A Machine Learning-Enhanced Hopf-Cole Formulation for Nonlinear Gas Flow in Porous Media

Abstract: Accurate modeling of gas flow through porous media is critical for many technological applications, including reservoir performance prediction, carbon capture and sequestration, and fuel cells and batteries. However, such modeling remains challenging due to strong nonlinear behavior and uncertainty in model parameters. In particular, gas slippage effects described by the Klinkenberg model introduce pressure-dependent permeability, which complicates numerical simulation and obscures deviations from classical Darcy flow behavior. To address these challenges, we present an integrated modeling framework for gas transport in porous media that combines a Klinkenberg-enhanced constitutive relation, Hopf-Cole-transformed mixed-form linear governing equations, a shared-trunk neural network architecture, and a Deep Least-Squares (DeepLS) solver. The Hopf-Cole transformation reformulates the original nonlinear flow equations into an equivalent linear system closely related to the Darcy model, while the mixed formulation, together with a shared-trunk neural architecture, enables simultaneous and accurate prediction of both pressure and velocity fields. A rigorous convergence analysis is performed both theoretically and numerically, establishing the stability and convergence properties of the proposed solver. Importantly, the proposed framework also naturally facilitates inverse modeling of pressure-dependent permeability and slippage parameters from limited or indirect observations, enabling efficient estimation of flow properties that are difficult to measure experimentally. Numerical results demonstrate accurate recovery of flow dynamics and parameters across a wide range of pressure regimes, highlighting the framework's robustness, accuracy, and computational efficiency for gas transport modeling and inversion in tight formations.

• The paper introduces a Hopf–Cole transformation that exactly converts the nonlinear Klinkenberg gas-flow equations into a linear Darcy-type mixed pressure–velocity system, with physical pressure recovered through the Lambert-W function.
• The DeepLS framework combines Fourier features, shared-trunk neural networks, adaptive loss weighting, and Adam–L-BFGS optimization to achieve accurate, stable, meshfree solutions for curved geometries, layered media, and anisotropic permeability fields.
• The convergence analysis establishes well-posedness and consistency as network capacity and collocation density increase, while benchmarks show close agreement with analytical and finite-element solutions, including 3D spherical flow trained in 8.26 minutes on a T4 GPU.

Motivation and problem setting

Gas flow through porous media at low pressures violates the classical Darcy model: wall slip produces an apparent permeability that exceeds the intrinsic permeability, a phenomenon captured empirically by the Klinkenberg correction. In this paper, Maduri and Nakshatrala formulate gas transport under the Klinkenberg model, in which the apparent permeability tensor takes the form $$\mathbf{K}_g(\mathbf{x}) = \mathbf{K}_0(\mathbf{x})(1 + \beta p_{\mathrm{atm}}/p)$$, making the governing equations quasilinear through pressure-dependent permeability. The authors' stated design goals are threefold: mitigate convergence and stability issues inherent to nonlinear models (G1), achieve high accuracy in the velocity field rather than only pressure (G2), and ensure strong numerical stability (G3). Their response is an integrated framework combining four ingredients: a Hopf–Cole transformation that linearizes the Klinkenberg equations, a mixed pressure–velocity formulation, a shared-trunk neural network architecture, and a Deep Least-Squares (DeepLS) solver. Notably, the framework extends the classical scalar Klinkenberg model to anisotropic and spatially heterogeneous permeability fields, and adopts a dimensionless $\beta$ in contrast to the conventional pressure-dimensioned form.

Hopf–Cole transformation and least-squares formulation

The central analytical device is the transformed pressure variable

$$\mathcal{P}(\mathbf{x}) = p(\mathbf{x}) + \beta\, p_{\mathrm{atm}} \ln[p(\mathbf{x})],$$

whose gradient satisfies $$\mathrm{grad}[\mathcal{P}] = (1 + \beta p_{\mathrm{atm}}/p)\,\mathrm{grad}[p]$$. Because the velocity field is invariant under this change of variables, the original nonlinear system reduces exactly to a linear Darcy-type problem in $(\mathcal{P}, \mathbf{u})$: a Darcy momentum balance with intrinsic permeability $\mathbf{K}_0$, mass conservation, and boundary conditions on $\Gamma_u$ and $\Gamma_p$. The physical pressure is recovered via the Lambert–$W$ function on its principal branch, which is well defined because the mapping $p \mapsto \mathcal{P}$ is continuous and strictly monotone on $\beta$0. A practical constraint follows directly from the logarithmic structure: prescribed pressures must be strictly positive absolute values, and nondimensionalization or reference-pressure shifts can invalidate the transformation if they produce nonpositive data.

The authors then construct a weighted least-squares functional over the residuals of the linearized system, including $\beta$1 penalties on both boundary segments. This choice is deliberate: unlike the Deep Ritz Method, whose mixed Darcy form yields a saddle-point system, or PINNs, which require adaptive collocation and loss balancing on strong-form residuals, the least-squares functional is nonnegative, symmetric, and positive definite, giving a well-conditioned optimization landscape. Training uses a shared-trunk network with Fourier feature lifting, Monte Carlo collocation for the integrals, adaptive loss weighting based on relative per-term convergence rates, and a two-stage Adam-then-L-BFGS optimizer.

Convergence analysis

The paper develops a complete functional-analytic convergence theory on the product space $\beta$2, where the augmented velocity space is required because the flux boundary residual is measured in $\beta$3; the authors discuss and reject weaker ($\beta$4 trace) and stronger ($\beta$5) alternatives. Under uniform ellipticity of $\beta$6 and $\beta$7, they prove coercivity of the bilinear form with explicit constants derived from carefully chosen weights $\beta$8, boundedness, strong convexity of $\beta$9, uniqueness of the global minimizer, and a quadratic-growth estimate. ReLU networks are shown to be dense in $$\mathcal{P}(\mathbf{x}) = p(\mathbf{x}) + \beta\, p_{\mathrm{atm}} \ln[p(\mathbf{x})],$$0 (via Sobolev approximation results), yielding a bias-gap bound $$\mathcal{P}(\mathbf{x}) = p(\mathbf{x}) + \beta\, p_{\mathrm{atm}} \ln[p(\mathbf{x})],$$1, and a uniform perturbation bound controls the discretization error from Monte Carlo sampling. Combining these with quadratic growth gives the total-error estimate

$$\mathcal{P}(\mathbf{x}) = p(\mathbf{x}) + \beta\, p_{\mathrm{atm}} \ln[p(\mathbf{x})],$$2

so the trained solution converges as network capacity and collocation count grow. Since the inverse transformation is continuous and the velocity is untouched by the transformation, convergence transfers to the physical pressure and velocity fields. One caveat is worth noting plainly: the density result guarantees existence of good approximants but provides no rates, so the analysis establishes consistency rather than quantitative error bounds for fixed architectures.

Numerical results

Four benchmarks validate the framework. For gas flow through concentric cylinders, DeepLS reproduces the closed-form Klinkenberg solution for both pressure and radial velocity using a 6-layer, 64-neuron ReLU network trained in 6.34 minutes on an NVIDIA T4 GPU; the solution correctly deviates from the Darcy prediction due to slip-enhanced permeability. Capacity studies show a sharp drop of several orders of magnitude in $$\mathcal{P}(\mathbf{x}) = p(\mathbf{x}) + \beta\, p_{\mathrm{atm}} \ln[p(\mathbf{x})],$$3 error beyond a moderate capacity threshold, with depth providing steeper gains once width bottlenecks are removed. The footing problem, featuring abrupt transitions between prescribed-pressure and no-flux segments on the top boundary, shows small errors against a stabilized mixed FEM reference, and increasing collocation density (3,000 to 12,000 points) systematically lowers terminal training loss. The layered medium benchmark demonstrates stable, oscillation-free resolution of step-like velocity profiles across five layers with sharp permeability contrasts—features that would typically require discontinuous Galerkin treatments in FEM. Finally, for gas flow through concentric spheres in 3D with a symmetry-plane no-flux condition enforced directly in the loss (avoiding Nitsche-type weak imposition), DeepLS achieves near-perfect agreement with the analytical solution in 8.26 minutes of training time, again without mesh generation.

Mechanics-based verification via Betti reciprocity

Beyond standard error metrics, the paper derives a Betti-type reciprocal relation for the Klinkenberg model expressed in terms of the transformed pressure, relating pairs of solutions driven by distinct boundary data. Applied to the footing problem—which admits no closed-form solution—the normalized reciprocity residual $$\mathcal{P}(\mathbf{x}) = p(\mathbf{x}) + \beta\, p_{\mathrm{atm}} \ln[p(\mathbf{x})],$$4 serves as an a posteriori error indicator requiring no reference solution. The empirical finding is architecturally informative: reciprocity error decreases with width but plateaus, while deeper networks achieve consistently lower errors at comparable widths, suggesting depth is the more parameter-efficient route to satisfying the underlying physics.

Limitations and open questions

Several limitations are acknowledged or evident. The transformation requires strictly positive absolute pressures, excluding gauge-shifted or nondimensionalized formulations where the datum may be nonpositive. The convergence theory relies on a uniform perturbation bound over the network class and delivers no explicit approximation rates; the bias-gap argument presumes the empirical minimizer is actually found by the optimizer, whereas the practical Adam/L-BFGS procedure offers no global optimality guarantee. All numerical validation is on synthetic benchmarks with unit viscosity and moderate $$\mathcal{P}(\mathbf{x}) = p(\mathbf{x}) + \beta\, p_{\mathrm{atm}} \ln[p(\mathbf{x})],$$5; performance under extreme parameter contrasts, body forces, or transient effects is untested. The claimed capability for inverse modeling of permeability and slippage parameters from limited observations is asserted in the abstract but not demonstrated numerically in the presented results. Open questions left by the paper include extension to multiphase flow with capillarity, integration of uncertainty quantification, and scaling behavior on high-performance computing platforms.

Conclusion

This work couples a Hopf–Cole linearization of the Klinkenberg model with a mixed-form DeepLS neural solver, supported by a rigorous well-posedness and convergence analysis and verified against analytical solutions, finite element references, and a Betti reciprocity identity. The main contributions are accurate simultaneous recovery of pressure and velocity—including in strongly layered media—meshfree handling of curved 3D geometries with embedded symmetry constraints, and quantified time-to-solution on modest hardware. The framework offers a mathematically grounded alternative to iterative nonlinear solvers for non-Darcy gas transport, though its demonstrated scope remains confined to steady, single-phase, positive-pressure regimes on synthetic problems.