Advancing multi-site emission control: A physics-informed transfer learning framework with mixture of experts for carbon-pollutant synergy

This paper presents a physics-informed transfer learning framework utilizing a carbon-pollutant mixture-of-experts model to enable scalable, regime-aware emission control and risk evaluation across 13 heterogeneous municipal solid waste incineration plants, achieving robust cross-facility transferability without complete re-learning.

The Gist

Imagine a city full of garbage trucks. Instead of just dumping trash in a landfill, these trucks deliver waste to massive incinerators (burning plants) that turn the trash into electricity. This is great for the city, but burning trash creates a messy cocktail of smoke: carbon dioxide, soot, and various toxic gases.

The problem is that every incinerator is different. One might burn wet food waste, another dry paper; one might have a giant furnace, another a smaller one. Because of these differences, a "recipe" for clean burning that works perfectly at Plant A often fails completely at Plant B. It's like trying to use a recipe for a perfect cake in a different kitchen with different ovens and ingredients—the result is usually a disaster.

Scientists have tried to use computers (AI) to predict how much pollution each plant will make. But these computers usually learn by heart only the specific plant they were trained on. If you move them to a new plant, they get confused.

This paper introduces a new, smarter way to teach these computers. Here is how it works, explained simply:

1. The "Expert Panel" (Mixture of Experts)

Instead of training one giant, confused brain to handle every situation, the authors built a team of four specialized "experts" (different types of AI models).

• The Long-Term Thinker: Good at spotting patterns over long periods.
• The Local Observer: Good at spotting quick, immediate changes.
• The Memory Keeper: Good at remembering what happened just a moment ago.
• The Steady Baseline: A simple, reliable predictor.

The system has a "Manager" (a gating network) that looks at what the plant is doing right now. If the plant is in a stable, slow-burning mode, the Manager might ask the "Steady Baseline" to do the work. If the plant is having a chaotic, high-heat moment, the Manager might call on the "Long-Term Thinker." This way, the system uses the right tool for the specific job, rather than trying to force one tool to do everything.

2. The "Physics Rulebook" (Physics-Informed)

Usually, AI learns just by looking at numbers. But numbers can be tricky; they might find fake patterns that don't make sense in the real world.
To fix this, the authors forced the AI to follow the Laws of Physics. They gave the computer a rulebook that says:

• "You can't create energy out of thin air."
• "If you put in more air, the fire changes in a specific way."
• "The amount of smoke coming out must match the amount of trash going in."

By forcing the AI to obey these rules, it learns the true logic of burning, not just the accidental patterns of one specific factory. This makes the AI much more reliable when it moves to a new plant.

3. The "Universal Translator" (Transfer Learning)

Once the AI learned the rules of burning at one "Reference Plant," the team wanted to see if it could understand 12 other plants without starting from scratch.
Think of it like learning to drive. If you learn to drive a car in New York, you can usually drive a car in London, even though the traffic rules and road layouts are different. You don't need to re-learn how to steer or brake; you just need to adjust to the new environment.

• The Result: The AI successfully "transferred" its knowledge. It didn't need to re-learn everything. It just adjusted its "Manager" to pick the right experts for the new plant's specific style of burning.
• The Proof: It predicted pollution levels accurately across all 13 plants, even though they were very different from each other.

4. The "Synergy Score" (CPSI)

Instead of just looking at one type of pollution (like just carbon or just soot), the team created a single "Synergy Score." This score acts like a health grade for the whole plant. It combines the carbon emissions and the toxic pollutants into one number to tell you how risky the plant is overall.
The AI learned to predict this single score very well, meaning it understands the whole picture of the plant's environmental impact, not just isolated parts.

5. The "Digital Twin" (The Map for the Future)

Finally, the authors turned this AI into a Digital Twin. Imagine a video game version of the real incinerator that runs in the computer.

• Because the AI understands the different "modes" of operation (the experts), the Digital Twin can simulate what would happen if the operators changed the air supply or the temperature.
• It acts like a GPS for the operators. Instead of guessing, they can ask the Twin: "If I do this, what happens to our pollution score?" The Twin can then suggest the best path to keep the plant running cleanly and safely.

The Bottom Line

The paper shows that by combining a team of specialized AI experts with the unbreakable rules of physics, we can build a smart system that understands how to burn trash cleanly. This system works not just in one factory, but can be easily adapted to dozens of different factories, helping cities manage waste and pollution more effectively without needing to start over every time.

Technical Summary

1. Problem Definition

Municipal Solid Waste Incineration (MSWI) is critical for waste management but faces significant challenges in controlling carbon emissions and air pollutants (CO, CO₂, NOₓ, SO₂, HCl, PM) simultaneously.

• Heterogeneity: MSWI plants vary widely in waste composition, furnace design, and operating strategies. This leads to distinct emission behaviors across facilities, making models trained on one plant fail when transferred to another.
• Limitations of Current Models: Existing data-driven models are typically:
  • Site-specific: They overfit to local correlations rather than learning underlying physical mechanisms.
  • Pollutant-decoupled: They treat pollutants independently, ignoring the coupled nature of combustion efficiency, temperature, and multi-pollutant formation.
  • Lacking Transferability: They struggle to generalize to new facilities without extensive retraining and lack a unified framework for evaluating system-level risk (carbon vs. pollutant trade-offs).

2. Methodology

The authors propose a Physics-Informed Transfer Learning (PITL) framework built upon a Carbon-Pollutant Mixture-of-Experts (CPMoE) architecture.

A. Carbon-Pollutant Mixture-of-Experts (CPMoE)

Instead of a single global mapping, the model recognizes that MSWI operates under multiple distinct regimes (e.g., drying, pyrolysis, combustion, burnout).

• Architecture:
  • State Recognition: A module using 1D-CNN and BiGRU with attention mechanisms to identify the current process phase.
  • Heterogeneous Experts: Four distinct neural network architectures (Transformer, CNN, LSTM, MLP) act as "experts," each capturing different temporal dependencies (long-range, local, recursive, or baseline).
  • Sparse Gating: A gating network dynamically routes input to the top-3 most relevant experts based on the current operating regime.
  • Multi-Task Output: Six independent heads predict the concentrations of CO, CO₂, NOₓ, SO₂, HCl, and PM.
• Carbon-Pollutant Synergistic Index (CPSI): A composite metric calculated from predicted pollutant concentrations to quantify integrated system-level environmental and health risk.

B. Physics-Informed Constraints

To ensure the model learns transferable physical laws rather than site-specific noise, three conservation constraints are embedded as soft regularization terms in the loss function:

1. Combustion-State Consistency: Aligns excess air ratios calculated from air supply/steam generation with those derived from flue gas oxygen levels.
2. Mass-Balance Constraint: Ensures the total pollutant output load is consistent with the inferred input load (based on elemental composition like Cl, S, N).
3. Energy Balance: Enforces consistency between input energy (fuel/air) and output energy (steam/flue gas heat).

C. Physics-Informed Transfer Learning (PITL)

The framework transfers a pre-trained model from a Source Domain (MSWI#1) to 12 Target Domains (different plants) with limited labeled data.

• Domain Alignment: Uses Maximum Mean Discrepancy (MMD) with multi-scale Gaussian kernels to minimize distribution differences in the latent feature space between source and target.
• Composite Loss Function: $L = L_{task} + \gamma L_{MMD} + \delta L_{phy}$, balancing prediction accuracy, domain invariance, and physical plausibility.

D. Interpretable Digital Twin

The trained model is extended into a digital twin for operational navigation. It maps process history to CPSI and regime states, allowing operators to simulate control actions (e.g., adjusting air supply or grate speed) to navigate toward lower-risk operating regimes without unsafe online exploration.

3. Key Contributions

1. Regime-Aware Architecture: Introduction of CPMoE, which explicitly models the heterogeneity of MSWI operations by combining state recognition with a mixture of diverse neural experts.
2. Physics-Guided Transfer: A novel transfer learning strategy that enforces conservation laws (mass, energy, combustion state) to anchor the model in physical reality, enabling robust cross-plant generalization.
3. System-Level Risk Metric: Development of the CPSI, shifting the focus from individual pollutant prediction to integrated carbon-pollutant synergy evaluation.
4. Interpretable Control: Transformation of the predictive model into a digital twin that supports regime-aware decision-making, bridging the gap between data-driven prediction and operational control.

4. Results

Experiments were conducted on data from 13 MSWI plants across China (Shandong, Zhejiang, Hebei).

• Source Domain Performance (Within-Plant):

  • CPMoE achieved average $R^2$ values of 0.668–0.904 for individual pollutants and 0.666–0.970 for CPSI.
  • It significantly outperformed single-architecture baselines (Transformer, CNN, LSTM, MLP), particularly in volatile plants where single models failed to capture complex regime shifts.
  • CO Prediction: While CO was difficult to predict in specific plants (MSWI#4–#6) due to combustion instability, CPMoE substantially reduced error peaks compared to baselines.

• Transfer Learning Performance (Cross-Plant):

  • After transferring from MSWI#1 to 12 target plants, the model maintained strong performance:
    • Average pollutant $R^2$: 0.661–0.842.
    • CPSI $R^2$: 0.610–0.841.
  • Stability: Physically anchored signals (CO₂, HCl) were highly transferable ($R^2 > 0.86$). Even volatile targets (SO₂, CO) showed bounded degradation.
  • Mechanism of Adaptation: Analysis of expert utilization revealed that adaptation occurs via structured re-weighting of operating regimes (e.g., CNN and MLP remained stable, while Transformer/LSTM usage varied) rather than complete model re-learning.

• Digital Twin Application:

  • The framework successfully identified distinct operating regimes and mapped them to risk levels, providing a structured basis for navigating control actions to minimize CPSI.

5. Significance

• Scalability: This work moves MSWI emission control from "one plant, one model" to a scalable, transferable paradigm, reducing the cost and time required to deploy control strategies across new facilities.
• Robustness: By embedding physical laws, the model avoids learning spurious correlations, making it more reliable under shifting operating conditions and for unseen facilities.
• Synergistic Control: The introduction of CPSI and the digital twin enables operators to make decisions that balance carbon reduction with pollutant mitigation, addressing the complex trade-offs inherent in waste-to-energy systems.
• Decision Support: It provides a pathway from pure prediction to actionable operational guidance, supporting the transition toward sustainable, low-carbon urban infrastructure.