HiPrFlame: Ab Initio High-Pressure Combustion

• HiPrFlame is an ab initio modeling framework that uses a third-order Virial equation of state and Enskog theory to capture real-fluid effects in high-pressure combustion.
• It replaces empirical methods with physics-based formulations and deep neural network surrogates trained on extensive property data to enable efficient CFD simulations.
• Demonstrations on hydrogen autoignition and laminar premixed flames show that HiPrFlame significantly improves predictive accuracy by addressing non-ideal thermodynamic and transport behaviors.

HiPrFlame is an ab initio-based modeling framework for high-pressure combustion intended for regimes in which operating pressures often exceed the critical point of working fluids and real-fluid effects fundamentally alter thermodynamic and transport properties. It integrates a third-order Virial equation of state derived from ab initio intermolecular potentials, real-fluid departure functions for thermodynamics, and Enskog theory for transport properties within an OpenFOAM architecture that can be used for 0-D to 3-D real-fluid modeling. To accelerate multidimensional simulations, the framework also employs artificial neural network surrogates trained on a comprehensive property database for efficient real-fluid property updating during CFD (Zhang et al., 8 Sep 2025).

1. Scientific setting and motivation

High-pressure combustion is central to modern propulsion and power-generation systems, and the motivating examples given for HiPrFlame include jet engines operating above $4\ \mathrm{MPa}$ and rocket chambers up to $30\ \mathrm{MPa}$. In such regimes, fluids exhibit strong real-fluid effects, with non-ideal thermodynamic and transport behaviors that profoundly influence combustion performance and emissions. The framework was introduced in response to the view that existing methods often use empirical equations of state such as SRK and Peng–Robinson, together with empirical transport correlations, and that these methods lack the accuracy, generality, and predictive capability required at high pressures or for novel fuel mixtures (Zhang et al., 8 Sep 2025).

Within this formulation, HiPrFlame is positioned as a more fundamental and broadly applicable alternative. The stated objective is not merely numerical convenience, but predictive real-fluid property evaluation based on quantum chemistry and statistical mechanics. This is particularly significant because the same summary reports that real-fluid effects can be substantial even below the critical pressure of air, so the motivating regime is broader than a narrowly defined supercritical limit (Zhang et al., 8 Sep 2025).

2. Molecular and thermophysical formulation

The thermodynamic core of HiPrFlame is a third-order Virial equation of state,

$\frac{p}{\rho RT} = 1 + B(T)\rho + C(T)\rho^2,$

where $B(T)$ is the second Virial coefficient and $C(T)$ is the third Virial coefficient. In the framework description, these coefficients are generated from ab initio quantum-chemistry potentials for key combustion species, including $\mathrm{H_2}$, $\mathrm{O_2}$, $\mathrm{N_2}$, and $\mathrm{H_2O}$. The same source states that higher-order terms saturate accuracy beyond third order, making the third-order truncation the operative choice (Zhang et al., 8 Sep 2025).

Thermodynamic state functions are then obtained through real-fluid departure functions. For enthalpy, the formulation is written as

$h = h_{ig} + \Delta h,$

with

$30\ \mathrm{MPa}$0

This is the mechanism by which the framework augments ideal-gas thermochemistry with real-fluid corrections (Zhang et al., 8 Sep 2025).

Transport properties are handled through Enskog theory, described as a non-empirical, molecular-based method derived from first principles via the Boltzmann equation. The reported constitutive forms are

$30\ \mathrm{MPa}$1

$30\ \mathrm{MPa}$2

$30\ \mathrm{MPa}$3

Here the variables $30\ \mathrm{MPa}$4 and $30\ \mathrm{MPa}$5 are functions of the Virial coefficients and their temperature derivatives. The framework description also states that, when Enskog theory is coupled with ab initio-based third-order Virial coefficients, real-fluid transport predictions match experiments within $30\ \mathrm{MPa}$6 error, compared with $30\ \mathrm{MPa}$7 for the empirical Chung method (Zhang et al., 8 Sep 2025).

Element	Basis	Function
Thermodynamics	Third-order Virial EoS from ab initio potentials	Real-fluid equation of state
Departure functions	Virial-based thermodynamic corrections	Enthalpy, heat capacity, density correction
Transport	Enskog theory	Viscosity, thermal conductivity, mass diffusivity
Mixture treatment	Physically motivated mixing rules	High-pressure combustion mixtures

The framework further states that it uses physically motivated mixing rules rather than simple mass-averaging to predict properties of combustion mixtures at high pressure (Zhang et al., 8 Sep 2025).

3. Surrogate acceleration and software architecture

A central feature of HiPrFlame is the replacement of direct on-the-fly evaluation of the ab initio property model by ANN surrogates. The reported workflow begins with a property database generated through Latin Hypercube Sampling over species composition, temperature, and pressure. More than $30\ \mathrm{MPa}$8 million data points were computed with the Virial-based methodology and used to train a deep neural network with five hidden layers and $30\ \mathrm{MPa}$9 nodes per layer, implemented in PyTorch and optimized with Huber loss and Adam (Zhang et al., 8 Sep 2025).

The surrogate takes temperature, pressure, and composition as inputs and returns thermodynamic and transport properties as outputs. The reported surrogate accuracy is below $\frac{p}{\rho RT} = 1 + B(T)\rho + C(T)\rho^2,$0 error in density and below $\frac{p}{\rho RT} = 1 + B(T)\rho + C(T)\rho^2,$1 in thermal diffusivity compared to NIST data. In the context of the paper, this ANN layer is what makes real-time, efficient property evaluation during CFD feasible, and thus unlocks multidimensional high-fidelity simulations that would otherwise be computationally impractical (Zhang et al., 8 Sep 2025).

HiPrFlame is implemented in OpenFOAM v7.0. The architecture supports 0-D, 1-D, 2-D, and 3-D real-fluid combustion simulations, and uses front-end fields for all properties, which are updated by the surrogate through a libtorch interface at each iteration. The reported numerical setup includes second-order upwind or central schemes for convection and shear, Euler time advancement, PISO for pressure–velocity coupling, and a highly resolved 1-D flame mesh with $\frac{p}{\rho RT} = 1 + B(T)\rho + C(T)\rho^2,$2 cells. The solver is also described as compatible with standard OpenFOAM boundary conditions, which permits transition from homogeneous autoignition calculations to multidimensional flame simulations without changing the core property infrastructure (Zhang et al., 8 Sep 2025).

4. Demonstration on hydrogen autoignition and laminar premixed flames

The initial demonstrations reported for HiPrFlame are homogeneous autoignition and one-dimensional laminar premixed hydrogen flames. In the 0-D autoignition study, the conditions are stoichiometric $\frac{p}{\rho RT} = 1 + B(T)\rho + C(T)\rho^2,$3 at $\frac{p}{\rho RT} = 1 + B(T)\rho + C(T)\rho^2,$4 and $\frac{p}{\rho RT} = 1 + B(T)\rho + C(T)\rho^2,$5. The paper summary states that at $\frac{p}{\rho RT} = 1 + B(T)\rho + C(T)\rho^2,$6, real-fluid and ideal-gas models yield similar ignition delay times and both match experiment, indicating minimal real-fluid influence under those conditions. At $\frac{p}{\rho RT} = 1 + B(T)\rho + C(T)\rho^2,$7, however, real-fluid effects become significant below $\frac{p}{\rho RT} = 1 + B(T)\rho + C(T)\rho^2,$8, where they reduce ignition delay times relative to ideal-gas predictions (Zhang et al., 8 Sep 2025).

The laminar premixed flame study uses $\frac{p}{\rho RT} = 1 + B(T)\rho + C(T)\rho^2,$9 at $B(T)$0 and $B(T)$1. At $B(T)$2, HiPrFlame and ideal-gas models both match experiments and each other, again indicating negligible real-fluid influence. At $B(T)$3, the reported result is qualitatively different: HiPrFlame predicts much higher laminar flame speeds than ideal-gas modeling, with an over $B(T)$4 increase at near-stoichiometric conditions, and these predictions match or surpass the experimental flame-speed data while the ideal-gas model severely underpredicts them (Zhang et al., 8 Sep 2025).

A further comparative result is especially important for model discrimination. The framework description states that simulations using Lennard-Jones versus ab initio potentials show that only ab initio-based models with a third-order Virial EoS properly capture real-fluid effects. That statement is used to argue not merely for non-ideality correction in general, but for the specific molecular and statistical-mechanical closure adopted in HiPrFlame (Zhang et al., 8 Sep 2025).

5. Relation to earlier high-pressure combustion frameworks

HiPrFlame emerged against a background of high-pressure combustion models built around flamelet formulations and cubic equations of state. One such earlier framework addressed high-pressure LOx/$B(T)$5 and LOx/hydrocarbon combustion for propulsion systems by embedding detailed chemistry in a RANS solver with a Low Reynolds number $B(T)$6-$B(T)$7 turbulence model and an extended flamelet/progress-variable formulation. In that model, all thermochemical quantities are evaluated by evolving the mixture fraction $B(T)$8 and a progress variable $B(T)$9, and a statistically most likely distribution approach was proposed to evaluate the joint PDF of $C(T)$0 and $C(T)$1 without assuming their functional form a priori (Coclite et al., 2014).

A later development extended the FPV approach to transcritical real-fluid combustion in fully compressible, explicit finite-volume solvers. That framework used the Peng–Robinson cubic equation of state, introduced a double-flux model to eliminate spurious pressure oscillations, formulated a hybrid scheme with entropy-stable flux correction to represent large density ratios, and pre-tabulated the parameters needed for real-fluid departure functions while using a quadratic expression to recover the attraction parameter. It was described as directly applicable to LES of real applications under transcritical conditions (Ma et al., 2017).

Against this background, HiPrFlame represents a different modeling choice. Rather than improving a cubic-EoS-centered workflow through new tabulation and numerical stabilization strategies, it replaces the property model itself by an ab initio-based third-order Virial EoS and Enskog transport closure, and then restores computational tractability through ANN surrogates in OpenFOAM (Zhang et al., 8 Sep 2025). A plausible implication is that HiPrFlame shifts the fidelity bottleneck from tabulation strategy and empirical parameterization toward the quality of the underlying molecular interaction data.

6. Scope, validation context, and recurring misconceptions

A recurring misconception in high-pressure combustion is that real-fluid effects become important only once operating conditions are clearly supercritical. The HiPrFlame laminar flame study explicitly argues otherwise, stating that real-fluid effects are already substantial below the critical pressure of air, not just at supercritical conditions (Zhang et al., 8 Sep 2025). A second misconception is that empirical cubic equations of state and fitted transport correlations are sufficient across species and extreme conditions. The rationale for HiPrFlame is explicitly framed as a response to the inadequacy of such methods for species coverage and extreme combustion conditions (Zhang et al., 8 Sep 2025).

The framework also sits within a broader high-fidelity combustion research environment. Direct Numerical Simulation of a premixed $C(T)$2-heptane flame at high Karlovitz number, $C(T)$3, was described as fully resolving all relevant spatial and temporal scales of turbulence and chemical reactions and offering a high-fidelity picture critical for HiPrFlame research. In that DNS, the progress-variable field became more diffuse than in a laminar flame, and the flame front showed significant wrinkling and thickening as turbulence penetrated the preheat and reaction zones (Bobbitt et al., 2013). More recently, DNS of turbulent lean premixed hydrogen flames at $C(T)$4, $C(T)$5, and $C(T)$6 found measurable differences in flame structure and turbulence-flame coupling driven by reduced turbulence dissipation within the flame at elevated pressure and temperature, while reporting that the surface-averaged tangential strain normalized by the Kolmogorov time remains approximately universal at $C(T)$7 (Kassar et al., 22 Jan 2026).

These surrounding studies help delimit the present scope of HiPrFlame. The cited HiPrFlame paper demonstrates the framework on homogeneous autoignition and one-dimensional laminar premixed flames, not yet on turbulent DNS or LES benchmarks (Zhang et al., 8 Sep 2025). This suggests that its primary encyclopedic significance lies in establishing a first-principles property model and computational pathway for high-pressure combustion, with turbulent and multidimensional validation forming the next logical stage rather than the completed one.

7. Significance for high-pressure combustion modeling

The framework description characterizes HiPrFlame as the first integration of ab initio-based third-order Virial EoS and molecular Enskog theory for both thermodynamics and transport in combustion CFD, and also as the first ANN-accelerated, OpenFOAM-based 0D–3D real-fluid capability built on that basis (Zhang et al., 8 Sep 2025). Within the paper’s own evidence, the significance of this integration is twofold: it improves property prediction relative to traditional empirical models, and it changes predicted combustion behavior in cases where ideal-gas modeling would imply qualitatively different reactivity.

For researchers in combustion, propulsion, and CFD, the importance of HiPrFlame therefore lies less in a single numerical algorithm than in a particular synthesis of molecular theory, surrogate modeling, and solver implementation. Its core claim is that accurate high-pressure combustion simulation requires simultaneous fidelity in thermodynamics and transport, and that this fidelity can be made computationally usable through database-driven ANN surrogates rather than by reverting to empirical equations of state or transport correlations (Zhang et al., 8 Sep 2025).