Use of Hydrogen in Industry as a Driver for Decarbonization: A Comprehensive Review

Abstract

The high levels of greenhouse gas emissions from energy-intensive industries have created an urgent need for decarbonization. As major sources of pollution, industries are increasingly being forced to reconsider the fuels they use in their processes. Interest in hydrogen combustion, particularly in high-temperature applications, has grown due to its physical properties and the absence of carbon dioxide emissions. Despite its potential, hydrogen combustion presents technical challenges, such as flame stability, burner adjustment requirements, control of nitrogen oxides (NO_x) emissions and material compatibility. This review examines the use of hydrogen as a fuel in industrial furnaces along three complementary axes. First, it analyzes combustion fundamentals, blending limits and mitigation strategies, including oxy-fuel and MILD combustion, and assesses their deployment maturity across the steel, cement, glass, ceramics, and refining and chemicals sectors, together with economic and regulatory constraints. Second, it reviews advances in Computational Fluid Dynamics (CFD) modeling of hydrogen flames, addressing turbulence-chemistry interaction, reaction kinetics, radiative heat transfer, NO_x formation and model validation. Third, it surveys camera-based diagnostics combined with Artificial Intelligence and computer vision for flame monitoring and combustion optimization. By synthesizing recent literature, the review identifies the principal knowledge gaps, notably standardized CFD validation datasets and robust monitoring under industrial conditions, providing a critical reference for researchers and industry.

1. Introduction

High-temperature industrial processes, such as iron and steelmaking, cement, glass, ceramics, refining, and chemicals, account for between a quarter and a third of global greenhouse gas (GHG) emissions and rely heavily on coal, coke, and natural gas to deliver process heat above 1000 °C [1,2,3]. As the European Union tightens emission limits under the European Green Deal, the “Fit for 55” package and the recent Clean Industry Pact, the industry is facing increasing regulatory and market pressure to decarbonize without losing competitiveness [2,4]. Direct electrification is technically viable for only part of this demand, as the requirements for radiative heat transfer, atmospheric chemistry and product quality in glass tanks, rotary kilns, steel reheating furnaces and ethylene crackers cannot always be met solely by electric resistance, plasma or induction heating [5,6,7].

In this context, hydrogen has become a strategic energy carrier capable of bridging the gap between industrial demand for high temperatures and climate targets [4,7,8]. Compared to fossil fuels, hydrogen offers three main advantages: zero direct CO₂ emissions when burned with air or oxygen [9], a high energy content of approximately 120 MJ/kg [10], and the possibility of being produced from water and renewable electricity via electrolysis, decoupling industrial heat from fossil fuel supply chains [11]. Historically, hydrogen has been used as a chemical feedstock in ammonia synthesis, methanol production and the hydrotreatment of crude oil [12], where global consumption currently exceeds 100 Mt per year [12]. However, recent advances in electrolyzer manufacturing, storage technologies, and burner design have expanded its potential as an alternative fuel [13,14].

The feasibility of replacing fossil fuels with hydrogen is currently being investigated across sectors where reducing emissions is difficult [15]. In the steel industry, direct reduction of iron with hydrogen (H₂-DRI), followed by electric arc melting, is being scaled up to demonstration and commercial levels and is considered the most viable route to producing primary steel with near-zero emissions [7,16,17,18]. In the cement and lime sector, mixtures of hydrogen and natural gas are being tested as substitutes for coal and petroleum coke in rotary kilns, although the impact on clinker mineralogy and CO₂ emissions from limestone calcination still needs to be addressed through carbon capture [6]. In the glass and ceramics sector, combustion using oxygen and hydrogen has been demonstrated at pilot scale in float and container glass furnaces [19,20]. In the refining and chemicals sector, the replacement of gray hydrogen with green hydrogen and the partial replacement of fuel gas with hydrogen in process heaters represent the greatest opportunity for decarbonization in the short term [8,21].

Figure 1 summarizes the current deployment status of hydrogen across the four industrial sectors mentioned in this review.

Figure 1. Hydrogen deployment status across the four hard-to-abate industrial sectors. Sources: [1,6,16,17,19,20,21].

Despite this progress, hydrogen combustion poses significant engineering challenges. The flame speed of hydrogen is approximately six to eight times higher than that of methane, its minimum ignition energy is approximately sixteen times lower, and its adiabatic flame temperature is about 150 °C higher under stoichiometric air combustion conditions [22,23]. These properties increase the risk of flame flashback in premixed burners and alter heat transfer to the load via radiation and convection. They also promote thermal NO_x formation, unless mitigation strategies such as lean premixed combustion, exhaust gas recirculation, staged combustion, oxy-fuel combustion, or moderate or intense low-oxygen dilution (MILD) combustion [24,25,26,27] are employed. Furthermore, hydrogen induces embrittlement in conventional steel pipes and reacts with refractory linings, requiring the qualification of materials throughout the entire value chain [28,29,30]. Its low volumetric energy density, three times lower than that of methane, necessitates the construction of oversized storage and transport infrastructure and complicates blending strategies in existing natural gas networks [29,31,32].

Several reviews have addressed specific aspects of this transition, but none cover it comprehensively from an industrial furnace perspective. On the supply and infrastructure side, Bhuiyan and Siddique [11] surveyed production, storage and transport challenges, while Martin et al. [29] examined the implications of converting natural gas networks. At the burner level, Levinsky [23] analyzed the constraints imposed by legacy infrastructure on fuel switching, and Zhu et al. [1] projected sector-wide emissions reductions of up to 32% for US industry by 2050. More specialized contributions have appeared on MILD and oxy-fuel combustion of hydrogen-enriched fuels [22,27], CFD modeling of hydrogen flames [33,34], and flame monitoring using machine learning [35,36]. However, these works address isolated dimensions of the problem. No existing review integrates combustion fundamentals, CFD modeling, and optical diagnostics within the operational context of high-temperature industrial furnaces, nor does it connect these technical aspects to the safety, regulatory, and economic constraints that govern the transition in energy-intensive sectors.

Given these challenges, blending hydrogen with natural gas has emerged as a near-term transitional strategy that allows partial decarbonization while operating within existing infrastructure constraints [37]. However, the combustion behavior of hydrogen-natural gas mixtures changes non-linearly with hydrogen fraction, with implications for burner stability, NO_x emissions and heat transfer that are sector-specific and not yet fully characterized for industrial furnaces [14,38].

1.1. Motivation and Contribution of This Review

Three converging observations motivate this work: Firstly, the volume of experimental and modeling studies on hydrogen combustion has grown considerably since 2019, but the results are scattered across combustion engineering, materials science, process control and automation. Secondly, industrial pilot projects such as HYDESS, TWINGHY, HyInHeat, GreenHeatEU and the Glass Futures hydrogen demonstrator are generating data that requires a unifying interpretative framework [19,20]. Thirdly, the convergence of CFD with data-driven monitoring, particularly through computer vision and deep learning, is opening a new paradigm for the design optimization and real-time control of hydrogen-fueled systems, but its specific application to industrial furnaces remains fragmented [38,39].

This review contributes to: (i) reconciling the fundamentals of hydrogen combustion, including its physicochemical properties and combustion regimes, with the specific constraints of energy-intensive industrial sectors; (ii) reviewing the combustion regimes and burner strategies relevant to hydrogen and hydrogen, natural gas firing, with attention to their implications for flame stability and NO_x formation; (iii) integrating CFD modeling approaches—including reaction mechanisms, turbulence, turbulence-chemistry interaction, heat transfer, and pollutant formation prediction—with their experimental validation requirements; and (iv) examining optical diagnostic and vision-based monitoring strategies for hydrogen combustion, paying particular attention to deep learning architectures used in flame analysis.

1.2. Structure of the Article

The article is structured as follows. Section 2 describes the methodology used to identify, select and classify the literature. Section 3 reviews the fundamentals of hydrogen as a combustion fuel, common to all its industrial uses, covering physicochemical properties, combustion behavior, hydrogen and natural gas blends, and combustion strategies. Section 4 examines the use of hydrogen as a fuel across the main energy-intensive industrial sectors, including sector-specific deployment, materials and infrastructure constraints, and the main technical challenges. Section 5 analyses CFD approaches to hydrogen combustion, covering reaction mechanisms, turbulence models, turbulence-chemistry interaction, heat transfer and pollutant formation. Section 6 examines hydrogen combustion monitoring, with emphasis on optical diagnostics and image-based methods using computer vision, machine learning and deep learning. The main conclusions, open challenges and recommendations are presented in Section 7.

Figure 2 provides a visual overview of the review structure, showing the three research axes around which the article is organized and the main conclusions and main challenges.

Figure 2. Conceptual structure of the comprehensive review.

2. Methodology

This paper was prepared as a structured narrative review whose methodology was inspired by selected PRISMA 2020 principles of transparency and reproducibility, adapted to the objectives of an engineering review article rather than applied as a formal systematic review or meta-analysis protocol. Accordingly, the review provides explicit research axes, databases consulted, reproducible search strings and qualitative inclusion criteria, but it does not claim to follow a complete PRISMA screening workflow with a formal flow diagram of identified, screened, excluded and included records [40]. Consistent with the PRISMA 2020 framework, the review defines explicit research questions, reproducible search strings (Appendix A, Table A1), predefined inclusion and exclusion criteria, and a transparent classification scheme. We note, however, that this work is deliberately a structured narrative (qualitative) review and not a meta-analysis: its aim is to synthesize heterogeneous engineering evidence across three research axes rather than to statistically pool comparable quantitative outcomes. The methodology was designed to identify, compare and synthesize scientific studies and relevant documentation directly related to hydrogen combustion in industrial contexts, articulated around three complementary research axes that structure the modern study of industrial combustion: (i) experimental analysis of hydrogen and hydrogen-enriched combustion in burners and furnaces, (ii) CFD modeling of hydrogen flames, and (iii) image-based diagnostics and machine learning techniques for hydrogen combustion monitoring. Section 4, Section 5 and Section 6 of this review correspond to these three axes.

2.1. Literature Sources

The literature research was conducted across different databases such as Scopus, Web of Science, ScienceDirect and Google Scholar, complemented by selected technical reports from the International Energy Agency, the European Commission, IRENA and recognized industrial associations. Beyond peer-reviewed sources, the review also incorporates selected outputs from publicly funded projects on hydrogen combustion and technical contributions from industrial associations and equipment manufacturers, even when not peer-reviewed or not fully accessible in open form. These sources are explicitly identified as such in the text and are used to complement, not to replace, the peer-reviewed evidence, particularly where they provide otherwise unavailable operational data from industrial demonstration trials.

2.2. Literature Selection Process

The research was structured around the three research axes of the review, with three thematic queries executed in Scopus and adapted to the syntax of the other three databases for the qualitative review. Each query was restricted to peer-reviewed research articles and review articles published between 2019 and 2025. The first query additionally restricted the subject areas to Engineering, Energy and Chemical Engineering. The full Scopus query strings are reported in Appendix A, Table A1. The main emphasis was placed on the 2019 to 2025 period, although earlier seminal works on combustion theory, MILD and oxy-fuel technology, chemical kinetics and combustion diagnostics were retained when of high relevance.

Figure 3 illustrates the evolution of scientific output along the three research axes over the 2019 to 2025 period, based on Scopus indexing. The Scopus searches used for the bibliometric figures were executed on 18 June 2026. Therefore, the 2025 values should be interpreted as a Scopus indexing snapshot at that date, since records for a given publication year may continue to be indexed after the end of the calendar year. The publication counts reported in the figures should therefore be interpreted as a scoping indicator of research activity rather than as the outcome of a formal PRISMA systematic selection process. The qualitative synthesis was based on the relevance of the selected literature to the three research axes of the review. The publication counts were obtained by executing the three thematic queries reported in Table A1, one per research axis. We emphasize that the publication figures are intended solely to provide a transparent, reproducible overview of research activity and thematic emphasis across the three axes, and are not presented as formal bibliometric analysis or as evidence supporting any technical conclusion. Because they rely on a single database and on queries based on keywords, these counts are subject to indexing and terminology limitations; they should therefore be interpreted as indicative trends rather than exhaustive measures. To limit selection bias, the qualitative synthesis itself draws on four databases (Scopus, Web of Science, ScienceDirect and Google Scholar) and retains seminal pre-2019 works where relevant. Scopus was chosen as the single source for this bibliometric characterization because it offers the most comprehensive coverage of peer-reviewed literature in the relevant engineering and energy subject areas, and because relying on a single database avoids the substantial double counting that would result from aggregating records across multiple indexes. The qualitative review, in contrast, draws on all four databases.

Figure 3. Annual publications by research theme indexed in Scopus (2019–2025).

Studies were prioritized when they met at least one of the following criteria: (i) they analyzed pure hydrogen or hydrogen-enriched combustion under industrially relevant conditions, including burner geometry, thermal load and atmosphere; (ii) they discussed burner, furnace or process implications for high-temperature industrial heat; (iii) they presented CFD modeling of hydrogen or hydrogen-methane flames for industrial applications; or (iv) they addressed combustion monitoring, optical diagnostics, computer vision or machine learning based flame assessment specifically applied to hydrogen or hydrogen-enriched combustion.

2.3. Classification of Literature

Following the selection process, the selected studies were organized into four thematic groups: (i) fundamental combustion properties, (ii) industrial applications, (iii) CFD modeling, and (iv) combustion monitoring, flame imaging and machine learning. Cross-cutting issues related to safety, materials and economics were retained as a transversal category. This classification allowed the review to maintain a process-oriented perspective while preserving a clear link to decarbonization outcomes.

3. Hydrogen Combustion: Fundamentals, Properties and Combustion Regimes

3.1. Physical and Chemical Properties of Hydrogen

Hydrogen is the lightest element, with a molecular mass of 2.016 g/mol and a density of approximately 0.0899 kg/m³ under standard conditions, about one-eighth that of methane and one-fourteenth that of air [41,42]. Its lower heating value (LHV) is 119.96 MJ/kg, the highest of any common fuel on a mass basis, but only 10.8 MJ/Nm³ on a volumetric basis, approximately one-third that of methane [10,42]. This asymmetry between gravimetric and volumetric energy density is the fundamental reason why hydrogen storage and transport pose distinct infrastructure challenges compared to natural gas.

From a transport perspective, hydrogen has thermal and mass diffusivities in air significantly higher than those of methane, with a Lewis number below unity under lean conditions [41,43]. This differential diffusivity enhances local flame curvature effects and contributes to the intrinsic instability of lean hydrogen flames, a behavior that is particularly relevant in the near-extinction regime associated with excess-air operation in industrial burners [43,44]. The viscosity of hydrogen at ambient conditions is approximately 8.8 × 10⁻⁶ Pa·s, around 20% lower than that of methane (11 × 10⁻⁶ Pa·s), which affects jet mixing characteristics and burner aerodynamics at equivalent volumetric flow rates [42].

Table 1 summarizes the key physical and thermochemical properties of hydrogen compared to methane and natural gas under standard reference conditions.

Table 1. Key physical and thermochemical properties of hydrogen, methane and natural gas under standard reference conditions (0 °C, 101.325 kPa unless otherwise stated). ^a Natural gas values are indicative ranges reflecting typical European grid composition (85–92 vol. % CH4). Sources: [10,22,41,42,44,45].

Property	Hydrogen	Methane	Natural Gas a
Molecular mass (g/mol)	2.016	16.04	17–20
Density (0 °C, 1 atm) (kg/m3)	0.0899	0.717	0.72–0.85
Lower Heating Value (LHV) (MJ/kg)	119.96	50.05	47–51
Lower Heating Value (LHV) (MJ/Nm3)	10.8	35.8	31–38
Flammability limits in air (vol. %)	4–75	5–15	5–15
Minimum ignition energy (mJ)	0.017	0.274	0.25–0.30
Autoignition temperature (°C)	500–577	540–597	537–597
Laminar flame speed (stoich., 25 °C, 1 atm) (m/s)	2.4–3.0	0.37–0.40	0.38–0.45
Adiabatic flame temperature (stoich., air) (°C)	~2107	~1957	1827–1957
Adiabatic flame temperature (stoich., O2)	~3080	~3050	-
Lewis number (lean, in air)	<0.4	~1.0	~1.0
Dynamic viscosity (20 °C, 1 atm) (×10−6 Pa·s)	8.8	11.0	10–11

3.2. Hydrogen Combustion Properties: Flame Speed, Adiabatic Temperature and Comparison to Conventional Fuels

3.2.1. Flammability Limits and Ignition

Hydrogen exhibits an exceptionally wide flammability range, from 4% to 75% by volume in air, compared with 5–15% for methane [42,44]. Its minimum ignition energy is approximately 0.02 mJ under stoichiometric conditions, nearly sixteen times lower than that of methane, corresponding to more than one order of magnitude [22,42]. The autoignition temperature of hydrogen in air is approximately 500–577 °C, comparable to that of methane (540–597 °C), so spontaneous ignition under typical industrial conditions is not a primary concern; however, the low ignition energy means that electrostatic discharges, hot surfaces or mechanical sparks represent credible ignition sources that must be addressed in burner and furnace design [28,44].

3.2.2. Laminar Flame Speed

The laminar burning velocity of hydrogen under stoichiometric conditions in air at 25 °C and 1 atm is approximately 2.4–3.0 m/s, between six and eight times higher than the corresponding value for methane (~0.37–0.40 m/s) [42,43,45]. The maximum laminar burning velocity of hydrogen occurs on the fuel-rich side, at an equivalence ratio of approximately 1.7–1.8, reaching values up to 3.5 m/s, whereas methane peaks very close to stoichiometric [43,45]. This asymmetric dependence on equivalence ratio has direct implications for burner stability: operation under lean conditions, as required for NO_x control, places the hydrogen flame in a region of steeply declining burning velocity, narrowing the margin between stable combustion and blow-off [19,45]. The high sensitivity of laminar flame speed to temperature means that air preheating, routinely used in regenerative industrial furnaces to recover flue gas enthalpy, substantially increases the risk of flashback in premixed hydrogen burners [23,44].

3.2.3. Adiabatic Flame Temperature

The adiabatic flame temperature of hydrogen burning in air under stoichiometric conditions is approximately 2107 °C, about 150 °C higher than that of methane under stoichiometric air combustion conditions, based on the representative values reported in Table 1 [42,45]. In pure oxygen, the adiabatic flame temperature can reach approximately 2807 °C [41]. This elevated flame temperature directly drives thermal NO_x formation through the Zeldovich mechanism, as the rate of N₂ dissociation is exponentially sensitive to temperature above 1527 °C [26,46]. Under lean premixed conditions and at low residence times, the N₂O-intermediate and NNH pathways become relatively more significant than under conventional diffusion-flame conditions, a distinction that is important for selecting reaction mechanisms in CFD modeling [44,46].

3.2.4. Radiative Properties and Heat Transfer Implications

The complete absence of carbon in hydrogen combustion eliminates soot and CO₂ formation, fundamentally altering the radiative characteristics of the combustion products. Soot is the dominant broadband radiative emitter in conventional hydrocarbon flames, and its absence in hydrogen flames produces a substantially lower flame luminosity in the visible spectrum [47,48]. The flue gases from hydrogen-air combustion consist predominantly of N₂ and H₂O, with the latter being a selective infrared emitter concentrated in bands at 1.4, 1.9 and 2.7 µm [47]. While H₂O emits more intensely than CO₂ in some near-infrared bands, the total emissive power of a hydrogen flame is generally lower than that of an equivalent hydrocarbon flame under the same thermal load, because the absence of gray-body soot radiation is not fully compensated by the selective gas-phase emission of H₂O [48,49]. This reduction in radiative heat transfer to the furnace load has been quantified in several experimental studies and must be accounted for in both burner design and CFD modeling of hydrogen-fired furnaces, as it implies that thermal equivalence at constant fuel power does not guarantee equivalent heat flux profiles to the product. In high-temperature industries such as glass manufacturing, altered heat flux distributions may generate local temperature gradients that affect melting homogeneity, fining efficiency, and glass surface quality. Although the literature consistently identifies this risk, quantitative datasets directly correlating hydrogen firing with defect rates remain scarce, highlighting a critical research gap for industrial deployment [47,48,49].

3.3. Hydrogen and Natural Gas Blends

Blending hydrogen with natural gas is the most widely discussed near-term strategy for decarbonizing existing combustion systems, because it allows partial CO₂ reduction without requiring the complete replacement of burners, pipelines or control infrastructure [37,50]. The combustion properties of hydrogen-methane mixtures vary non-linearly with the hydrogen volumetric fraction, and this non-linearity has important practical consequences for industrial furnace operation. Therefore, the blending limits, NO_x trends and flame-stability behavior reported in the literature should not be interpreted as universal thresholds, since they depend strongly on burner geometry, equivalence ratio, operating pressure, fuel composition, air preheating, dilution strategy and furnace configuration.

3.3.1. Effect on Flame Speed and Stability

The onset of flame instability is not defined by a single hydrogen fraction but depends on the burner configuration. The laminar burning velocity of CH₄/H₂ blends increases non-linearly with hydrogen content, with the rate of increase accelerating above approximately 40–50% H₂ by volume [45,51]. For blends up to 20–30% H₂, the increase in flame speed is modest, and most existing industrial burners can accommodate the change with limited control adjustments [37,50]. Beyond 40% H₂, flame stabilization becomes a critical design parameter: the risk of flashback in premixed configurations increases substantially, whilst in non-premixed configurations the flame shortens and shifts closer to the burner tip, increasing thermal loading on refractory materials [23,44,51].

3.3.2. Effect on NOx Emissions

The relationship between hydrogen blending ratio and NO_x emissions is strongly dependent on burner geometry, equivalence ratio and operating temperature, making generalization across furnace types unreliable [14,38]. A meta-analysis of NO_x emissions from gas burners reported that, for a 5% hydrogen blend, the change in NO_x relative to pure natural gas ranged from −12 to +39% across fourteen studies, with an average increase of approximately 8% [50], reflecting the dominant influence of local flame temperature and residence time over fuel composition per se. At higher blending ratios, Cozzi and Coghe demonstrated that in non-premixed swirl flames, when the total volumetric fuel flow rate was held constant, both CO and NO_x increased progressively up to 80% H₂ content, an effect attributed to reduced thermal input and residence time at constant volumetric flow rather than to hydrogen chemistry itself [52]. Conversely, Mao et al. showed that targeted nozzle modifications in a non-premixed industrial burner can reduce NO_x by 42.8% when firing hydrogen-enriched methane by shortening the residence time at peak temperature while maintaining temperature uniformity within an 8% deviation [26].

3.3.3. Effect on Heat Transfer

As the hydrogen fraction increases, the progressive reduction in soot and CO₂ concentration shifts the balance from broadband gray-body radiation to selective H₂O emission, altering the spatial distribution of radiative heat flux within the furnace [47,48]. This effect becomes significant above approximately 30–40% H₂ by volume and must be incorporated into furnace thermal models to accurately predict product heating trajectories [49,51]. In high-temperature furnaces operating with significant air preheat, the increase in adiabatic flame temperature with hydrogen content partially offsets the reduction in flame emissivity, but the net effect on the load heat flux is geometry-dependent and cannot be assumed without case-specific modeling or measurement [38,49].

3.4. Hydrogen Combustion Strategies

The combustion strategy, defined by the state of mixing between fuel and oxidant before they enter the reaction zone and by the dilution, oxidant composition and temperature of the resulting mixture, governs flame stability, peak temperature, residence time and pollutant formation. Five strategies are particularly relevant for industrial hydrogen combustion.

• Premixed: Fuel and oxidant are mixed upstream of the burner. The reaction front propagates into a homogeneous reactive mixture at a velocity governed by the laminar flame speed. Premixed operation enables compact flames and tight control of the equivalence ratio for NO_x mitigation, but the high laminar flame speed and low ignition energy of hydrogen narrow the margin against flashback and burner tip overheating, particularly under the air-preheat conditions typical of regenerative industrial furnaces [53].
• Non-premixed or diffusion: Fuel and oxidant are injected separately, and mixing occurs in the reaction zone, which intrinsically eliminates flashback risk. The flame is governed by the scalar dissipation rate rather than by laminar flame speed. The locally stoichiometric conditions at the reaction layer maximize peak temperature and, consequently, thermal NO_x formation. This is the dominant configuration in high-temperature industrial burners (steel reheating, glass and cement) precisely because flashback is unacceptable in those applications.
• Partially premixed: Fuel is mixed with part of the oxidant upstream of the burner, while the remaining oxidant is introduced separately in the combustion chamber. This configuration combines the NO_x-control benefits of premixed operation with a reduced flashback risk, at the cost of more complex burner aerodynamics. Many industrial burners labeled as low-NO_x, including most ultra-low NO_x burners and staged designs, operate in this regime [54].
• Moderate or Intense Low-oxygen Dilution (MILD) or flameless: Exhaust gases are intensely recirculated within the furnace, preheating the reactants above their autoignition temperature while reducing the oxygen concentration below ~5% volume. Under these conditions, combustion occurs in a distributed volumetric mode rather than along a thin flame front. The temperature field becomes more uniform and peak temperatures are reduced. Thermal NO_x can be reduced to single-digit ppm levels even at full firing rate. The absence of a discrete flame mitigates the flashback concern characteristic of hydrogen premixed operation [55]. From a deployment standpoint, MILD combustion is not simply a burner swap: the regime relies on sufficiently high momentum injection to entrain and recirculate flue gases internally, together with reactant preheating above the autoignition temperature. These conditions depend strongly on the furnace geometry and enclosure, so achieving unconditional MILD operation with hydrogen, which requires a higher critical recirculation ratio than methane, may not be attainable in every existing unit. In practice, partial retrofitting is feasible where high-velocity or regenerative burners can be integrated into furnaces of adequate aspect ratio, and commercial flameless regenerative burners already accommodate fuels which contain hydrogen. However, furnaces whose geometry cannot sustain the required recirculation would need substantial redesign rather than a simple burner replacement.
• Oxy-combustion and oxygen-enriched combustion: Air is replaced, fully or partially, by oxygen, eliminating or decreasing nitrogen ballast. For pure hydrogen, this yields an essentially H₂O-only exhaust, while for hydrogen-natural gas blends a CO₂-and H₂O-rich stream suitable for capture is obtained. Flame temperatures rise substantially, up to approximately 2807 °C for stoichiometric H₂-O₂, but the absence of N₂ in pure oxy-fuel operation suppresses the thermal NO_x pathway altogether. This regime is widely used in glass, ceramics and steel processing, where very high product temperatures are required, and is the basis of the hydrogen oxy-fuel demonstrations recently reported for float and container glass furnaces (e.g., H2GLASS project [56]).

All these regimes for hydrogen and hydrogen-natural gas blends have been studied experimentally and computationally in the literature. Figure 4 shows the results from a targeted Scopus search on these regimes. Literature on premixed combustion is dominant, followed by non-premixed combustion. The scientific interest in partially premixed combustion has increased in the last year. Literature on flameless combustion and oxycombustion remains sparse compared to premixed and non-premixed combustion, likely due to the lower Technology Readiness Level (TRL) of these approaches.

Figure 4. Annual hydrogen and hydrogen-methane blends combustion publications for different combustion strategies in Scopus (2019–2025).

4. Industrial Applications of Combustion of Hydrogen and Hydrogen-Methane Blends

The deployment of hydrogen as a combustion fuel in industrial furnaces is at different stages of maturity across sectors, ranging from early-stage pilot testing to demonstration at commercial scale. Steel, cement, glass, ceramic, and refining and chemicals industries are considered in this review. Figure 5 shows the results of a Scopus search on hydrogen combustion for these industries. The number of publications has been increasing steadily in recent years. Most publications are related to the steel industry, reflecting their higher decarbonization pressure and technological maturity. The rest of the industries show a lower, similar number of publications to each other.

Figure 5. Annual hydrogen and hydrogen-methane blends combustion publications for different industries in Scopus (2019–2025).

The following overview synthesizes current experience across the main energy-intensive sectors.

4.1. Steel Industry

The steel industry represents the most advanced deployment of hydrogen in industrial thermal processing. In the direct reduction of iron (H₂-DRI) route, hydrogen serves primarily as a chemical reductant rather than a combustion fuel, reacting with iron ore in a shaft furnace at temperatures of 800–1000 °C to produce direct reduced iron (DRI) for subsequent electric arc furnace (EAF) melting [8,16,57]. Flagship projects, including HYBRIT in Sweden, H₂ Green Steel, and Salzgitter SALCOS, have demonstrated this route at pilot and pre-commercial scales, and several facilities are progressing towards full commercial operation [16,17,57]. In reheating furnaces, where steel slabs are brought to rolling temperature (1150–1250 °C), hydrogen combustion is being tested as a direct fuel substitute. The main challenges are the altered flame radiation profile, the risk of surface oxidation by elevated H₂O partial pressures in the furnace atmosphere, and the need to maintain temperature uniformity within tight metallurgical tolerances [47,58]. The HYDESS [59] and HyInHeat projects within the European Union are generating operational data on these furnaces with hydrogen fractions ranging from 20% to 100% [60].

4.2. Cement Industry

Rotary kilns for cement clinker production are among the most demanding combustion environments for hydrogen substitution. The kiln flame must deliver intense radiative heat transfer to the clinker bed at temperatures exceeding 1450 °C, and the flame shape, length and emissivity directly influence clinker mineralogy by affecting the material’s temperature-time profile [5,61]. Hydrogen flames, being shorter, hotter and less luminous than coal or petcoke flames, alter this profile in ways that affect the alite/belite ratio and potentially the clinker’s burnability [61,62]. Industrial trials at Hanson’s Ribblesdale plant in the United Kingdom demonstrated that hydrogen-biofuel blends can produce clinker that meets compositional and strength requirements, with CFD pre-analysis predicting a larger, hotter flame envelope and a higher temperature near the burner tip under hydrogen firing [61]. Critically, CO₂ emissions from limestone calcination, which account for approximately 60–65% of total cement process emissions, are not affected by fuel switching and require carbon capture regardless of the combustion fuel employed [5,63].

4.3. Glass Industry

The glass industry is particularly sensitive to combustion atmosphere, as the furnace atmosphere directly influences glass color, redox state of multivalent ions and the rate of volatilization of batch components such as sulphates and nitrates [64,65]. Oxy-hydrogen combustion has been demonstrated at pilot scale in float glass and container glass furnaces, including the Glass Futures demonstrator in St. Helens, United Kingdom, and the TWINGHY project [19,20]. The elimination of atmospheric nitrogen in oxy-fuel combustion resolves the NO_x problem but introduces high H₂O concentrations in the furnace atmosphere, which affect glass surface quality by increasing hydroxyl incorporation and modifying convective heat transfer patterns [64,65]. Air-hydrogen combustion in glass tanks has seen limited industrial deployment due to high NO_x emissions and altered radiative heat transfer, and it remains an active area of research to overcome this issue [65]. Furthermore, the reduced flame luminosity and modified radiative heat transfer associated with hydrogen combustion may alter furnace thermal homogeneity, potentially increasing the risk of defects associated with non-uniform melting and temperature gradients, although published industrial quantifications of these effects remain limited.

4.4. Ceramic Industry

In tunnel and periodic kilns for technical ceramics and sanitaryware, hydrogen substitution is constrained by the atmospheric sensitivity of the fired products and the difficulty of achieving a uniform temperature distribution with a shorter, hotter flame [6]. The lime industry, which operates rotary kilns at 900–1100 °C to calcine limestone to lime, is considered a more accessible entry point for hydrogen blending because the process is less sensitive to flame radiation profiles than cement or glass production, and product coloration is not a concern [5,6].

4.5. Refining and Chemicals Industry

Process heaters in petroleum refining, which fire the feedstock streams for crude distillation, hydrocracking and catalytic reforming, represent a near-term opportunity for hydrogen substitution, as many refineries already receive hydrogen as a by-product or co-product of their operations [1,21]. The replacement of refinery fuel gas with hydrogen-enriched streams requires attention to flame stability in multi-fuel burners, NO_x compliance, and the effect of elevated H₂O content in flue gases on heat exchanger fouling [1,21]. In ethylene crackers and steam methane reformers, where the process chemistry is tightly coupled to the furnace heat flux profile, the altered radiation characteristics of hydrogen flames represent a more significant engineering constraint [5,7].

4.6. Natural Gas Infrastructure Suitability for Hydrogen Blending

From the perspective of existing natural gas infrastructure, blends up to approximately 20% H₂ by volume are generally considered compatible with standard burners, meters and pipelines without modification, based on Wobbe index criteria and material compatibility assessments [37,50]. At higher fractions, burner redesign, valve adjustment and material qualification become necessary, particularly for components subject to hydrogen embrittlement and permeation [28,31]. Detailed material-qualification requirements for hydrogen service are established in dedicated standards, including ISO 11114-4 [66], ASME B31.12 [67] and ISO/TR 15916 [68]. Several European pilot projects, including HyDeploy in the United Kingdom and the H2-MIX project in Germany, have demonstrated the operational feasibility of blending at the distribution network level, while acknowledging that industrial end-use applications impose more stringent constraints than domestic appliances due to higher operating pressures and temperatures [69,70].

4.7. Economic and Regulatory Considerations

Although the technical feasibility of hydrogen combustion has been demonstrated across several sectors, the main constraint to its industrial adoption is economic rather than technical. The switch to hydrogen as a fuel increases operating costs, as the cost of hydrogen per unit of energy, once supplied, currently far exceeds that of natural gas [71]. Reported production costs range from 0.75 to 2.08 USD/kg for conventional ‘gray’ hydrogen produced by steam reforming of methane without mitigation measures [72], and from 4.99 to 12.18 USD/kg for electrolytic green hydrogen in Europe [73]. In the case of green hydrogen, the levelized cost is determined mainly by the price of electricity and the electrolyzer capacity factor, making it highly location dependent.

By contrast, the capital cost of the combustion equipment is comparatively modest, since the principal capital burdens lie in material qualification (Section 4.6), storage and delivery infrastructure, and safety systems rather than in the burner itself. This cost gap is expected to narrow as carbon prices rise under the EU Emissions Trading System and the Clean Industry Pact [74], but a full techno-economic assessment is site-specific and beyond the scope of this review.

4.8. Technical Challenges and Limitations

Across all sectors, a consistent pattern emerges: hydrogen combustion has been demonstrated at pilot or semi-industrial scales, but systematic experimental datasets that simultaneously characterize heat transfer, emissions, and product quality across a range of hydrogen fractions and excess air ratios remain sparse. Most published studies focus on a single output variable, typically NO_x or temperature, and do not provide the coupled heat-flux and combustion-monitoring data needed to validate CFD models or to develop real-time control strategies. This gap is particularly pronounced for furnaces in the 50–500 kW thermal range, which are representative of pilot-scale industrial equipment but are underrepresented in the literature relative to both laboratory-scale burners below 20 kW and full-scale industrial installations above 5 MW.

The adoption of hydrogen as a combustion fuel in industrial furnaces is constrained by a set of interrelated technical challenges that span burner design, process integration, emissions control and system monitoring. This section synthesizes the main limitations identified in the literature, distinguishing between challenges inherent to the physics of hydrogen combustion and those arising from interactions between hydrogen and existing furnace infrastructure.

Flashback, the upstream propagation of the flame into the burner mixing zone or fuel supply line, is the most safety-critical challenge specific to premixed hydrogen burners. The high laminar flame speed of hydrogen, combined with its low minimum ignition energy and wide flammability range, means that the margin between the flow velocity required for stable flame anchoring and the velocity at which flashback occurs is substantially narrower than for methane [23,44]. This margin is further reduced by air preheat, which is standard practice in regenerative industrial furnaces, and by the low-pressure-drop conditions that are typical of large industrial burner arrays [23,75]. Non-premixed (diffusion) burner configurations avoid flashback by design but introduce trade-offs in NO_x formation, since the locally stoichiometric conditions at the flame front maximize thermal NO_x production [26,46]. Partially premixed and staged combustion designs represent engineering compromises that are actively being developed for industrial hydrogen applications [26,44,75].

Thermal NO_x formation is the dominant emissions challenge for hydrogen combustion in air. The absence of fuel-bound nitrogen and prompt NO formation pathways that are characteristic of hydrocarbon fuels means that, in principle, hydrogen combustion should produce lower NO_x than coal or heavy fuel oil under the same temperature conditions. In practice, however, the higher adiabatic flame temperature of hydrogen leads to substantially higher thermal NO_x than that of natural gas under comparable excess-air conditions, particularly in conventional diffusion flames [26,46,76]. Several mitigation strategies have been demonstrated at laboratory and pilot scale: lean premixed combustion reduces peak flame temperature at the cost of increased flashback risk; exhaust gas recirculation (EGR) dilutes the oxidant stream and lowers flame temperature; staged combustion separates the fuel and air injection to avoid locally stoichiometric zones; and MILD combustion, achieved by operating with highly preheated and diluted reactants above autoignition temperature, has been shown to achieve ultra-low NO_x with hydrogen fuels when properly implemented [24,25,26,46]. The challenge for industrial furnaces is that these strategies are geometry-specific, and their effectiveness is sensitive to the precise flow field, which varies across furnace types and cannot be reliably predicted without validated CFD models [33,38,77]. Consequently, hydrogen blending limits and NO_x mitigation strategies must be evaluated case by case, rather than transferred directly from one burner or furnace configuration to another.

The substitution of hydrocarbon fuels with hydrogen modifies the spectral and total emissive power of the flame and combustion gases, with consequences for the heat flux distribution to the furnace load [47,48,49]. In furnaces where product quality is controlled by the temporal temperature profile, such as glass tanks, steel reheating furnaces and ceramic kilns, these changes cannot be compensated solely by adjusting fuel power: the spatial redistribution of heat flux may require burner repositioning, flame geometry redesign or active control of the equivalence ratio profile along the furnace length [48,58,64]. Current engineering tools for predicting these effects, including zone models and full CFD simulations, require spectral radiative property data for H₂O-N₂ mixtures at high temperature that are not yet validated with the same confidence as the corresponding data for CO₂-containing combustion products [47,49].

Hydrogen induces embrittlement in ferritic and martensitic steels through the adsorption and diffusion of atomic hydrogen into the metal lattice, reducing ductility and fracture toughness and promoting subcritical crack growth [28,29,30]. The susceptibility to embrittlement increases with hydrogen partial pressure, temperature cycling and the presence of stress concentrations, all of which are present in industrial burners and piping systems [28,30]. Austenitic stainless steels offer better resistance but are not immune, particularly at elevated temperatures and pressures [29]. Refractory materials in furnace linings may be affected by high H₂O partial pressures in hydrogen-fired atmospheres, which can cause corrosion of silica-containing refractories through hydroxyl volatilization and accelerate sintering of alumina-based materials [30,77]. The qualification of materials throughout the entire fuel supply and combustion system value chain is a prerequisite for the safe industrial deployment of hydrogen and is an area where standardized testing protocols and regulatory frameworks are still being developed at European and international levels [28,29,77].

Conventional combustion monitoring in industrial furnaces relies on thermocouples, suction pyrometers and extractive gas analyzers, supplemented in some installations by optical probes for flame detection [35,78]. These instruments were designed and calibrated for hydrocarbon fuels and present limitations when applied to hydrogen flames: the near-invisible nature of hydrogen-air flames in the visible spectrum makes conventional camera-based flame detection unreliable, while the absence of CO₂ invalidates some infrared-based sensors [35,48,78]. UV-based detection of OH* chemiluminescence, a marker of the reaction zone in both hydrocarbon and hydrogen flames, provides a more robust diagnostic for hydrogen-rich combustion and has been validated in several laboratory studies [36,78,79]. The integration of RGB and UV camera systems with machine learning algorithms for real-time flame characterization is an emerging approach that is discussed in detail in Section 5. The challenge for industrial deployment is the development of monitoring systems that can operate reliably under the harsh thermal and mechanical conditions of high-temperature furnaces while providing the spatial and temporal resolution required for model-based control of NO_x and heat flux [35,36,79].

4.9. Research Gaps

Three gaps emerge consistently from the literature reviewed in this section. First, experimental datasets that simultaneously characterize heat transfer, NO_x emissions and flame structure across a range of hydrogen fractions and excess-air ratios in a single furnace geometry are rare, limiting the cross-validation of CFD models. Second, the transition from laboratory flame studies to pilot-scale furnace experiments has been made in relatively few studies, and the scaling behavior of NO_x mitigation strategies from burner to furnace level is not well established. Third, integrated monitoring systems combining UV and visible optical diagnostics with real-time data processing have not been systematically evaluated in industrial hydrogen-fired furnaces, leaving a gap between the diagnostic capabilities demonstrated in research settings and the monitoring tools available to plant operators.

4.10. Cross-Sector Comparison of Hydrogen Combustion Deployment

The feasibility of hydrogen combustion differs markedly across the industrial sectors reviewed above. In steelmaking, hydrogen deployment is comparatively advanced when associated with direct reduction routes and reheating applications, where demonstration and commercial initiatives are already being developed. However, implementation remains constrained by material compatibility, hydrogen embrittlement, refractory degradation, burner adaptation and the availability of low-carbon hydrogen supply [7,16,17,18,28,29,30,53]. In this sector, hydrogen therefore represents one of the most technically credible decarbonization routes, but its deployment requires coordinated changes in fuel supply, furnace operation and materials qualification.

In cement and lime production, hydrogen can reduce pollutant emissions and may be integrated into rotary kiln firing systems, but it cannot eliminate the process CO₂ released during limestone calcination. Therefore, its decarbonization value is structurally limited unless combined with carbon capture or other process mitigation strategies [6,61,62,63]. The main implementation barriers are associated with kiln atmosphere, flame shape, heat-transfer distribution, clinker quality and the fact that combustion-related emissions represent only part of the total carbon footprint.

Glass and ceramics occupy an intermediate position in terms of deployment maturity. Pilot and demonstration activities have shown that hydrogen and hydrogen-enriched combustion can be technically feasible, particularly when combined with oxy-fuel operation [19,20,56,64,65]. Nevertheless, these sectors are especially sensitive to temperature uniformity, radiative heat transfer and atmosphere composition. The lower luminosity of hydrogen flames and the modified contribution of water vapor radiation can alter the heat flux profile inside the furnace, potentially affecting melting homogeneity, fining, surface quality and defect formation [23,49,51]. Consequently, the main barrier in glass and ceramics is not only whether hydrogen can be burned safely, but whether equivalent product quality can be maintained under modified thermal boundary conditions.

Refining and chemical industries offer a comparatively near-term opportunity because hydrogen is already produced, handled or consumed in many sites, and partial substitution of fuel gas in process heaters may require fewer radical changes than in sectors where hydrogen infrastructure is absent [8,21]. However, implementation remains constrained by fuel logistics, leakage control, safety management, burner compatibility, NO_x formation and the non-linear behavior of hydrogen-natural gas blends [23,29,37,38,50]. In these applications, retrofit complexity may be lower than in steel, cement or glass, but operational risk management and infrastructure compatibility remain decisive.

Overall, this comparison shows that hydrogen combustion should not be interpreted as a uniform decarbonization solution for all high-temperature industries. Its implementation readiness depends on combinations of fuel availability, retrofit complexity, operational barriers, safety requirements, commercialization status and tolerance to changes in heat transfer profiles. The dominant limiting factor shifts from materials and process-route transformation in steel, to process chemistry in cement, to product-quality control in glass and ceramics, and to fuel logistics and safety management in refining and chemicals. This explains why future deployment pathways must be assessed sector by sector rather than generalized across all energy-intensive industries.

5. Computational Models of H₂ Combustion in Industrial Furnaces

Computational Fluid Dynamics (CFD) constitutes a key tool for design and assessment of the performance of combustion systems. CFD simulation results can provide detailed information on flame structure, temperature distribution, pollutant formation, heat transfer, and other relevant phenomena that are complex to measure experimentally.

The wide range of physical and chemical combustion phenomena across multiple spatial and temporal scales is not computationally practical to be resolved directly in industrial engineering simulations. Closure models are required to represent phenomena such as turbulence, turbulence-chemistry interaction, radiative heat transfer, transport processes, and pollutant formation. Closure models derived for hydrocarbon fuels are oftentimes based on assumptions that do not apply to hydrogen flames, due to characteristics such as high diffusivity, low Le, high flame speed, and thermo-diffusive instability. Hence, most recent scientific literature on hydrogen combustion CFD modeling has focused on evaluating the applicability of existing models to such fuel, as well as on reviewing assumptions and modifying models to achieve this applicability.

In this chapter, we present key insights from recent scientific literature on derivation and validation of reaction mechanisms and CFD closure models for combustion of hydrogen and hydrogen-methane blends in industrial furnaces. The applicability, characteristics and limitations for each of the models are highlighted, and research trends and gaps are identified.

5.1. Reaction Mechanisms

The development of reaction mechanisms adequate for the required accuracy and available computational resources is critical for the industrial use of CFD for hydrogen combustion studies. In recent years, several authors have worked on deriving and validating such mechanisms, both for hydrogen and hydrogen-methane combustion.

For natural gas combustion, the GRI-Mech 3.0 is the industry standard [80]. Other reaction mechanisms, such as USC mech 2.0, San Diego mech, and Aramco mech 1.3 (C3 version), are also widely used. These reaction mechanisms have been evaluated for chemical kinetics of hydrogen-methane blends, using lab-scale spherical flame experimental results, and CHEMKIN PRO as software [81]. The Aramco mech 1.3 shows the best performance in predicting laminar burning velocity for most conditions, followed by GRI-Mech 3.0 (except for high pressure). The USC Mech 2.0 works better for hydrogen blends with low hydrogen content, while the San Diego Mech works better for high hydrogen content. None of these mechanisms can predict hydrogen-rich or pure hydrogen combustion at elevated pressures. To cover for this deficiency, several detailed, skeletal, reduced, and global mechanisms have been derived and assessed in recent years for mixtures up to 100% hydrogen, with satisfactory results on accuracy and computational cost [82,83].

A 10-step mechanism is recommended to capture all qualitative features of pure hydrogen combustion systems [84]. Single-step mechanisms capable of obtaining reasonably accurate predictions of laminar flame speed, auto-ignition time and adiabatic flame temperature have been developed in recent years [85]. Work has also been done on the use of machine learning for derivation of kinetic mechanisms for hydrogen combustion [86]. Via this technique, the error in the prediction of delay times and laminar flame speeds is decreased by ~25%. The use of machine learning for deriving kinetic mechanisms shows great potential for further improvement of hydrogen combustion models.

5.2. Turbulence Models

The dominant turbulence models used for research and industrial CFD in recent years are Reynolds-Averaged Navier–Stokes (RANS), its unsteady version (URANS), Large-Eddy Simulation (LES), and Detached Eddy Simulation (DES).

The rate of publication of scientific papers using turbulence models for hydrogen and hydrogen-methane combustion studies has been growing in recent years. From a Scopus search on the use of turbulence models in recent years (Figure 6), the use of LES is predominant in hydrogen combustion studies. Publications using RANS and URANS are also common. The use of DES is marginal.

Figure 6. Annual CFD hydrogen combustion publications by turbulence model in Scopus (2019–2025).

RANS is the standard for simulation of combustion in industrial systems, due to its reasonable computational cost. RANS methods solve for time-averaged flow, eliminating the impact that fluctuations have on the reaction and on the flow. This limitation plays a relevant role for hydrogen combustion. Diffusion effects are an example of this. Preferential and differential diffusion depend on local, instantaneous strain and curvature of the flame front, which are not resolved on RANS. Hence, they are generally included via corrections. Thermo-diffusive instability (i.e., the flame front generating small-scale wrinkling due to low Le effects) is inherently a fluctuation phenomenon, not directly represented in RANS simulation results. Additionally, transient phenomena such as ignition delay, flashback, and local extinction and reignition cannot be predicted using RANS. Further, regarding flame heat transfer, Turbulence-Radiation Interaction cannot be captured by RANS simulations due to its dependence on local flame structure and wrinkling, leading to underprediction of radiative heat transfer. This is extremely relevant for the use of CFD for industrial furnace design and performance assessment, as generally heat transfer is one of the Key Performance Indicators of the system. All these limitations make RANS use limited to applications in which the assessment of system-level quantities, such as total heat release or average temperature, is sufficient. Even for such applications, adequate corrections and closure models shall be applied to account for the global impact of the phenomena mentioned above. Relevant operating conditions such as local wall heat flux, flame stability margins, or pollutant formation cannot be predicted reliably by RANS.

RANS k–ε with a steady diffusion flamelet model has been used to characterize steady combustion performance of non-premixed turbulent hydrogen-methane blends in non-ferrous melting furnaces for casting [87]. The RNG k-epsilon and standard k-epsilon approaches have been compared in terms of accuracy with respect to experimental data, concluding that the RNG model is superior to the standard one [88]. URANS has been used in the literature to simulate industrial furnaces in which large-scale transient behavior of hydrogen flames influences flame stabilization and heat transfer. The k–ω Shear Stress Transport (SST) URANS method has been used for hydrogen-air diffusion flames, obtaining accurate predictions on axial velocity and mixture fraction profiles [89]. Scale-Adaptive Simulation (SAS) provides the most accurate temperature distribution predictions in URANS simulations when compared to realizable k–ε, k–ω SST, and Reynolds Stress Models (RSM) [90].

As previously indicated, LES is the dominant turbulence model in recent scientific literature. LES resolves large turbulent scales directly and models the small sub-grid ones. LES is increasingly used for industrial combustion simulations, due to its high accuracy compared to RANS. However, its computational cost (1–2 orders of magnitude higher than RANS) remains prohibitive for many applications. As discussed above, for studies in which transient phenomena are targeted (e.g., flashback prediction, flame stability margins, thermo-diffusive flame front breakup) or affect the predicted results (e.g., local wall heat flux), RANS generally cannot be used. Hence, LES becomes the chosen alternative if allowable by computational resources, domain size, and geometry complexity.

The accuracy of LES models for hydrogen combustion prediction has been widely assessed in the literature [91,92,93,94]. Improvement of existing LES models for hydrogen combustion applications has also been pursued by authors (e.g., novel explicit spatial filtering for turbulent premixed hydrogen-air combustion [91]). Novel approaches to improve the prediction of phenomena such as thermo-diffusion, differential diffusion, and preferential diffusion (e.g., for studies in micromix combustors [92], lean premixed low swirl lifted flames [93]) have been developed.

LES is increasingly used in hydrogen industrial furnace applications for detailed studies of flame dynamics (e.g., lean premixed low swirl lifted flames [91]), burner design (e.g., micromix combustors [92]), lifted hydrogen-nitrogen vitiated coflow jet flames [94], flame stability studies (e.g., ignition delay in non-premixed auto-igniting hydrogen flames [95]), and diffusion phenomena (e.g., differential diffusion impact on hydrogen-methane oxycombustion diffusion flames [96]). Under the scope of EU-funded projects, LESs for hydrogen combustion are being performed in the HyInHeat Project [60]. LES simulations accurately predict flame stability and diffusion phenomena in hydrogen flames.

The use of DES methods is much less extended in combustion studies than RANS and LES. DES was originally developed for high-Re external aerodynamics, to capture large-scale vortex shedding while reducing boundary layer treatment cost. The application of this approach to combustion is limited, as the RANS/LES interface is grid-dependent, which can create ambiguities in the local application of TCI closures. There are recent studies using DES for high-speed [97] and supersonic [98] hydrogen combustion, but no studies with application of DES to industrial furnaces have been identified.

5.3. Chemical Source Term Closure Models—Turbulence-Chemistry Interaction (TCI)

The dominant chemical source term closure models used for research and industrial CFD in recent years are the Eddy Break-Up (EBU) model, the Eddy Dissipation Model (EDM), the Eddy Dissipation Concept (EDC) model, the Flamelet-Generated Manifold (FGM) closure, the Turbulent Flame Closure (TFC) model, and the Transported Probability Density Function (TPDF) closure. The selection of the most appropriate closure model for turbulence-chemistry interaction depends on the combustion regime, the desired output quantities and the required balance between computational cost and predictive accuracy.

Figure 7 shows the results of a Scopus search on the use of TCI models for hydrogen combustion in scientific literature in recent years. There is a clear dominance of the use of flamelet methods, followed by PDF/TPDF closures. The use of EDM and EDC has been lower, although in the last year it has gained traction.

Figure 7. Annual CFD hydrogen combustion publications by TCI model in Scopus (2019–2025).

For hydrogen combustion, EBU has mainly been used for simulation of large-scale hydrogen deflagration experiments in the literature [99], with little application for industrial furnaces. Similarly to the EBU model, authors have used EDM for premixed hydrogen deflagration modeling. The combination of EDM and TFC models to adequately characterize the mean reaction rate can take advantage of the capabilities of both models [100].

EDC is the industry standard for non-premixed combustion simulations, especially for jet diffusion flames in which mixing controls the reaction rate. This model applies finite-rate Arrhenius kinetics to the turbulent structures of each cell, using two empirical constants related to the reacting fraction and residence time, which are dependent on the local turbulence field. EDC has lower computational cost than FGM and PDF, making it a suitable solution for industrial furnace applications. Further, ignition delay and NO_x reaction kinetics are captured by this model. However, the use of EDC requires recalibration of the constants for combustion of hydrogen and hydrogen-methane blends. The EDC model constants have been recalibrated for specific combustion applications of hydrogen and hydrogen-methane blends, such as flameless combustion [101].

FGM is widely used for turbulent premixed flames, as it captures flame stretch and preferential diffusion effects by parameterizing the flamelet manifold using a reaction progress variable. Extended manifolds use the mixture fraction as a second control variable for parameterization of partially premixed and diffusion flames. This method provides detailed chemistry results at reasonable computational cost, as the chemical reactions are tabulated rather than solved for every timestep. The FGM closure has been extensively used in recent years in combination with LES to model a wide variety of hydrogen combustion cases: micromix combustors [92], lean premixed hydrogen-air low swirl lifted flames [93], turbulent lifted hydrogen-nitrogen vitiated coflow jet flames [94], non-premixed auto-igniting hydrogen flames [95], laminar premixed flames [102], turbulent hydrogen-air premixed flames [103], partially premixed hydrogen-air flames [104], and ultra-lean premixed hydrogen-air flames [105]. Authors highlight the accuracy of the results obtained with this methodology in relation to DNS and experimental campaigns. FGM assumes a 1D flamelet structure, which does not always hold when phenomena such as thermo-diffusive instability or strong local extinction, typical of hydrogen flames, appear. This approach can account for flame stability and diffusion aspects via corrections for non-unity Le, preferential diffusion, flame stretch, and thermo-diffusive effects.

The PDF/TPDF model does not make an assumption on the combustion mode, as the mean reaction rate is closed through a composition statistical distribution. This model is frequently chosen for simulation of partially premixed flames. The model has also been extensively used for diffusion flames. This model has been used in the literature for hydrogen and hydrogen-methane combustion applications such as industrial melting furnaces for secondary aluminum production [106] and micromix combustion for boiler burners [107], and non-industrial applications such as hydrogen combustion in gas turbine engines [108] and rocket engines [109]. The computational cost of PDF/TPDF methods is generally prohibitive for simulation of industrial systems, as the full composition PDF needs to be solved for the entire system. The combination of RANS and TPDF methods is proposed as an accurate, lower-cost alternative to LES and DNS for modeling non-premixed turbulent hydrogen-methane flames [110]. However, this approach is limited by the incapability of RANS to capture unsteady local fluctuations, which drive diffusion phenomena, ignition and extinction.

5.4. Diffusion Phenomena Models

Differential diffusion, preferential diffusion, Soret diffusion, and thermodiffusive instability are highly relevant characteristics of hydrogen combustion. These effects are derived from the non-unity Le characteristic of hydrogen, which substantially impacts its diffusion. These phenomena significantly impact flame structure and stability.

The inclusion of these diffusion effects for hydrogen combustion is a relevant focus of recent scientific literature on CFD modeling, with 110 publications in recent years according to Scopus (Table A1). All these phenomena have had similar scientific relevance in recent years, as seen in Figure 8.

Figure 8. Annual CFD hydrogen combustion publications by diffusion phenomena in Scopus (2019–2025).

Hydrogen-specific transport phenomena have been studied in the literature for specific configurations, such as micromix combustors [92], lean premixed low swirl lifted flames [93], non-premixed auto-igniting turbulent hydrogen flames [95], and hydrogen-methane oxy-combustion diffusion flames [96].

As previously discussed, these phenomena cannot be directly captured by RANS, as they are dependent on unsteady local flame curvature and strain. Further, the mixture-averaging of diffusion coefficients does not allow for capturing these phenomena either. Capturing these phenomena requires empirical corrections applied during post-processing of the mean field solution, multicomponent diffusion models, and/or the use of LES/DNS.

5.5. Heat Transfer Models

Predictive heat transfer modeling is a critical capability for comparative industrial furnace studies considering changes in fuel. Due to their different heat release characteristics, methane and hydrogen flames will lead to different temperature distributions within the furnace. Conductive, convective and radiative heat transfer need to be re-evaluated, ensuring that the underlying assumptions are valid for the study cases. These heat transfer mechanisms can be comparatively quantified using CFD simulations.

According to a targeted Scopus search (Table A1), since 2019, >50 publications related to CFD modeling of heat transfer in hydrogen-fired furnaces have been found. A total of 27 publications related to radiation and 14 related to convection have been identified.

Overall, authors conclude that convective heat transfer is increased for hydrogen combustion compared to methane. Over 5% peak flame temperature increase, and higher temperature uniformity is obtained by blending methane with hydrogen up to 2% in volume in industrial furnaces [111]. Hydrogen combustion can obtain up to 13% faster convective heat transfer rates of samples in industrial furnaces compared to natural gas combustion, according to simulations [112]. For Electric Arc Furnaces with hydrogen-methane non-premixed impinging flames, flame temperature and convective heat transfer increase with hydrogen blending percentage [113].

For accurate radiative heat transfer prediction, gas radiative properties need to be adequately characterized. The weighted-sum-of-gray-gases (WSGG) approach is generally used for determining the absorption coefficient of the gas mixture. This approach has been recently recalibrated for gas compositions corresponding to reactants and products of combustion of hydrogen and hydrogen-methane blends [114,115]. However, comprehensive validation of this model using industrial-scale simulation remains a gap. Its accuracy has only been assessed for a given range of pressures and blend ratios, not for the full range of system sizes and operating conditions.

Including radiation in CFD combustion studies is generally required for achieving accuracy in the prediction of flame temperature and structure. During model validation, authors found that relative error in temperature prediction decreased by 5% when radiation is considered [116].

In turbulent high-temperature flames, heat transfer by radiation interacts with turbulence, affecting flame characteristics. This phenomenon is referred to as Turbulence-Radiation Interaction (TRI). RANS CFD models generally do not account for TRI, as this phenomenon is dependent on instantaneous temperature, species, and radiation intensity fluctuations. Generally, RANS models underestimate radiative heat transfer [117], as mean variable fields do not account for this effect. TRI can increase radiative heat transfer from the flames to the walls or loads by 40%, also affecting CO and NO_x formation predictions [117,118], as both are sensitive to local instantaneous temperature.

5.6. NOx Formation Modeling

There have been extensive efforts in recent years to develop accurate models and quantify NO_x emissions of hydrogen and hydrogen-natural gas blends combustion in the scientific literature. Since 2019, there have been >200 scientific publications related to NO_x formation modeling in hydrogen and hydrogen-natural gas blends combustion, according to a targeted Scopus search (Table A1).

Most NO_x emissions by hydrogen combustion are generated via the thermal pathway, due to the high local temperatures. The NNH and N₂O mechanisms are also relevant, especially at regimes with low local temperatures, such as flameless combustion. The fuel-bound and prompt pathways are not present, as they appear due to hydrocarbon fuel composition.

The thermal pathway is generally the focus of CFD simulations for prediction of NO_x emissions during combustion of hydrogen and hydrogen-methane blends, except for flameless combustion. This modeling is implemented via correlation of RANS temperature maps to empirical NO_x formation models [119] or transport equations [92]. The postprocessing approach generally underestimates NO_x formation, as instantaneous local temperature fluctuations and TRI are not considered, which significantly affect this formation.

Several CFD studies for diffusion combustion of hydrogen-natural gas blends for industrial furnaces (ceramic [111], heat treatment of aluminum ingots [119], and other applications [120]) reported NO formation increases of 35–150% for blends between 20 and 60% of hydrogen in volume. Certain combustor concepts, such as micromix [92] and flameless [121] combustors, are reported to present a significant reduction in NO_x emissions compared to existing diffusion combustors, due to their lower local peak temperatures. However, as discussed, other NO_x pathways, such as NNH and N₂O mechanisms, shall be considered to evaluate this reduction in formation.

5.7. Flame Stability Modeling

Flame stability phenomena in hydrogen combustion via CFD simulations have been widely covered by scientific literature in recent years. As seen in Figure 9, in recent years, around 50 publications per topic have been released regarding relevant flame stability phenomena: flashback, ignition delay or autoignition, and quenching or local extinction.

Figure 9. Annual CFD hydrogen combustion publications related to flame stability phenomena in Scopus (2019–2025).

Relevant efforts have been invested in recent years towards achieving models with predictive flashback capabilities for hydrogen combustion. Flashback prediction requires sufficient mesh resolution for resolving boundary layer and vortex core dynamics, as these phenomena dominate flashback mechanisms. Hence, the recent literature has generally been focused on model development and validation using reference flames, due to the extremely high computational cost of models capable of predicting flashback for industrial-scale furnaces. Flashback prediction studies have been performed on simplified cases such as premixed turbulent swirling combustors with central air injection using LES and flamelet models [122], premixed hydrogen-methane-air laminar flames using finite rate chemistry [123], laminar lean premixed hydrogen-air jet flames [124], and premixed hydrogen-air flames confined in burner ducts [125]. Hydrogen combustion flashback has also been studied recently in industrial use cases using CFD, such as for premixed hydrogen-air combustion in gas condensing boilers [126], although industrial-scale studies remain scarce.

Ignition delay and autoignition affect stability and propagation of the flame. These phenomena have been studied in the literature using CFD. They depend on local mixture composition and temperature history. Ignition delay and self-ignition for diffusion hydrogen flames have been predicted accurately in the literature using LES-FGM [95], as this approach captures the temperature history through the storage of the transported progress variable. For hydrogen-methane flames, using a RANS k-ε with modified EDC approach, it is concluded that increasing hydrogen blending percentage, increasing oxygen content in air, and using MILD combustion reduce ignition delay time when compared to methane-air flames [127].

Flame quenching and local extinction are highly relevant phenomena related to flame stability. They occur when local strain rate or heat loss exceeds the limit that the local flame reaction rate can sustain. The different diffusivity and flame speed of hydrogen compared to other hydrocarbons make quenching limits not directly possible to extrapolate. Quenching limits of hydrogen diffusion flames have recently been characterized for laminar, simplified configurations [128]. However, these limits shall be characterized further at turbulent, industrial-scale conditions. Numerical studies on quenching of laminar premixed hydrogen-methane flames have been performed and validated in the recent literature [129].

5.8. Verification, Validation, and Uncertainty Quantification

Predictive reliability remains one of the principal challenges in hydrogen combustion modeling. Hence, verification, validation quality and Uncertainty Quantification (UQ) play a key role in assessing the suitability of different methods.

Mesh convergence studies are standard practice for CFD simulations. Specific characteristics of hydrogen combustion, such as low laminar flame thickness, local phenomena such as thermo-diffusive instabilities, and boundary layer flashback, create special emphasis on the need for local refinements (e.g., close to the flame front and to walls). The method by Roache [130], based on the use of a Grid Convergence Index, is the standard used in industry to this end. However, the application of this method is not generalized in the hydrogen combustion literature.

Both experiments and DNS have been used in the recent literature to validate hydrogen and hydrogen-methane combustion CFD models.

For experimental hydrogen combustion model validation, in recent years, canonical flames such as lean premixed low swirl lifted flame [93] and vitiated coflow jet flame [94] have been used. However, hydrogen combustion benchmarking remains limited compared to hydrocarbon combustion. For hydrocarbon combustion, benchmarks such as Sandia and DLR flames have been built up over more than three decades of work, via communication and agreement between research groups. Hence, turbulence and TCI models for natural gas combustion have been extensively calibrated and cross-validated using a systematic database of flames. Since 2023, the TNF workshop [131] has developed its own hydrogen-specific dataset initiative, based on a set of EU projects funded under the Clean Aviation Program (HYLON, AHEAD, NTNU). This dataset shall be further standardized, developed, and used for cross-validation to become a benchmarking standard for CFD validation.

DNS is generally used as a representative baseline for comparison of combustion simulations using turbulence (e.g., RANS, LES) and reaction (e.g., flamelet, EDC) models of combustion. Its extremely high computational cost, especially at high Reynolds numbers (such as in industrial furnaces), makes this approach infeasible for engineering applications, being restricted to research or validation processes within small, well-defined domains. In the recent literature on hydrogen combustion, DNS has been used to obtain a better physical understanding of complex models and difficult-to-measure phenomena, such as turbulence models (e.g., evaluation of novel spatial filtering for LES in turbulent premixed hydrogen-air combustion [91]), diffusive phenomena (e.g., preferential diffusion and thermo-diffusive instability in lean hydrogen flames [105]), turbulence-chemistry interaction (e.g., effect of turbulence on reaction rates in swirling hydrogen-air premixed flames [132]), and flame stretch (e.g., effect of pressure on flame structure in lean premixed hydrogen-syngas spherical flames [133]).

Literature on validation of CFD models at semi-industrial and industrial scales is scarce. Authors have performed model validation using natural gas combustion in semi-industrial furnaces, to then use the validated models to assess CFD modeling approaches for hydrogen combustion. An example of this is the work by Daurer et al. [134], who performed several turbulence and TCI model assessments for hydrogen oxyfuel combustion in a semi-industrial furnace (180 kW), after validation considering experimental results from natural gas combustion. The applicability range of this validation should be assessed carefully to ensure that the results are reliable.

UQ helps assess the accuracy of CFD results and identify where the focus should be allocated to improve it. For combustion of hydrogen and hydrogen blends, authors have worked towards UQ of ignition delay times and NO_x emissions using novel methods such as Bayesian analysis integrated with Artificial Neural Networks [135], and non-intrusive polynomial chaos expansion based on the point collocation method with the Markov Chain Monte Carlo method [136].

5.9. Trends and Research Gaps on CFD Models for Hydrogen and Hydrogen-Methane Blends Combustion

As previously reported, there is a growing interest in CFD for hydrogen combustion in the scientific literature. This is justified by recent industrial decarbonization efforts and by the increase in model and computational capabilities. However, there are some research gaps in phenomena modeling that should be covered in future research.

The derivation of low-cost, high-accuracy kinetic mechanisms for hydrogen and hydrogen-methane combustion remains a relevant topic for research. The use of machine learning for optimization of reaction mechanisms is a promising approach to achieve the accuracy and computational cost required by the application. For industrial CFD applications, novel single-step reaction mechanisms with increased accuracy can significantly increase the fidelity of results, while keeping computational cost reasonable. For hydrogen-methane blends, reaction mechanisms valid for wide ranges of blends should be derived. This will ensure high fidelity and reasonable computational cost of CFD studies on hydrogen blending for industrial furnaces.

Overall, LES-FGM models are the most widely used in the recent literature. These models have shown promising accuracy in reference flames. Nevertheless, their computational cost remains excessive for most studies on industrial furnaces, as they require a computational mesh size that is not feasible to simulate within industrial furnace design timelines. Accurate models with lower computational cost, improvements in numerical solving methods, and the use of high-performance computing should be developed to transfer the small-scale CFD modeling capabilities to the industrial framework. An alternative to these highly detailed CFD models is surrogate models, which have been developed and assessed in recent years for hydrogen combustion [137,138,139]. Methods based on Flamelet Manifold Neural Networks have achieved relative errors below 2% while reducing computational time by a factor of 4 compared to FGM tables [140].

Most hydrogen CFD studies in the literature are focused on the detailed modeling and individual evaluation of turbulence, chemistry, turbulence-chemistry interaction, or heat transfer models. To keep computational cost reasonable, simplifications are made on the rest of the models, which limit the overall accuracy of the complete model. Coupled modeling and assessment of all these phenomena is yet to be adequately covered in the literature.

Diffusion and flame stability phenomena are highly relevant in turbulent hydrogen flames modeling. Most CFD models for existing fuels rely on assumptions that do not apply to hydrogen combustion. While DNS can predict these phenomena, LES and RANS models need explicit accounting for them. Future research should incorporate these effects into CFD models, so that they can be used for detailed burner and furnace design.

While CFD models predict global flame temperature well for hydrogen combustion, local wall heat flux is not yet predicted accurately by most models. While recalibration of WSGG radiation models has been performed for combustion of hydrogen and hydrogen-methane blends, comprehensive validation of these adapted models shall be performed for broader operating temperature and pressure ranges. Further work on convective and radiative heat transfer predictive capabilities on CFD models of hydrogen and hydrogen-methane flames is required to use these models for predicting heat transfer industrial loads. This development is key for the use of CFD as a support design tool for retrofitting industrial furnaces with the use of hydrogen.

6. Hydrogen Combustion Monitoring Using Computer Vision and Artificial Intelligence

Camera-based combustion monitoring has become an important tool in hydrogen combustion research because it provides spatially distributed, time-resolved and non-intrusive information which is difficult to obtain from other measurement probes alone [141,142,143]. This capability is particularly relevant for hydrogen flames, whose weak visible luminosity, high diffusivity, fast propagation and sensitivity to thermo-diffusive effects complicate the interpretation of conventional local measurements [144,145,146]. Figure 10 presents the temporal trend in scientific publications on hydrogen combustion focused on flame monitoring, highlighting the growing research interest in optical diagnostics and artificial intelligence-based approaches between 2019 and 2025.

Figure 10. Annual hydrogen combustion publications related to flame monitoring based on optical features and artificial intelligence (2019–2025).

In experimental systems, such as pilot furnaces [147], dual fuel engines [142] and burners of gas turbines [148], imaging enables the simultaneous observation of flame stabilization, global morphology, heat release patterns and flow interactions.

In multi-element [149], micromix multi-nozzle array [150] and scramjet combustors [151], high-speed optical measurements are especially valuable because flame topology, flamelet interaction, and instability growth can evolve over short time scales and across strongly three-dimensional flow fields.

Pure hydrogen flames emit much weaker broadband visible radiation than sooting hydrocarbon flames, so visible imaging must be interpreted with careful attention to sensor sensitivity, exposure, background radiation, and spectral selection [141,144]. However, visible imaging remains useful when hydrogen is blended with methane, ammonia, carbon monoxide, or natural gas because the resulting flame morphology can still reveal changes in propagation and stability [152], stretch sensitivity [153], and fuel–air mixing [154].

Beyond the visible spectrum, OH* chemiluminescence is widely used as an optical marker of the reaction zone and heat-release dynamics, although its intensity depends on local chemistry, pressure, temperature, mixture composition, optical collection, and line-of-sight integration [143,155,156]. On the other hand, OH Planar Laser Induced Fluorescence (OH-PLIF) provides a more spatially localized description of the reaction layer than passive chemiluminescence because it interrogates a laser sheet rather than collecting path-integrated spontaneous emission [151,157,158,159]. Particle Image Velocimetry (PIV) extends flame imaging from morphology to flow-field quantification by linking flame-front deformation, vortex dynamics, recirculation, and local strain to measured velocity fields [160,161]. The combined use of OH-PLIF, OH* chemiluminescence, and PIV implies multimodal measurements that connect reaction-zone topology, heat-release dynamics, and flow structure within a single experimental framework [150].

Infrared imaging complements visible and ultraviolet diagnostics by revealing hot regions, wall interaction and thermal evolution that may not be apparent from the weak hydrogen flame luminosity alone [142,162]. Moreover, multispectral imaging can also be used for hydrogen combustion diagnostics, as visible luminosity [144], radical emission [142], and thermal radiation [163] provide complementary information about the flame state. Other techniques, like Schlieren [145,152] and shadowgraph imaging [164], are particularly suitable for weakly luminous hydrogen flames because they visualize density-gradient fields without relying on radiative flame emission.

Regarding the equipment configuration, the required camera frame rate must be matched to the characteristic time scale of the phenomenon, since ignition and flame acceleration [155], thermoacoustic response and vortex–flame interaction [149], and supersonic combustion [151] may occur at millisecond scales.

Digital image processing converts raw optical recordings into reproducible observables through background correction, normalization, filtering, segmentation, contour extraction, and feature calculation [158,165,166,167]. Geometric descriptors such as projected area, flame length, centroid position, lift-off height, perimeter, curvature, and eccentricity provide compact indicators of flame stabilization, spreading, wrinkling, and dynamic response [149,150,168]. In spherical hydrogen–air flames, image-based reconstruction of cellular structures is essential because thermo-diffusive instabilities can increase flame surface area, accelerate burning, and modify the apparent propagation rate [145,165,168].

Time-resolved flame contours allow the extraction of propagation velocity, growth rate, oscillation frequency, and phase relationships with acoustic or pressure signals, which are key quantities for studying combustion instability [149,151,155]. Three-dimensional reconstruction methods address the limitations of single-view projections by recovering volumetric flame structure or emission fields from multi-view images and tomographic algorithms [156,169,170].

Visual information can also be used to infer combustion variables that are not directly measured by the camera, provided that sufficient prior data, calibration, physical modeling, or statistical learning methods are available [162,171,172]. Classical optical inference relies on known relationships between radiation and physical variables. For example, two-color pyrometry estimates temperature from radiation intensities measured at two spectral bands, assuming an appropriate emissivity model and optical calibration [162]. Chemiluminescence imaging can be used as an indicator of heat-release distribution, although the relationship between radical emission and heat release is not universal and must be interpreted with care [155,171,172]. Similarly, OH-PLIF provides information about reaction-zone structure but not directly about heat release, species concentration, or temperature unless additional calibration and correction procedures are applied [143,158,159,173]. These examples illustrate an important point: camera images are rich in information, but the measured intensity is not automatically equal to the desired combustion variable. A physically meaningful interpretation requires understanding the optical signal, the chemistry of the emitting or fluorescing species, the camera response, and the experimental geometry.

Machine learning has emerged as a powerful alternative for extracting latent information from flame images and for modeling nonlinear relationships between visual signatures and combustion variables. Machine learning is a branch of artificial intelligence that learns input-output relationships from data rather than relying exclusively on explicit analytical models. In combustion diagnostics, the inputs may include raw images, processed contours, intensity maps, spectral bands, temporal sequences, acoustic signals, pressure traces, or multimodal combinations of several sensors [165,168,171,172,174,175,176]. The outputs may include fuel-blend proportion, combustion regime, stability class, flame surface area, NO_x concentration, temperature-related indicators, reconstructed flame fields, or future flame evolution [165,168,171,172,174,175,176]. The appeal of machine learning lies in its capacity to handle high-dimensional data and nonlinear dependencies that are difficult to represent with simple correlations. Hydrogen combustion is a particularly suitable field for these methods because flame structure depends on coupled chemical, transport, acoustic, and turbulent processes. However, this same complexity also makes validation essential: a model trained under one burner geometry, camera exposure, fuel composition, or pressure condition may not generalize to another configuration.

Recent studies illustrate the rapid expansion of image-based and multimodal machine learning in hydrogen combustion. Classification models have been developed for premixed flames with different hydrogen-natural gas proportions using image and audio data [174]. In such cases, the model does not merely detect whether a flame is present; it attempts to identify fuel-composition classes or combustion regimes from observable signatures. This is relevant for industrial monitoring because changes in hydrogen blending can modify flame speed, stability, acoustics, and emissions. Other works have used deep learning to analyze cellular structures in spherical premixed hydrogen-air flames, where automatic extraction of flame-front features can support quantitative descriptions of instability development [165,168]. In lean direct injection systems, flame imaging combined with statistical and machine learning models has been used to predict NO_x emissions from hydrogen and hydrogen/natural gas blends [171]. This is scientifically important because NO_x is not directly visible in ordinary flame images, but image features may contain indirect information about reaction intensity, hot zones, and operating conditions that influence thermal NO_x formation. In scramjet combustors, neural networks have been applied to predict flame development and reconstruct flame fields from multi-source information, addressing highly transient and compressible combustion environments [169,175,176]. These works show that imaging is no longer limited to post-experiment visualization; it can become part of predictive monitoring, control, and digital-twin frameworks.

Despite these advances, several limitations must be considered. First, image-based models can be sensitive to camera settings, lens aperture, exposure time, optical filters, viewing angle, background illumination, and sensor saturation. A machine learning model may therefore learn experimental artifacts rather than physical combustion features if the dataset is not carefully designed. Second, flame images often provide projected information, so line-of-sight integration can hide three-dimensional effects. Third, training datasets may be unbalanced or too narrow, causing poor extrapolation outside the tested operating range. It is therefore important to emphasize that machine learning cannot, by itself, reconstruct three-dimensional flame information that is physically absent from a single line-of-sight projection: when the training data are limited to 2D views, the model can only infer 3D structure indirectly through statistical correlations learned from configurations where volumetric ground truth was available. Reliable compensation thus requires that the training set itself incorporates volumetric information—obtained from multi-view imaging, tomographic reconstruction, or high-fidelity simulations (e.g., DNS/LES)—so that the network learns physically consistent projection-to-volume relationships rather than dataset-specific artifacts. Consequently, the robustness of any image-based model is bounded by the diversity and dimensionality of its training data, reinforcing the need for physics-informed learning, uncertainty quantification, and cross-configuration validation. Fourth, many high-performing deep learning models are difficult to interpret, which limits their usefulness for scientific explanation. For this reason, future progress should combine data-driven methods with physical constraints, uncertainty quantification, interpretable features, calibration standards, and cross-validation across different burners and operating conditions. Hybrid approaches are especially promising: image processing can extract physically meaningful variables, machine learning can model nonlinear relationships, and optical diagnostics can provide independent validation data. In this context, the state of the art is moving from qualitative flame observation toward calibrated, multispectral, high-speed, three-dimensional, and data-driven optical diagnostics capable of supporting combustion characterization, emission prediction, instability analysis, and real-time control of hydrogen combustion systems.

7. Conclusions, Main Challenges and Recommendations

This paper reviewed the state-of-the-art of hydrogen combustion in industrial furnaces from three complementary perspectives: its fundamental combustion properties and industrial deployment, its computational modeling through CFD, and its monitoring through computer vision and artificial intelligence. The study of this subject covers multiple levels, ranging from the physicochemical characterization of hydrogen flames and their behavior under different combustion regimes, through the specific constraints of sectors such as steelmaking, cement, glass, ceramics and refining, to the modelling and diagnostic tools required to design and operate hydrogen-fired systems. The three research axes examined here are strongly interdependent yet remain largely disconnected in the literature. Future research should prioritize this integration.

The main conclusions and recommendations are summarized below.

Hydrogen as an Industrial Fuel

• The combustion properties of hydrogen differ substantially from those of natural gas. Its laminar flame speed is approximately six to eight times higher than that of methane, its minimum ignition energy is approximately sixteen times lower, and its adiabatic flame temperature is about 150 °C higher under stoichiometric air combustion conditions. These differences drive the two main technical barriers: flashback in premixed configurations and increased thermal NO_x emissions. These barriers cannot be addressed by linear scaling of existing natural gas burners.
• Deployment maturity depends largely on the sector. The steel industry has tested hydrogen as fuel for reheating processes. Glass, cement and ceramics remain at pilot or demonstration scale. Industrial implementation of hydrogen as fuel for industrial thermal processing remains constrained by the modified flame heat profile and its impact on product quality.
• Hydrogen and natural gas blends provide a transitional pathway, but their combustion behavior changes non-linearly with hydrogen fraction. Blends up to approximately 20% by volume are generally compatible with existing infrastructure, whereas higher fractions require burner redesign, material qualification and specific NO_x mitigation strategies.
• Materials compatibility, in particular hydrogen embrittlement of ferritic and martensitic steels and refractory corrosion under high H₂O partial pressure, remains a prerequisite for safe deployment and requires standardized qualification protocols that are still under development at European and international level.
• Economic rather than technical viability is the primary barrier to industrial adoption. The delivered cost of hydrogen per unit of energy currently exceeds that of natural gas by a large factor, dominated by the electricity price and the electrolyzer capacity factor, whereas the capital cost of the combustion equipment itself is comparatively modest. Fuel switching becomes competitive only above carbon-price levels well beyond current EU allowance prices, so deployment will depend on the joint evolution of hydrogen cost, carbon pricing and targeted support instruments rather than on combustion performance alone.

CFD modeling of hydrogen combustion

• Closure models developed for hydrocarbon fuels frequently rest on assumptions that do not hold for hydrogen. The review of these assumptions and subsequent model adaptation brings LES and RANS predictions closer to DNS and experimental results. This is essential for using CFD as a tool for industrial furnace design. However, closure model calibration requires adequate benchmarking and validation, which is not yet widely available. This remains a research gap for future work.
• LES-FGM is currently the most widely used and best-validated approach. Its computational cost remains prohibitive for industrial-scale domains at high Reynolds numbers. Reduced-cost models, improved numerical solvers and high-performance computing are required to transfer small-scale modeling capability to industrial design timelines. The use of surrogate models and Flamelet Manifold Neural Networks is a promising research direction to fulfill this need.
• Radiative heat transfer models, such as WSGGM, were originally fitted for hydrocarbon combustion species. When considering water vapor as the main product, as for hydrogen combustion, these models should be recalibrated for several pressure ranges and blend ratios. Further, comprehensive validation against industrial furnace data should be performed to ensure the accuracy of radiative heat transfer calculations for hydrogen combustion. This is a key step for the applicability of CFD for heat transfer modeling in industrial furnaces with hydrogen combustion.
• Coupled modeling of turbulence, TCI, diffusion phenomena, flame stability, and heat transfer to industrial loads should be tackled in future research.
• For model benchmarking and validation, comprehensive, standardized datasets on hydrogen combustion shall be developed, based on canonical flames, lab-scale and semi-industrial scale combustion setups. The recent TNF initiative should continue its work and expand to additional research groups to become a standard benchmark for hydrogen combustion model cross-validation and calibration.
• CFD simulation of hydrogen and hydrogen-methane combustion in full industrial furnaces remains a key research direction that requires further work.

Combustion monitoring through computer vision and AI

• Camera-based monitoring provides spatially distributed, time-resolved, non-intrusive measurements. This information is particularly valuable for hydrogen flames, whose weak visible luminosity and sensitivity to thermo-diffusive effects complicate conventional local measurements. UV detection of OH* chemiluminescence has proven a robust diagnostic where visible imaging fails.
• Machine learning and deep learning architectures have shown high accuracy in flame classification, regime identification and anomaly detection under controlled conditions, but their specific application to hydrogen combustion remains an emerging and fragmented field, as reflected by the comparatively small body of literature identified in this review.
• The main barrier to industrial deployment is the scarcity of open, curated, multi-modal datasets specific to hydrogen combustion under representative furnace conditions, together with the limited validation of these systems in the harsh thermal and optical environment of real furnaces.
• Because research gaps are identified across the industrial-application, CFD-modeling and combustion-monitoring sections, Table 2 consolidates the main barriers discussed in this review and links each of them to a recommended research action. This synthesis is intended to support future experimental campaigns, model-development efforts and research calls focused on industrial hydrogen combustion.

Table 2. Consolidated research gaps and recommended research actions for hydrogen combustion in industrial furnaces.

Research Gap	Specific Barrier Identified	Recommended Research Action	Related References
Industrial-scale validation of hydrogen combustion in furnaces	Most available evidence remains limited to laboratory studies, pilot trials or sector-specific demonstrators, while long-duration full-scale furnace data is still scarce.	Develop industrial validation campaigns under representative production loads, furnace atmospheres and operating constraints, including direct assessment of product quality, emissions and operational stability.	[19,20,23,37,38,50]
Sector-specific product-quality impacts	In glass and ceramics, the lower luminosity of hydrogen flames and modified radiative heat transfer may affect melting homogeneity, temperature uniformity, surface quality and defect formation.	Couple hydrogen-firing trials with product-quality metrics, defect tracking, spatial heat-flux measurements and furnace-temperature mapping.	[19,20,23,49,51]
Retrofit complexity and burner compatibility	Hydrogen substitution is not a simple fuel replacement because flame speed, Wobbe index, flashback risk, heat release and burner stability change non-linearly with H2 fraction.	Define sector-specific retrofit envelopes for burners, valves, control systems, refractories, fuel trains and safety systems under increasing hydrogen fractions.	[23,37,38,50,52]
NOx control under hydrogen and H2/NG blends	Higher flame temperature and changes in residence time can increase thermal NOx, while mitigation effectiveness depends strongly on burner geometry and furnace configuration.	Test combined NOx-abatement strategies, including staged combustion, exhaust-gas recirculation, MILD/flameless combustion, oxy-fuel operation and residence-time control.	[24,25,26,27,50,52]
Heat-transfer equivalence between natural gas and hydrogen firing	Thermal equivalence at constant fuel power does not guarantee equivalent radiative and convective heat-flux profiles to the load.	Generate validated datasets comparing natural gas, H2/NG blends and pure H2 under equivalent operating conditions, including radiative properties and wall/load heat fluxes.	[23,49,51]
CFD validation at semi-industrial and industrial scale	CFD studies are extensive, but validation against semi-industrial or full industrial hydrogen-fired furnace measurements remains limited.	Develop open benchmark cases with geometry, boundary conditions, temperature fields, heat fluxes, species concentrations, velocity data and NOx measurements.	[25,33,34,38]
Selection of CFD models for industrial furnace applications	High-fidelity methods improve physical description but remain computationally demanding for full-furnace design, optimization and control.	Compare RANS, LES, flamelet, finite-rate chemistry, reduced-order and surrogate models using common industrial benchmark cases and clearly reported computational costs.	[33,34]
Uncertainty quantification in hydrogen combustion modeling	Kinetic, turbulence-chemistry, radiation, boundary-condition and operating uncertainties can strongly affect predicted temperature, flame stability and NOx emissions.	Integrate uncertainty quantification into CFD workflows and report sensitivity analyses and prediction intervals rather than only deterministic results.	[25,133,134]
Lack of hydrogen-specific optical and AI datasets	Open, curated and multimodal datasets for hydrogen flames under representative furnace conditions remain scarce, limiting the generalization of AI-based monitoring.	Create shared datasets combining visible imaging, UV/OH* chemiluminescence, infrared imaging, multispectral imaging, PIV/PLIF and synchronized process data.	[35,36,39,139,147,152,153,154,155,156,157,158,159,160,161]
Robustness of image-based monitoring in industrial environments	Camera-based and AI methods often perform well under controlled conditions but may not generalize to high-temperature, optically harsh and geometry-specific furnace environments.	Validate models across burners, fuels, loads, viewing angles, optical access conditions and furnace atmospheres, including uncertainty estimation and out-of-distribution detection.	[35,36,39]
Safety and material compatibility	Hydrogen leakage, wide flammability limits, low ignition energy, embrittlement and refractory interactions require additional qualification compared with natural gas systems.	Combine combustion trials with materials testing, leakage detection, sensor validation, ventilation assessment and safety-case development for industrial retrofits.	[28,29,30,41,42]
Integrated techno-economic and deployment assessment	Deployment feasibility depends not only on combustion performance, but also on hydrogen availability, CAPEX/OPEX, retrofit complexity, NOx compliance, safety upgrades and product-quality risk.	Perform sector-specific techno-economic assessments linking emissions reduction, hydrogen supply, retrofit cost, operational risk, product quality and commercialization pathway.	[1,5,7,8,21,29,37]

Cross-cutting challenges and recommendations

Based on this review, we identify four priorities for future work:

• Coupled experimental campaigns that simultaneously characterize heat flux, emissions and flame structure across hydrogen fractions in single furnace geometries.
• Reduced-cost, fully coupled CFD models with validated hydrogen-specific transport and radiation submodels, applicable to industrial-scale furnaces.
• Open multi-modal datasets and physics-informed learning architectures for hydrogen flame monitoring.
• Integrated studies involving experimental, computational, and monitoring activities on hydrogen flames, with interrelated conclusions and validation of the approaches.