Thermodynamics-Informed Input Reparameterization for Neural Prediction of Real-Fluid Thermodynamic Properties in Supercritical Combustion

arXiv:2607.19241 · cs.LG, physics.comp-ph, physics.flu-dyn · Submitted 2026-07-21 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Thermodynamics-Informed Input Reparameterization for Neural Prediction of Real-Fluid Thermodynamic Properties in Supercritical Combustion".

Jane: The gist The work introduces a thermodynamics-informed input reparameterization strategy, termed target-aligned input reparameterization (TAIR),

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, to wrap up this paper on "Thermodynamics-Informed Input Reparameterization for Neural Prediction of Real-Fluid Thermodynamic Properties in Supercritical Combustion", the main thing they proved is that replacing the raw enthalpy input with target-matched coordinates like T-tilde and rho-tilde significantly reduces prediction errors.

Jane: They achieved a reduction in held-out test-set root mean squared error by factors of approximately one point five for temperature, two point zero for density, and seven point five for compressibility <ref:2607.19241#pg3>. That’s a tangible improvement over the raw-input model when tested on supercritical methane–oxygen counterflow flame data against target-inconsistent controls <ref:2607.19241#pg3>.

Lu: The implication is that this deterministic preconditioning of the input coordinates helps neural networks handle real fluid departures more effectively <ref:2607.19241#pg3>. It shows that matching the input to the target's ideal-gas dependence yields better training losses under the same network setup <ref:2607.19241#pg3>.

Meng: For us on the production side, if we can achieve a seven point five times reduction in error for compressibility, that’s a lot of computational savings when we scale up these simulations <ref:2607.19241#pg3>. But we have to be careful because they mentioned their analysis only characterizes one representative state in each temperature bin <ref:2607.19241#pg3>.

Tom: Exactly. It’s a strong result based on targeted input design, but it’s not a magic fix for every single physical situation without further tuning and validation <ref:2607.19241#pg3>. That’s the core of what they demonstrated with this TAIR approach.

Conclusion: Tom: So, we're wrapping up our look at this paper, "Thermodynamics-Informed Input Reparameterization for Neural Prediction of Real-Fluid Thermodynamic Properties in Supercritical Combustion." Basically, they figured out a way to make those neural networks better at predicting what things are like when they aren't ideal gases.

Jane: Yeah, the title sounds super technical, but the core idea is about fixing how you feed the data into the model. They took raw enthalpy and replaced it with something called a target-matched coordinate.

Lu: What that means practically is they’re not feeding in just one raw number for everything; they're using a math transformation based on fluid physics to guide the network toward what it *should* be learning, which is the real fluid behavior.

Meng: From my side, I'm thinking about deployment. If we can get that kind of accuracy improvement, it means we might cut down on running those expensive iterative thermodynamic calculations a lot less often before hitting our production models.

Tom: Exactly. The results show they got these big gains in error reduction—like a factor of seven for compressibility—and the overhead on the actual compute is tiny, just about eight point eight percent more time for the whole test set run.

Jane: And it’s important because they showed this method works even when you introduce different strain rates or different flame structures, not just in one specific test case.

Lu: They interpret this as a deterministic preconditioning of the input coordinates; they aren't adding new physical information, but they are exposing the leading ideal-gas dependence so the network can focus on learning the departures from that baseline.

Tom: So what it changes for us is that we don't just rely on whatever raw data we feed in; we engineer a better input coordinate based on how thermodynamics actually works.

Jane: It opens up a path where you can build more robust prediction tools for these complex combustion processes because the model is essentially getting a head start on the physics.

Meng: It’s promising, but I still need to see how stable that transformation is if we try it on something completely outside their augmented envelope; right now, it’s proven across those conditions.

Lu: That's the next hurdle; proving transferability beyond the tested conditions is where the real research goes next.

Haoze Zhang, Han Li, Ke Xiao, Yangchen Xu, Runze Mao, Zhi X. Chen

State Key Laboratory of Turbulence and Complex Systems, Peking University · AI for Science Institute (AISI)

cs.LG, physics.comp-ph, physics.flu-dyn

Submitted: 2026-07-21

Updated: 2026-07-21

Journal ref: Physics of Fluids 38(10), 107502 (2026)

DOI: 10.1063/5.0352874

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 81/100

The gist: The gist The work introduces a thermodynamics-informed input reparameterization strategy, termed target-aligned input reparameterization (TAIR), to improve neural prediction of real-fluid

Key concepts

Target-Aligned Input Reparameterization (TAIR): TAIR is a strategy that modifies the input coordinate used by neural networks. Instead of using raw enthalpy for all properties, it replaces it with a specific thermodynamic coordinate tailored to each target property, guiding the network to learn real-fluid deviations more effectively.
Computational Cost of Cell-Local Thermodynamic Closure: Evaluating the full thermodynamic closure is computationally expensive because it requires repeated calculations involving real-fluid state evaluations. TAIR avoids this by using simplified ideal-gas estimates for temperature and density, which are derived from readily available solver variables.
Target-Matched Thermodynamic Coordinate: This is a specific coordinate transformation applied to the raw enthalpy. For example, the temperature network uses a transformed coordinate that neglects real-fluid enthalpy departure and approximates species heat capacities as constants, focusing on the leading ideal-gas dependence.

Summary

The gist The work introduces a thermodynamics-informed input reparameterization strategy, termed target-aligned input reparameterization (TAIR), to improve neural prediction of real-fluid thermodynamic properties in supercritical combustion by replacing raw enthalpy with target-matched thermodynamic coordinates.

How it works

The core problem addressed is the computational cost of evaluating the cell-local thermodynamic closure, which requires repeated real-fluid state evaluations from a solver-available state (h, p, Y) through enthalpy–temperature inversion and repeated real-fluid equation-of-state evaluations

TAIR replaces the raw enthalpy coordinate of each property network with a target-matched thermodynamic coordinate

The temperature network uses a temperature estimate obtained by inverting a constant-cp ideal-gas mixture enthalpy approximation, while the density and compressibility networks use an ideal-gas density estimate

These explicit algebraic transformations use only solver-available variables and species constants, guiding the networks to learn realfluid departures from ideal-gas-based baselines rather than reconstructing the full closure from raw enthalpy

Input Coordinates and Transformations

The raw-input baseline uses the solver-native enthalpy coordinate h for all three targets, resulting in a mapping (T,ρ,ψ) = Fθ (h, p,Y)

TAIR modifies only this thermodynamic coordinate while retaining pressure, composition, network architecture, and output definitions

For the temperature network, the transformed coordinate is T˜ = T0 + h−Ns∑iYih◦f,iNs∑iYic◦p,i

The density and compressibility networks use an ideal-gas density estimate ρ˜ = p/(RmT˜)

The resulting target-specific assignments are summarized in Table I, showing that TAIR assigns each network the target-matched thermodynamic coordinate that represents the leading ideal-gas dependence of its target

Performance and Accuracy Gains

The method is assessed using supercritical methane–oxygen counterflow flame data against a raw-input baseline and targetinconsistent cross-reparameterization controls

TAIR reduces held-out test-set root-mean-squared error (RMSE) by factors of approximately 1.5, 2.0, and 7.5 for T, ρ, and ψ, respectively

For an unseen-strain-rate flame within the augmented thermodynamic envelope, the corresponding factors are 3.6, 14.5, and 6.0

The target-inconsistent controls perform markedly worse, indicating that the gains arise from thermodynamically matched input design rather than generic preprocessing

For temperature, the RMSE decreases from 0.371K for the raw-input model to 0.250K for TAIR, an improvement factor of approximately 1.5

For density, the raw-input model yields an RMSE of 0.0746 kgm−3, whereas TAIR reduces it to 0.0368 kgm−3, corresponding to an improvement factor of approximately 2.0

The strongest improvement is obtained for compressibility, where TAIR reduces the RMSE from 3.19 × 10−8 s squared m−2 for the raw-input model to 4.24 × 10−9 s squared m−2, corresponding to a reduction factor of approximately 7.5

Computational Cost and Transferability

The TAIR workflow requires an additional 0.014 s over the full test set relative to the raw-input neural model, corresponding to an observed overhead of approximately 8.8%

This overhead is therefore small relative to the iterative thermodynamic calculation that the surrogate replaces

The preservation of the TAIR accuracy advantage at this separately generated flame demonstrates transfer across strain-rate conditions within the augmented thermodynamic envelope

The unseen-strain-rate case demonstrates transfer to a different flame structure and operating condition within the augmented thermodynamic envelope, rather than unrestricted extrapolation to arbitrary thermodynamic states

Thermodynamic Interpretation

TAIR is best interpreted as a deterministic thermodynamic preconditioning of the input coordinates

The estimated temperature T˜ is constructed by neglecting the real-fluid enthalpy departure and approximating the ideal-gas species heat capacities as constants at the reference temperature

TAIR does not introduce additional thermodynamic-state information, as it retains p and Y, and h and T˜ are related by an invertible affine transformation

The target-matched coordinates expose the leading ideal-gas-based dependence of each target, leaving a more departure-like input–output relation for neural regression

The crossreparameterization results further show that replacing enthalpy by an analytical variable does not by itself guarantee improved optimization

The target-matched coordinates are associated with tighter projected relations and lower attainable training losses under the same network capacity and optimization procedure

The analysis characterizes only one representative state in each temperature bin and only the selected first input coordinate

Conclusion

TAIR reduces the held-out test-set RMSE by factors of approximately 1.5, 2.0, and 7.5 for T, ρ, and ψ, respectively

TAIR also achieved a closure-level single-core speed-up of approximately 89.

Improvements for AI systems

What TAIR can do: It allows neural networks to learn real-fluid departures from ideal-gas-based baselines rather than reconstructing the full closure from raw enthalpy, leading to improved accuracy by providing a lowcost thermodynamic transformation of the input coordinates.
What TAIR can do: The method provides specific performance gains, reducing held-out test-set root-mean-squared error (RMSE) by factors of approximately 1.5, 2.0, and 7.5 for T, ρ, and ψ for supercritical methane–oxygen counterflow flame data.
What TAIR can do: It enables closure-level computational speed-up; specifically, the TAIR workflow achieves a closure-level single-core speed-up of approximately 89.4× relative to the iterative Cantera reference.

Sources

Gaussian Error Linear Units (GELUs)