresses the necessity to process spatially and temporally entangled degradations, with each ingredient benefiting from targeted memory and feature learning.

4. Loss Functions and Training Regimes

DIDBlock consistently employs multi-objective loss architectures comprising:

• Data-Fitting: Direct error between prediction and ground truth ($L_\mathrm{data}$).
• Domain-Specific Regularization: For batteries, monotonic smoothness for thermodynamic loss ($L_\mathrm{thermo}$), Arrhenius scaling for kinetics ($L_\mathrm{kinetic}$) (Tao et al., 1 Jun 2024).
• Decoupling/Orthogonality: Covariance minimization (battery: $L_\mathrm{dec}$), or cosine similarity (vision, $L_d$).
• Auxiliary Pathway Losses: Supervision on intermediate outputs (video), or maintenance of clean/degrad artifact independence (image).

Hyperparameters are tuned via cross-validation or grid search, with typical regularization weights $10^{-3}$ to $10^{-1}$ and supervised by early stopping on validation error. Optimizers include Adam with carefully selected learning rates, and in video, cosine annealing scheduling.

5. Empirical Characteristics and Performance

DIDBlock demonstrates high quantitative fidelity and transferability in all examined domains:

Application Area	Performance Highlights	Key Gains
Battery lifetime modeling	95.1% capacity prediction accuracy (MAPE 4.9%) from first 50 cycles;	25× speed-up over full cycle testing
	Cross-temperature domain generalization; Arrhenius adaptation
Image restoration	+4.18 dB PSNR over NAFNet; distinct clustering of degradation types	Strong multi-degradation handling
UDC video restoration	Robust recovery under entangled haze/flare; temporally consistent outputs	Superior to SOTA video/UDC baselines

DIDBlock’s separation of degradation ingredients not only improves restoration or prediction accuracy but also yields interpretable ingredient trajectories (thermo/kinetic loss in batteries, artifact signature embeddings in imaging) that match physical analyses (e.g., FEA, incremental capacity benchmarks for batteries).

6. Implementation Considerations

Modularity: DIDBlock can be cleanly slotted into encoder stages, recurrent units, or processing pipelines, requiring only domain-appropriate featurization and matching channel dimensions.

Physical/Statistical Constraints: For maximal effectiveness, application-specific regularization (e.g., Arrhenius scaling, monotonicity, or orthogonality) should be tuned to reflect the separation of physically plausible ingredient mechanisms.

Computational Cost: The overhead introduced by parallel subnetworks, frequency decomposition, or multi-path processing is marginal relative to baseline architectures but yields substantial gains in representation clarity and restoration adaptability.

Transferability: Stepwise featurization (battery SCA), dynamic filtering (vision), and explicit masking (UDC video) make DIDBlock adaptable across domains given appropriate preprocessing and regularization.

7. Significance and Applications

DIDBlock embodies a cross-domain approach to the decoupling of temporally and spatially entangled degradation processes, serving as a cornerstone in modern interpretable and adaptive system modeling. It underpins:

• Ultra-early verification of battery prototype reliability, supporting rapid manufacturing and recycling decisions (Tao et al., 1 Jun 2024).
• Multi-artifact restoration and adaptation in challenging imaging settings (rain, haze, noise) (Gao et al., 7 Nov 2025).
• Artifact disentanglement for UDC video, achieving temporally stable recovery under device-specific degradations (Liu et al., 8 Mar 2024).

A clear implication is that explicit ingredient decoupling enables not just improved accuracy but interpretable diagnosis, facilitating adaptation to new environments, physical domains, or artifact compositions without reengineering the entire inferential pipeline.