Combustion’s AI Edge: Modeling Fire with Machine Learning

by

Author: Denis Avetisyan

This review examines how artificial intelligence is accelerating combustion research by enabling faster, more accurate, and scalable modeling of complex phenomena.

A critical assessment of AI-powered surrogate modeling techniques for multiscale combustion, including physics-informed learning and opportunities for workflow automation and digital twin development.

Despite increasing computational power, fully resolving multiscale combustion phenomena remains a significant challenge, necessitating innovative modelling approaches. This review, ‘AI-Powered Surrogate Modelling for Multiscale Combustion: A Critical Review and Opportunities’, comprehensively assesses the rapidly evolving field of artificial intelligence-driven surrogate modelling as a means to bridge scales and accelerate combustion research. Current advancements demonstrate substantial gains in predictive accuracy and computational efficiency across applications ranging from chemical kinetics to engine simulations, yet challenges persist regarding model transferability and physical consistency. How can we develop more robust, scalable, and physically grounded AI frameworks to enable truly predictive digital twins for next-generation combustion technologies?

The Inherent Complexity of Combustion: A Multiscale Challenge

Combustion processes inherently span a remarkable range of scales, from the molecular interactions governing chemical kinetics to the macroscopic turbulence within an engine cylinder. Accurately modelling this multiscale behaviour presents a formidable computational challenge; simulating all relevant phenomena simultaneously requires resolving timescales differing by orders of magnitude – milliseconds for ignition down to nanoseconds for molecular collisions. Consequently, traditional computational fluid dynamics (CFD) approaches, relying on fine-grained meshes and explicit time integration, demand immense processing power and memory. Even with high-performance computing, fully resolving these details remains impractical for many engineering applications, particularly those involving complex geometries or transient conditions. This computational burden restricts the ability to explore a wide range of design parameters or perform real-time predictive analysis, hindering advancements in engine efficiency and emissions control.

Simultaneous resolution of turbulence, chemical kinetics, and engine dynamics presents a formidable challenge in combustion modelling. These phenomena operate across vastly different timescales and length scales; turbulent fluctuations occur rapidly, while chemical reactions and piston motion unfold much slower. Capturing their complex interplay requires models that can accurately represent the full spectrum of these processes, but doing so often leads to prohibitively high computational costs. Existing approaches frequently rely on simplifying assumptions or averaging techniques, which can sacrifice accuracy or fail to capture crucial details of the combustion process. The difficulty lies not only in the computational demands, but also in the mathematical and algorithmic complexities of coupling these disparate physical regimes within a single, cohesive framework. Consequently, achieving a truly predictive and reliable multiscale combustion model remains a central goal of ongoing research.

The intensive computational demands of multiscale combustion modelling significantly restrict the scope of engineering design and predictive capabilities. Current limitations impede detailed analysis of numerous design iterations, hindering the optimization of combustion engines and systems for improved efficiency and reduced emissions. Consequently, engineers often rely on simplified models or empirical correlations, sacrificing accuracy for computational feasibility. This trade-off impacts the development of next-generation technologies, as the ability to virtually prototype and refine designs before physical testing is severely constrained, increasing development costs and timelines. Ultimately, a lack of efficient, high-fidelity modelling restricts innovation and the potential for breakthroughs in combustion technology.

Combustion modelling’s practical application hinges on the development of surrogate models capable of delivering predictions with acceptable speed and accuracy, yet current techniques often fall short. While reduced-order modelling and data-driven approaches offer potential, they frequently struggle with maintaining predictive power across a wide range of operating conditions or exhibit limited scalability when applied to complex geometries and turbulent flows. The core challenge lies in efficiently capturing the intricate interplay between fluid dynamics, chemical kinetics, and heat transfer – processes that span multiple scales in both time and space. Consequently, engineers are actively pursuing innovative strategies, including machine learning and advanced statistical methods, to create surrogates that balance computational efficiency with the fidelity required for detailed engine analysis and optimized design exploration.

Embracing the Data: An AI-Driven Paradigm for Combustion

AI-powered surrogate modelling addresses the computational demands of combustion simulation by constructing data-driven approximations of complex physical phenomena. Traditional combustion simulations, reliant on direct numerical simulation (DNS) or large eddy simulation (LES), require significant processing time and resources. Surrogate models, trained on data generated by these high-fidelity methods, learn the underlying relationships governing combustion behavior and can subsequently predict outcomes with acceptable accuracy – often within established error tolerances – at a fraction of the computational cost. This approach enables faster design cycles, improved optimization studies, and real-time applications where traditional methods are impractical due to time constraints.

AI-powered surrogate models enable prediction of combustion behavior by leveraging data generated from computationally intensive direct numerical simulations (DNS). Rather than solving the governing equations of fluid dynamics and chemistry online for each scenario, these models are trained on DNS datasets to approximate the input-output relationships of the combustion process. This allows for rapid prediction of combustion characteristics – such as flame speed, pollutant formation, and heat release – for new operating conditions without repeating the full DNS calculation. The accuracy of the surrogate model is directly dependent on the quantity and quality of the training data, but successful implementations demonstrate the potential to significantly reduce computational cost while maintaining acceptable levels of predictive capability.

Deep Neural Networks (DNNs) are a prevalent architectural choice for surrogate models due to their demonstrated capacity for universal function approximation. This capability stems from the network’s layered structure and non-linear activation functions, allowing it to learn and represent highly complex relationships between input parameters and combustion outputs. Specifically, DNNs excel at identifying intricate patterns within high-fidelity simulation data, effectively mapping input features – such as fuel composition, temperature, and pressure – to corresponding combustion characteristics like flame speed or pollutant formation. The number of layers and neurons within each layer are key parameters adjusted during training to optimize the model’s ability to accurately represent the underlying, often unknown, function governing the combustion process.

Studies indicate that the implementation of AI-powered surrogate models yields potential workflow acceleration of up to 10x when applied to tasks such as chemistry evaluation. This acceleration stems from the models’ ability to provide rapid predictions based on learned data, circumventing the computational demands of full-scale simulations. Current research demonstrates a trajectory of rapid evolution in these models, with ongoing efforts focused on expanding their applicability to a broader range of combustion scenarios and improving predictive accuracy through architectural advancements and training methodologies.

Anchoring Prediction in Reality: Physics-Guided Learning and Operator Discovery

Physics-Guided Learning enhances machine learning model performance by incorporating established physical laws and constraints directly into the training process. This is achieved through the formulation of loss functions that penalize deviations from known physical principles, such as conservation of mass, momentum, and energy. By enforcing these constraints, the model is encouraged to learn solutions that are not only consistent with the training data but also adhere to fundamental physical realities. This approach yields improved accuracy, particularly when extrapolating beyond the range of observed data, as the model is less likely to produce physically implausible predictions and demonstrates enhanced generalization capabilities across a wider range of operating conditions.

Operator learning in combustion modeling involves identifying and learning the mathematical operators that govern the underlying physical processes, rather than directly mapping inputs to outputs. This approach utilizes techniques such as neural operators to approximate these operators – including those representing transport, diffusion, and reaction – from limited data. By learning these fundamental relationships, the model can generalize to unseen data and predict system behavior outside the training dataset, improving extrapolation capabilities and ensuring physical consistency. This contrasts with traditional data-driven methods which may struggle with out-of-distribution predictions and can produce non-physical results, as the learned relationships are not explicitly constrained by the governing physics.

Convolutional Neural Networks (CNNs) are effectively utilized for analyzing spatial combustion data due to their ability to identify patterns and features within images or spatial grids representing combustion fields, such as temperature or species concentration. These networks employ convolutional layers to extract localized features, enabling robust performance even with variations in input data. Conversely, Recurrent Neural Networks (RNNs), particularly Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) variants, are well-suited for processing the temporal aspects of combustion processes. Their recurrent connections allow them to maintain an internal state, capturing dependencies and evolution over time, which is crucial for modeling dynamic combustion phenomena like flame propagation and instabilities. The combination of CNNs for spatial analysis and RNNs for temporal modeling provides a powerful approach for comprehensive combustion data processing and prediction.

Current implementations of physics-guided learning and operator learning techniques have achieved accuracy improvements up to 92% when applied to specific combustion applications and datasets. While these results demonstrate significant potential, ongoing research and development are prioritizing two key areas: reduction of training time and enhancement of model robustness. Efforts to decrease training time include algorithmic optimizations and leveraging parallel computing architectures. Robustness is being addressed through techniques such as data augmentation, uncertainty quantification, and the development of models less sensitive to noisy or incomplete input data.

Beyond Prediction: Towards Autonomous Combustion Discovery

The development of AI-powered surrogate models is revolutionizing combustion research by enabling the creation of comprehensive Virtual Laboratories and Digital Twins. These computational replicas, trained on high-fidelity simulation data, accurately predict system behavior with significantly reduced computational cost, allowing engineers to rapidly prototype and optimize designs without the limitations of physical experimentation. By effectively mimicking complex combustion processes, these models facilitate the exploration of a vast design space, identifying optimal operating conditions and novel configurations far more efficiently than traditional methods. This capability not only accelerates the development of cleaner and more efficient engines, but also opens doors to customized combustion systems tailored to specific applications, from power generation to materials processing.

The pursuit of efficient combustion processes is being revolutionized by automated experimentation, driven by sophisticated AI-powered surrogate models. These models predict system behavior, allowing researchers to intelligently navigate the vast parameter space of combustion conditions – fuel mixtures, temperatures, pressures, and ignition timings – without relying on exhaustive physical trials. By iteratively proposing, simulating, and analyzing experiments in silico, these systems drastically reduce the time and resources needed to identify optimal operating conditions for maximizing energy output and minimizing harmful emissions. This approach not only accelerates the discovery of superior combustion strategies but also diminishes the demand for expensive and potentially hazardous physical experiments, paving the way for more sustainable and cost-effective research and development cycles in areas ranging from engine design to power generation.

The integration of Reduced-Order Modelling (ROM) with artificial intelligence significantly enhances the practicality of complex combustion simulations. Traditional, high-fidelity models, while accurate, are often computationally expensive, hindering real-time applications. ROM techniques create simplified representations of these systems, capturing essential dynamics with far fewer degrees of freedom. When coupled with AI – particularly machine learning algorithms – these ROMs become even more powerful. AI can learn to predict the behavior of the full system based on the reduced model, compensate for any loss of accuracy inherent in the simplification, and even adapt the ROM itself to changing conditions. This synergy unlocks the potential for real-time control of combustion processes, enabling dynamic optimization for improved efficiency, reduced emissions, and rapid response to fluctuating demands – critical for applications ranging from internal combustion engines to power generation and beyond.

The future of combustion research is rapidly evolving with the emergence of agentic artificial intelligence. These autonomous systems represent a paradigm shift, capable of independently formulating experimental plans, executing those plans through robotic control of laboratory equipment, and then thoroughly analyzing the resulting data – all without human intervention. This self-directed research cycle dramatically accelerates the pace of discovery, moving beyond the limitations of traditional, manual experimentation. Such systems aren’t simply automating existing processes; they are actively learning and adapting, allowing for the efficient exploration of vast chemical spaces and the identification of optimal conditions for combustion. Initial applications demonstrate the potential to synthesize and characterize hundreds of stable compounds in material discovery campaigns – a throughput previously unattainable – promising breakthroughs in areas ranging from fuel efficiency to novel material design.

The pursuit of AI-powered surrogate modelling in combustion science, as detailed in this review, echoes a fundamental truth about complex systems. Every attempt to create a digital twin, to accurately represent reality with computational efficiency, necessitates accepting a degree of approximation. As Pyotr Kapitsa observed, “It is necessary to learn to see the forest as a whole, and not just the trees.” This holistic view is crucial; the article emphasizes the need for multiscale modelling and robust workflows, acknowledging that complete fidelity is often unattainable. Instead, the focus shifts to understanding the inherent limitations and building models that age gracefully, offering valuable insights even as they diverge from perfect representation. The advancement of agentic AI within these models allows for a more adaptive approach, embracing the inevitable decay of precision with intelligent recalibration.

What Lies Ahead?

The pursuit of AI-powered surrogate modelling in combustion, as this review demonstrates, is not about achieving a final, perfect representation. It is, rather, a continuous negotiation with inherent limitations. Every abstraction carries the weight of the past, each simplification a compromise against the full complexity of reacting flows. The current emphasis on accuracy, while laudable, risks obscuring a more fundamental concern: longevity. Models, like all systems, decay. The true metric of progress will not be peak performance, but graceful degradation-the ability to maintain useful predictions even as underlying data shifts or scales increase.

The promise of agentic AI and workflow automation remains largely unrealized. Current efforts often resemble elaborate scaffolding, requiring constant manual intervention. A truly robust system will demand self-diagnosis, adaptive refinement, and the capacity to learn from its own failures-a kind of digital homeostasis. The field must move beyond simply accelerating existing workflows and confront the challenge of building workflows that are fundamentally resilient to unforeseen circumstances.

Ultimately, the value of these models will be judged not by their ability to replicate reality, but by their capacity to extend the lifespan of existing knowledge. The pursuit of ‘digital twins’ is, in essence, a form of applied entropy management. Only slow change preserves resilience. The challenge is not to build perfect representations, but to build representations that age well.

Original article: https://arxiv.org/pdf/2604.25617.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-04-30 05:26