AI-Powered Surrogate Modelling for Multiscale Combustion: A Critical Review and Opportunities

This review critically assesses the state-of-the-art in AI-powered surrogate modeling for multiscale combustion, comparing various learning approaches across scales from chemical kinetics to engine systems while highlighting key challenges like transferability and extrapolation errors, and identifying future opportunities for developing reliable, physically grounded frameworks.

The Gist

The Big Picture: Why We Need a "Speed Boost" for Fire Science

Imagine you are trying to design a cleaner, more efficient engine for a car or a plane. To do this, you need to understand exactly how fuel burns, how heat moves, and how pollutants (like smog) are created.

Scientists currently use super-computers to simulate these fires. Think of these simulations as extremely detailed, slow-motion movies of every single molecule in the fire dancing, colliding, and reacting. While these movies are incredibly accurate, they take forever to render. If you want to test 100 different fuel blends to find the best one, you might have to wait years for the computer to finish the calculations.

The Problem: The world needs cleaner energy now. We can't wait years to test new fuels like hydrogen or ammonia.

The Solution: This paper reviews a new tool called AI-powered Surrogate Modelling. Think of this as training a smart, fast apprentice to watch the slow-motion movie once, learn the patterns, and then predict what happens next in a split second, without needing to re-calculate every single molecule.

---

How It Works: The Three Levels of the Fire

The paper looks at how this "smart apprentice" (AI) helps at three different sizes, from the tiniest atoms to the whole engine.

1. The Micro Level: The Molecular Dance Floor

• The Old Way: To see how atoms bond and break, scientists used to use "Quantum Mechanics" (super accurate but painfully slow) or "Classical Mechanics" (fast but often inaccurate). It was like choosing between a slow-motion 4K camera or a blurry sketch.
• The AI Fix: The paper describes using AI to create a "Smart Map" of the molecular dance floor. The AI learns from the slow, accurate quantum data and builds a map that is just as accurate but runs as fast as the sketch.
• The Result: Scientists can now simulate how new fuels (like ammonia) break down and create pollutants without waiting months for the computer to finish.

2. The Middle Level: The Camera Lens (Experiments)

• The Problem: When scientists look at real fires in a lab, they often can't see everything. Some parts are too dark, too fast, or blocked by soot. It's like trying to guess the shape of a cloud by only seeing a few edges.
• The AI Fix: The AI acts like a super-powered photo editor.
  • Denoising: If the camera image is grainy (noisy), the AI cleans it up to reveal the true flame shape.
  • Virtual Sensing: If scientists can only measure temperature at one spot, the AI uses that data to guess the temperature of the entire fire, filling in the blanks.
  • 3D Reconstruction: If they only have 2D photos from different angles, the AI stitches them together to build a 3D model of the fire instantly.

3. The Macro Level: The Engine Simulator (CFD)

• The Problem: When simulating a whole engine, the computer has to solve complex math equations for millions of tiny grid points. The "chemistry" part (calculating how fuel burns) is the bottleneck, taking up 90% of the time.
• The AI Fix: Instead of solving the hard math equations every time, the AI uses a pre-learned shortcut. It's like a GPS app that doesn't calculate the physics of every car on the road; it just knows the fastest route based on past data.
• The Result: The simulation runs 10 to 20 times faster. This allows engineers to test many more designs in the same amount of time.

---

The "Apprentice" vs. The "Master"

The paper compares different types of AI "apprentices":

• The Basic Apprentice (Standard AI): Good at memorizing patterns it has seen before. If you ask it about a fire it hasn't seen, it might guess wrong.
• The Physics-Guided Apprentice (PINNs): This apprentice is given a rulebook (the laws of physics, like conservation of energy). It can't just guess; it must follow the rules. This makes it much more reliable and less likely to make "silly" mistakes when facing new situations.
• The Operator Learner: This is a special type of apprentice that learns the rules of change rather than just static pictures. It's like learning how a river flows rather than just memorizing a photo of the river at one moment.

---

The Catch: It's Not Perfect Yet

The paper is very honest about the limitations. Just because the AI is fast doesn't mean it's always right.

1. The "Out-of-Distribution" Trap: If you train the AI on a small candle flame, it might fail miserably when you ask it to predict a massive jet engine fire. It hasn't seen that "world" before.
2. Inconsistent Reporting: Some studies say their AI is "100 times faster," but they are comparing it to a very slow computer. Others compare it to a fast one. It's hard to know who is actually winning because everyone uses different rules.
3. The "Black Box" Problem: Sometimes the AI gives the right answer, but we don't know why. In engineering, knowing why is just as important as the answer.

---

The Future: The "Self-Driving Lab"

The paper ends with a vision for the future called "Agentic AI."

Imagine a self-driving laboratory. Instead of a human scientist spending weeks setting up experiments, cleaning data, and running simulations, an AI "agent" does it all.

• It plans the experiment.
• It runs the simulation.
• It checks if the results make sense.
• If the results are weird, it automatically adjusts the plan and tries again.
• It keeps a perfect log of everything it did so humans can check its work.

This isn't just about making things faster; it's about creating a reliable, automated loop where AI helps humans discover cleaner fuels and better engines much faster than ever before.

Summary

This paper is a review of how Artificial Intelligence is being used to speed up the science of fire. It turns slow, expensive computer simulations into fast, accurate predictions. It helps scientists see hidden details in experiments and test new fuels quickly. However, the field still needs better standards to ensure these AI tools are trustworthy and work in real-world situations. The ultimate goal is to build automated "virtual labs" that can help us solve the climate crisis by designing cleaner energy systems.

Technical Summary

1. Problem Statement

Combustion science faces a critical bottleneck in the transition to low-carbon energy systems. The decarbonization of aviation, marine, and industrial sectors requires the adoption of alternative fuels (e.g., hydrogen, ammonia, e-fuels) and complex multi-objective optimization (balancing efficiency, stability, and emissions like NOx and particulate matter).

• Computational Bottleneck: High-fidelity simulations (DNS, LES, CFD) coupled with detailed chemical kinetics involve solving stiff, high-dimensional ordinary differential equations (ODEs) across millions of cells. This is computationally prohibitive for design loops and real-time control.
• Data & Validation Gaps: While high-fidelity data exists from simulations and experiments, there is a lack of standardized benchmarks, inconsistent reporting of accuracy vs. speed, and poor validation of model robustness under distribution shifts (e.g., new fuels or operating conditions).
• Workflow Fragmentation: Current AI applications are often isolated "curve-fitting" exercises rather than integrated, trustworthy components of engineering workflows, leading to issues with physical consistency (e.g., mass conservation) and extrapolation errors.

2. Methodology

The authors conducted a critical review and quantitative synthesis of 125 peer-reviewed studies (2020–2026) covering AI applications across the combustion multiscale hierarchy. The methodology involved:

• Taxonomy Construction: Categorizing studies into model families (Supervised, Unsupervised, Physics-Informed, Operator Learning, Reinforcement Learning) and application domains (Molecular/Atomistic, Experimental Diagnostics, Continuum CFD).
• Quantitative Synthesis: Analyzing reported metrics (accuracy, speedup, inference latency) to identify trade-offs. The authors normalized data where possible to compare disparate studies and constructed Pareto frontiers to identify non-dominated model families.
• Scale-Bridging Analysis: Examining how AI connects scales:
  • Atomistic: ML potentials (MLIPs) for reactive MD.
  • Experimental: AI for diagnostics, denoising, and virtual sensing.
  • Continuum: AI for turbulence closures, chemistry acceleration, and digital twins.
• Future Framework Proposal: Introducing the concept of "Agentic AI" and "Virtual Labs" to automate the entire research lifecycle (design, simulation, analysis, and decision-making).

3. Key Contributions

A. Comprehensive Multiscale Review

The paper systematically maps AI applications across three distinct scales:

1. Molecular/Atomistic Scale:
   • ML Interatomic Potentials (MLIPs): Using Deep Potential and Neural Network Potentials (NNPs) to achieve near-DFT accuracy at MD computational costs, enabling reactive MD for complex fuels (e.g., ammonia-hydrogen blends).
   • Reaction Network Discovery: Using ML to decode trajectories from reactive MD to identify dominant reaction pathways and extract kinetic parameters (activation energies, rate constants) for reduced mechanisms.
2. Experimental Scale:
   • Diagnostic Enhancement: Using CNNs for denoising (e.g., flame emission spectroscopy) and super-resolution.
   • Virtual Sensing: Inferring hard-to-measure quantities (e.g., temperature fields, soot volume fractions) from accessible signals (e.g., chemiluminescence, pressure) using GANs and U-Nets.
   • Tomography: Accelerating 3D reconstruction of flow fields from sparse projections using physics-informed neural networks (PINNs).
3. Continuum (CFD) Scale:
   • Chemistry Acceleration: Replacing stiff ODE integration with neural surrogates (MLPs, DeepONets) or operator learning to predict reaction rates and source terms.
   • Turbulence-Closure: Using CNNs and symbolic regression to learn subfilter-scale closures for LES, improving upon traditional flamelet models.
   • Digital Twins: Developing non-intrusive surrogates that map boundary conditions to full flow fields for rapid design optimization.

B. Quantitative Performance Synthesis

The authors provide a data-driven assessment of the "Accuracy-Speed Trade-off":

• Model Families: Deep regressors (DNNs, CNNs) and Physics-Informed Neural Networks (PINNs) generally offer the best balance of high accuracy and significant speedup (10x–100x) compared to classical ML (SVM, Random Forest) which are limited to low-dimensional tasks.
• Operator Learning: Architectures like DeepONet and Fourier Neural Operators (FNO) show promise for learning solution operators across varying initial conditions, offering mesh-independent surrogates.
• Robustness: Physics-constrained models (PINNs, conservation-aware tabulation) demonstrate tighter performance dispersion and better extrapolation capabilities compared to purely data-driven models.

C. Identification of Critical Gaps

The review highlights that the field is hindered by:

• Inconsistent Benchmarks: Lack of standardized datasets and reporting protocols makes cross-study comparison difficult.
• Workflow Friction: Manual pre-processing and post-processing dominate the time cost, negating the speed gains of the AI models themselves.
• Trust & Uncertainty: Insufficient reporting of uncertainty quantification (epistemic vs. aleatoric) and lack of validation under distribution shifts (new fuels/conditions).

D. Proposal for Agentic AI Virtual Labs

The paper proposes a paradigm shift toward Agentic AI:

• Concept: Autonomous AI agents that plan, execute, and analyze combustion simulations/experiments in a closed loop.
• Application: A "Virtual Lab" for cleaner combustion where agents orchestrate ReaxFF/ML-MD simulations, validate safety constraints, perform active learning to select the next most informative simulation, and standardize post-processing.
• Goal: To transform combustion research from manual, fragmented workflows into automated, reproducible, and auditable pipelines.

4. Key Results

• Speedup vs. Accuracy: The synthesis reveals that while many studies report massive speedups (up to 20x–100x), these are often baseline-dependent. However, physics-constrained and operator-learning approaches consistently occupy the Pareto-optimal frontier, offering high accuracy with robust extrapolation.
• Deployment Efficiency: Inference latency is a critical bottleneck for real-time control. GPU-accelerated deep surrogates and PINNs are identified as the most viable for "inner-loop" solver integration.
• Experimental AI: AI-driven diagnostics have successfully demonstrated the ability to reconstruct 3D soot fields and infer emissions (NOx, CO) from sparse or noisy data, reducing the need for expensive laser diagnostics.
• Alternative Fuels: AI surrogates are proving essential for modeling ammonia and hydrogen combustion, where traditional mechanisms struggle with complex nitrogen chemistry and low-temperature oxidation pathways.

5. Significance

• Paradigm Shift: The paper marks the transition of AI in combustion from a peripheral "post-processing" tool to a core workflow component capable of accelerating the design loop for decarbonization.
• Standardization Call: It provides a strong argument for the community to adopt standardized benchmarks, reporting metrics (including uncertainty and conservation checks), and open data practices to enable true scientific progress.
• Future Roadmap: By introducing the "Agentic Virtual Lab," the authors outline a future where AI not only speeds up calculations but fundamentally changes how combustion science is conducted—making it more autonomous, reproducible, and capable of handling the complexity of next-generation fuels.
• Practical Impact: The review serves as a guide for engineers and researchers to select appropriate AI models based on specific constraints (e.g., choosing PINNs for stability in long-time integrations vs. CNNs for image-based diagnostics), accelerating the deployment of clean combustion technologies.