betaFeedback

WorksPhysics and Astronomy

Regional accuracy of physics-informed neural networks (PINNs) for recirculation in a deep lid-driven cavity

Abstract

Deep lid-driven cavities contain strong motion beneath the lid and much weaker recirculation at depth, so agreement in bulk velocity is an incomplete assessment of a computed flow. We examine data-free physics-informed neural networks (PINNs) for steady, two-dimensional incompressible flow in a cavity of depth-to-width ratio five at Reynolds numbers 100, 250, 500 and 1000. An established hard-constrained streamfunction formulation is combined with prescribed regional collocation, transferred initialization, and selected residual-scaling and envelope variants. Independent spectral/hphp calculations with the same Gaussian-regularized lid provide the reference fields. For a cavity of depth-to-width ratio D/W=5D/W=5 at Reynolds numbers Re=100\mathrm{Re}=100, 250, 500, and 1000, evaluation on a common 121×605121\times605 grid gives interior relative velocity discrepancies of 5.130%–7.850%. Over the same cases, floor-region relative velocity discrepancies remain below 5%, with absolute RMS discrepancies ranging from 1.255×10−10U1.255\times10^{-10}U to 2.332×10−9U2.332\times10^{-9}U, where UU is the lid-velocity scale. In a separate 161×801161\times801 assessment at \Rey=1000, the refined workflow lowers interior velocity discrepancy from 5.624% to 2.658% and gauge-aligned pressure RMS discrepancy from 2.793×10−32.793\times10^{-3} to 1.167×10−3ρU21.167\times10^{-3}ρU^2. The improvements compare accumulated training workflows and do not isolate regional sampling. Configuration rankings depend on both region and observable. These finite-budget comparisons establish where the retained workflows agree with the reference and where rotational-field discrepancies remain substantial.

Introduction

A deep lid-driven cavity tests more than the reproduction of its dominant vortex. Motion imposed at the lid drives a sequence of much weaker recirculating vortices below it, while the strongest velocity gradients remain close to the moving wall and its junctions with stationary sidewalls. A numerical approximation can therefore agree in bulk velocity while differing in local shear or in weak motion at depth. This study asks how consistently a data-free physics-informed neural network (PINN) reproduces velocity, pressure and sampled rotational structure across these regions.

The physical configuration is established. Burggraf connected viscous cavity eddies with the rotational core that develops as Reynolds number increases [1]; Pan and Acrivos examined rectangular cavities, including creeping-flow calculations at depths up to five widths [2]. Moffatt’s corner-eddy analysis explains a local hierarchy of diminishing viscous structures [3], which must be distinguished from the larger vortices stacked through a deep cavity. Cheng and Hung studied 0.1≤D/W≤70.1 \le D/W \le7 and 0.01≤Re⁡≤50000.01 \le\operatorname{Re} \le5000 [10]. Zhuo et al. included D/W=5D/W = 5 at Re⁡=100\operatorname{Re} = 100, 500 and 1000 among broader ranges [11]. The present work does not claim the first simulation of this geometry or Reynolds number. Its subject is the regional accuracy of a neural solution procedure for this established flow.

Classical square-cavity benchmarks also show why the boundary-value problem and assessment protocol matter. Multigrid, continuation and spectral calculations provide accurate reference solutions [5, 6, 7, 8], while steady and time-dependent studies distinguish solution accuracy from physical stability [4, 9]. In particular, regularizing the moving lid changes the problem at the wall junctions. Here the PINN and reference solve the same Gaussian-regularized-lid problem; comparisons with conventional discontinuous-lid studies are qualitative. A reported steady field or maximum studied Reynolds number is not a measured instability threshold. Table 1 records these distinctions for the deep-cavity literature.

StudyVerified geometry and Reynolds scopeFlow regime and relevance
Pan and Acrivos [2]Creeping-flow calculations through D/W=5D/W = 5; experiments at Re⁡=20\operatorname{Re} = 20–4000Steady-flow structure; 4000 is an experimental range limit, not a measured Hopf threshold.
Cheng and Hung [10]D/W=0.1D/W = 0.1–7; Re⁡=0.01\operatorname{Re} = 0.01–5000Steady vortex structures, including deep vortices; the reported range is not an onset criterion.
Zhuo et al., Part I [11]D/W=1.5D/W = 1.5, 3, 5, 7; Re⁡=100\operatorname{Re} = 100, 500, 1000, 3200, 5000Steady-flow positions, sizes and strengths; directly relevant geometric precedent.
Zhuo et al., Part II [12]First Hopf bifurcations for D/W=1D/W = 1–5Periodic and quasiperiodic states above geometry-dependent critical values; not a universal turbulence threshold.
Lin et al. [13]Transition study at D/W=1D/W = 1–3Documents depth-dependent Hopf onset; also discusses earlier unsteady cases at Re⁡=7500\operatorname{Re} = 7500 for D/W=1.5D/W = 1.5 and 4. These are not depth-five thresholds.

Table 1. Deep-cavity literature: a maximum studied Reynolds number is not a transition threshold. Depth is normalized by lid width.

Neural trial functions can enforce boundary conditions by construction [14]; modern PINNs train through differential-equation residuals [17, 18, 19]. For incompressible flow, variable choice, loss weighting and transfer affect optimization and differentiation [20]. Gradient imbalance, conditioning and unequal convergence rates complicate the relation between a small training loss and field accuracy [21, 22, 23, 34]. A deep cavity makes this distinction physically concrete: an objective dominated by strong upper motion need not adequately constrain weak lower recirculation. Hard constraints, localized representations and adaptive sampling offer different responses [25, 26, 28, 30, 33]; they should not be conflated with the prescribed regional sampling used here.

Whereas McDevitt et al. [39] examined square and shallower trapezoidal cavities, the present study extends the assessment to a rectangular cavity with D/W=5D/W = 5 at Re⁡=100\operatorname{Re} = 100, 250, 500, and 1000. Its distinguishing contribution is a matched, quantitative assessment of velocity, gauge-aligned pressure, and sampled vorticity across the strong upper circulation and weak floor recirculation, together with an evaluation of regional training, transferred initialization, and selected scaling variants.

Table 2 places these extensions alongside the reported geometry, training-data requirements and assessment scope of related cavity PINN studies. It distinguishes the present deep-cavity, region-resolved comparison from earlier applications; it does not rank solvers whose boundary conditions and evaluation protocols differ.

StudyReported scopeDistinction of the present work
McDevitt et al. [39]Square cavities: data-free cases through Re = 5000 and sparse-data-assisted cases through Re = 21000; parametric trapezoidal cavities with 0.5≤D/W≤10.5 \le D/W \le1. Velocity, streamfunction and pressure fields are examined.Extends the assessment to a rectangular cavity with D/W=5D/W = 5. Quantifies regional velocity, gauge-aligned pressure and sampled vorticity discrepancies against a matched spectral/hp reference, including the weak floor recirculation.
Fowler et al. [40]Two-dimensional natural convection in a thermal cavity, with Rayleigh number included as a network input.Addresses mechanically driven recirculation rather than buoyancy-driven flow. The fixed-depth Reynolds-number series examines regional agreement across the strong upper circulation and weak lower vortices.
Takao and Ii [41]Rectangular cavities with D/W=0.5D/W = 0.5–22 at Re = 100; coordinate transformation and fine-tuning using interior velocity observations.Examines the deeper geometry D/W=5D/W = 5 through Re = 1000, without supervised interior solution values in training. Transferred initialization is obtained from earlier PINN states.
Wang et al. [33]Data-free square-cavity calculations through Re = 1000, using adaptive sampling.Examines deep-cavity flow using prescribed regional collocation, rather than adaptive sampling, and evaluates workflow changes through separate upper, interior and floor metrics.
Present workData-free rectangular-cavity assessment at D/W=5D/W = 5 and Re = 100, 250, 500 and 1000, with detailed workflow comparisons at Re = 1000.Combines regional collocation, configuration-matched transferred initialization and selected residual-scaling and envelope variants with a matched, region-resolved assessment of velocity, pressure and sampled vorticity. Identifies changes in workflow ranking between regions and observables.

Table 2. Scope of related cavity PINN studies and the extensions examined in the present work.

Recent cavity-specific reviews and reliability studies emphasize that formulation, boundary treatment and evaluation metrics prevent simple rankings across implementations [43, 44]. We address this issue with matched lid data and independent Nektar++ spectral/hp fields [47, 48]. Velocity, pressure after a specified gauge alignment, and vorticity computed from both sampled velocity fields are assessed separately. The last quantity tests sampled derivative agreement, rather than the native automatic-differentiation vorticity of the network.

The contribution is a matched workflow assessment at D/W=5D/W = 5 and Re = 100, 250, 500 and 1000, with detailed comparisons at 1000. We evaluate prescribed regional collocation, configuration-matched transferred initialization, and selected residual-scaling and envelope variants within the inherited formulation. The comparisons identify improved velocity and pressure measures, and changes in configuration ranking across regions. All four detailed workflows share the same final sampling prescription but have different preceding histories. Their comparison therefore assesses accumulated workflows rather than isolating the benefit of sampling. The evidence supports conclusions for these retained states, one geometry and the stated numerical reference.

Related neural PDE approaches include the deep Galerkin and Deep Ritz methods [15, 16], variational domain decomposition [27], and the DeepXDE software framework [29]. Gradient-enhanced losses, self-adaptive weighting and repeated parameter initialization offer other training strategies [31, 32, 38]; they are not components of the present implementation. Hard-boundary Navier–Stokes formulations provide a related comparison [42], while error estimates for Navier–Stokes PINNs distinguish analytical accuracy questions from empirical residual reduction [24].

The paper first specifies the matched problem, inherited representation and assessment protocol. The results then present the common-grid Reynolds-number series before the detailed Re⁡=1000\operatorname{Re}=1000 workflow comparisons. The discussion separates observed regional performance from reference sensitivity, optimization progress and dynamical stability. Recorded loss histories are retained in an appendix.

Physical problem and cavity campaign

Throughout, Re⁡\operatorname{Re} denotes Reynolds number. Unless a dimensional quantity is explicitly identified, the equations use nondimensional variables. In the method name spectral/hphp, hh means element size and pp means polynomial degree; these are distinct from the plotting height hh and the pressure pp.

Governing equations and matched lid

Consider a two-dimensional rectangular cavity of width WW and depth DD. Coordinates are 0≤x≤W0 \le x \le W and 1−D≤y≤11-D \le y \le1 in the implementation, with the lid at y=1y=1. For figures, h=y+D−1h=y+D-1 denotes height above the floor. Length and velocity are nondimensionalized by WW and the lid velocity scale UU, respectively; Re⁡=UW/ν\operatorname{Re}=UW/\nu and U=W=1U=W=1 in the cases considered here. The steady incompressible equations are

ux+vy=0,(1)u_x+v_y=0, \tag*{(1)}
Rx=uux+vuy+px−Re⁡−1(uxx+uyy)=0,(2)R_x=uu_x+vu_y+p_x-\operatorname{Re}^{-1}(u_{xx}+u_{yy})=0, \tag*{(2)}
Ry=uvx+vvy+py−Re⁡−1(vxx+vyy)=0.(3)R_y=uv_x+vv_y+p_y-\operatorname{Re}^{-1}(v_{xx}+v_{yy})=0. \tag*{(3)}

Here uu and vv are the horizontal and vertical velocity components, pp is pressure, and RxR_x and RyR_y are the corresponding steady momentum residuals. Subscripts denote differentiation: for example, ux=∂u/∂xu_x=\partial u/\partial x and uxx=∂2u/∂x2u_{xx}=\partial^2u/\partial x^2. The parameter ν\nu is kinematic viscosity. In Eq. (4), UlidU_{\mathrm{lid}} is the prescribed nondimensional lid speed and δx\delta_x its horizontal smoothing length. Pressure is defined up to an additive constant. The side and bottom walls are stationary. The moving lid is smoothly brought to rest at its endpoints:

u(x,1)=Ulid(x)=[1−e−x2/δx2][1−e−(x−W)2/δx2],v(x,1)=0,δx=0.001 W.(4)u(x,1)=U_{\mathrm{lid}}(x)=\left[1-e^{-x^2/\delta_x^2}\right]\left[1-e^{-(x-W)^2/\delta_x^2}\right],\qquad v(x,1)=0,\qquad\delta_x=\sqrt{0.001}\,W. \tag*{(4)}

The smoothing in Eq. (4) removes the discontinuity at the lid endpoints. We use this same profile in the PINN and CFD calculations, so the reported discrepancies compare solutions of the same boundary-value problem. A broader transition could ease approximation near the corners, but would change the prescribed lid and is not introduced in this comparison. Standard cavity results, including the high-Reynolds-number solutions of Erturk et al. [8], provide useful context, although their lid conditions must be considered when comparing fields. Earlier finite-volume calculations with different lid profiles are excluded from the quantitative comparison below.

Nondimensionalization, pressure and vorticity

An asterisk identifies a dimensional quantity; tt denotes time, p0p_0 an arbitrary constant reference pressure, and u=(u,v)\mathbf{u}=(u,v) the velocity vector. The gradient is ∇=(∂x,∂y)\nabla=(\partial_x,\partial_y) and the Laplacian is ∇2=∂xx+∂yy\nabla^2=\partial_{xx}+\partial_{yy}. In the component sum in Eq. (6), x1=xx_1=x, x2=yx_2=y, u1=uu_1=u, u2=vu_2=v, and both indices i,ji,j run from 1 to 2. The streamfunction ψ\psi in Eq. (7) satisfies u=ψyu=\psi_y and v=−ψxv=-\psi_x. For a fluid of constant mass density ρ\rho and kinematic viscosity ν\nu, the dimensional variables are transformed as x=x∗/Wx=x^*/W, y=y∗/Wy=y^*/W, t=t∗U/Wt=t^*U/W, u=u∗/U\mathbf{u}=\mathbf{u}^*/U and p=(p∗−p0)/(ρU2)p=(p^*-p_0)/(\rho U^2). Thus the Reynolds number measures the ratio of inertial to viscous effects on the cavity-width scale, rather than on its depth. Varying D/WD/W at fixed Re⁡\operatorname{Re} changes the geometry without changing this reference viscosity. In the shifted coordinates used by the neural implementation, the moving wall remains at y=1y=1 and the floor is at y=1−Dy=1-D; the plotting coordinate h=y+D−1h=y+D-1 is a translation and has the same spatial derivatives as yy.

The independent CFD calculation advances the unsteady equations

∂tu+u⋅∇u=−∇p+Re⁡−1∇2u,∇⋅u=0.(5)\partial_t\boldsymbol{u}+\boldsymbol{u}\cdot\nabla\boldsymbol{u}=-\nabla p+\operatorname{Re}^{-1}\nabla^2\boldsymbol{u},\qquad\nabla\cdot\boldsymbol{u}=0. \tag*{(5)}

The PINN seeks a steady field, for which the first term vanishes. No volume force or interior velocity observations are included in its momentum objective. The lid supplies energy through tangential motion, while viscous stress transfers this motion into the interior and dissipates it at the stationary walls. As the aspect ratio increases, the upper circulation drives progressively weaker motion at depth. Resolving the gradients associated with that motion is important even where the local speed is small.

Pressure is an independent output in the neural formulation. It provides the gradient required to balance convection and diffusion, rather than being reconstructed from velocity after training. Taking the divergence of Eq. (5) and using continuity gives

∇2p=−∑i,j∂ui∂xj∂uj∂xi=−[ux2+2uyvx+vy2].(6)\nabla^2p=-\sum_{i,j}\frac{\partial u_i}{\partial x_j}\frac{\partial u_j}{\partial x_i}=-\left[u_x^2+2u_yv_x+v_y^2\right]. \tag*{(6)}

This identity explains the relationship between pressure curvature and velocity gradients. It is not imposed as an additional PINN loss: the two momentum equations already couple the pressure output to velocity. An additive pressure constant leaves both equations unchanged. For field comparisons, the candidate-minus-reference mean pressure over the common comparison region is removed. This aligns their gauges while retaining spatial discrepancies, gradients and extrema. It is particularly important when comparing pressure contours from solvers with different treatments of the constant mode.

The out-of-plane vorticity, positive for counterclockwise rotation, is

ω=vx−uy=−∇2ψ.(7)\omega=v_x-u_y=-\nabla^2\psi. \tag*{(7)}

Taking the curl of the momentum equations eliminates pressure and yields

∂tω+uωx+vωy=Re⁡−1(ωxx+ωyy).(8)\partial_t\omega+u\omega_x+v\omega_y=\operatorname{Re}^{-1}\left(\omega_{xx}+\omega_{yy}\right). \tag*{(8)}

In a steady cavity, convection redistributes the vorticity generated at the walls while diffusion spreads it into the interior. Vorticity maps therefore complement streamlines: similar circulation patterns can coexist with different shear-layer thicknesses or corner gradients. Pressure and vorticity also probe different aspects of the solution. Pressure reflects momentum balance across the cavity, whereas vorticity emphasizes derivatives of velocity and is more sensitive to local resolution. For the paired sampled-field plots, both velocity fields are differentiated with the same spatial operator on the same grid; this makes their vorticity discrepancy comparable without mixing neural automatic differentiation with a different CFD plotting derivative.

Cavity cases

Table 3 records the retained campaign inventory. The matched multiphysics assessment uses the D/W=5D/W=5 cases at four Reynolds numbers, with detailed workflow comparisons at Re⁡=1000\operatorname{Re}=1000.

ReD/WD/WRole and evidence used here
1001, 2, 3, 4, 5, 6Retained depth ladder; D/W=5D/W = 5 supplies the lower-Reynolds-number assessment.
2501, 5Intermediate-Reynolds-number cases with a matched lid in the deep cavity.
5001, 5Intermediate-Re development and initialization for the principal deep-cavity control.
10001, 5Square-cavity context and principal deep-cavity study. Detailed high-Re results below concern D/W=5D/W = 5.

Table 3. Retained campaign inventory. Only the D/W=5D/W = 5 cases enter the matched four-Reynolds-number assessment.

At greater depth, the upper flow dominates both the velocity field and an unweighted momentum residual. The lower recirculation therefore requires particular attention. Moffatt’s corner analysis [3] gives a classical example of weak structures spanning widely separated scales. The local corner eddies in that analysis should be distinguished from the stacked recirculation occupying the interior of a deep cavity.

Scale-aware neural formulation and training

Hard-constrained streamfunction–pressure representation

The trainable weights and biases form the parameter vector θ\boldsymbol{\theta}. The scalar outputs NψN_{\psi} and NpN_{p} are the neural streamfunction correction and pressure output, respectively. The lifting ψlid\psi_{\mathrm{lid}} supplies the prescribed wall motion, BB is a boundary-vanishing factor, and EE is the depth envelope. The local coordinates ξ\xi and η\eta range from zero to one across the width and depth. The vertical lifting length is δy\delta_y. Depth below the lid is dd; d0d_0 is the envelope transition depth, κ\kappa its decay rate, and β\beta the inverse length controlling the smoothness of the softplus function sβs_{\beta}. Its argument zz is a scalar, and σ\sigma denotes the logistic function. Two network outputs Nψ(x,y;θ)N_{\psi}(x,y;\boldsymbol{\theta}) and Np(x,y;θ)N_{p}(x,y;\boldsymbol{\theta}) define

ψ=ψlid+B2E(y)Nψ,p=E(y)Np,(9)\psi= \psi_{\mathrm{lid}} + B^2 E(y)N_{\psi}, \qquad p = E(y)N_p, \tag*{(9)}
u=∂yψ,v=−∂xψ,(10)u = \partial_y\psi, \qquad v = -\partial_x\psi, \tag*{(10)}
B=16ξ(1−ξ)η(1−η),ξ=x/W,η=[y−(1−D)]/D.(11)B = 16\xi(1-\xi)\eta(1-\eta), \qquad\xi= x/W, \qquad\eta= [y-(1-D)]/D. \tag*{(11)}

The analytic lifting is

ψlid=(y−1)η2[1−e−(x−W)2/δy2][1−e−x2/δy2]e−(1−y)2/δy2,δy=0.1W.(12)\psi_{\mathrm{lid}} = (y-1)\eta^2\left[1-e^{-(x-W)^2/\delta_y^2}\right]\left[1-e^{-x^2/\delta_y^2}\right]e^{-(1-y)^2/\delta_y^2}, \qquad\delta_y = 0.1W. \tag*{(12)}

The factor B2B^2 makes both the neural correction and its first derivatives vanish on the walls. Equation (12) provides the prescribed tangential motion at the lid while satisfying the other stationary-wall velocities. Incompressibility follows identically from Eq. (10). Thus neither continuity nor velocity boundary conditions need separate penalty weights. Hard constraints reduce one source of optimization competition; they do not by themselves ensure the momentum equations are accurately solved.

The depth envelope inherited from the deep-cavity baseline is

E(d)=exp⁡{−κ[sβ(d−d0)−sβ(−d0)]},d=1−y,(13)E(d) = \exp\left\{-\kappa\left[s_{\beta}(d-d_0)-s_{\beta}(-d_0)\right]\right\}, \qquad d = 1-y, \tag*{(13)}
sβ(z)=β−1log⁡(1+eβz),β=10/W,d0=0.25W,κ=4.212/W.(14)s_{\beta}(z) = \beta^{-1}\log(1+e^{\beta z}), \qquad\beta= 10/W, \qquad d_0 = 0.25W, \qquad\kappa= 4.212/W. \tag*{(14)}

The retained value κ=4.212/W\kappa= 4.212/W is motivated by the real decay exponent in the parallel-wall Stokes-flow limit discussed by Moffatt [3]. Here it sets the representation scale; it is not fitted to the CFD comparison and is not assumed to be an exact decay rate for the finite-Reynolds-number cavity. The function E(y)E(y) in the field representation means E(d)E(d) evaluated at d=1−yd = 1-y. The envelope supplies a smooth depth scale without prescribing vortex locations or forcing a particular number of vortices. Its derivatives are retained in automatic differentiation: the physical velocity is the derivative of the complete transformed streamfunction, not simply an envelope times an untransformed neural velocity. Stable evaluation of the softplus is important because the momentum residual involves high derivatives of ψ\psi.

Differentiation of the transformed fields

The network receives the physical nondimensional coordinates (x,y)(x,y) and returns two scalar functions. It does not predict velocity components independently. With A=B2A = B^2, differentiation of the full representation gives

u=ψlid,y+AyENψ+AEyNψ+AE∂yNψ,(15)u = \psi_{\mathrm{lid},y}+A_yEN_\psi+AE_yN_\psi+AE\partial_yN_\psi, \tag*{(15)}
v=−ψlid,x−AxENψ−AE∂xNψ,(16)v = -\psi_{\mathrm{lid},x}-A_xEN_\psi-AE\partial_xN_\psi, \tag*{(16)}

because EE depends only on depth. These product-rule terms are part of the physical field. In particular, changing the envelope changes velocity even when the network parameters are unchanged. The analytic lifting and the first derivatives of BB and EE are evaluated explicitly in the implementation, and network derivatives are obtained by automatic differentiation. Momentum residuals then require first and second derivatives of velocity, hence derivatives of the neural streamfunction through third order. The pressure terms require first derivatives of its transformed output.

For the smooth envelope in Eq. (13), its derivative is

Ey=κ σ(β[d−d0])E,σ(z)=11+e−z.(17)E_y = \kappa\,\sigma\left(\beta[d-d_0]\right)E,\qquad\sigma(z)=\frac{1}{1+e^{-z}}. \tag*{(17)}

The positive sign follows from d=1−yd=1-y: the envelope increases toward the lid. The subtraction of sβ(−d0)s_\beta(-d_0) makes E(0)=1E(0)=1. Below the transition region, the network represents a slowly scaled correction while EE carries its depth dependence. We evaluate softplus through a smooth log-add-exp expression, which avoids overflow and retains consistent higher derivatives near the envelope transition. Six tanh hidden layers supply smooth derivatives of the required order; the two-output network has 21,122 trainable parameters and uses double precision throughout.

On each wall, B=0B=0 and ∇(B2)=2B∇B=0\nabla(B^2)=2B\nabla B=0. The neural correction consequently contributes neither penetration nor tangential velocity there, regardless of the learned weights. At the lid, ∂yψlid\partial_y\psi_{\mathrm{lid}} supplies the matched Gaussian-regularized speed. On the stationary walls its first derivatives vanish. The identity ux+vy=ψyx−ψxy=0u_x+v_y=\psi_{yx}-\psi_{xy}=0 enforces continuity throughout the cavity. These properties leave the optimizer to fit momentum balance rather than trade boundary or continuity penalties against momentum residuals. The streamfunction lifting, amplitude envelope and optimizer implementation were supplied by C. J. McDevitt; the regional training distributions and the increased-Reynolds-number deep-cavity assessment are the extensions examined here.

Residual scaling and corner treatment

In Eq. (18), JJ is the empirical loss, NN the number of collocation points, and ii their index. The quantities MiM_i, EiE_i, Rx,iR_{x,i} and Ry,iR_{y,i} are the mask, envelope and momentum residuals evaluated at point ii. The mask MM uses the corner length rmr_m defined in Eq. (19); the notation MRMR below means the product of the mask and either momentum residual. The training objective is

J(θ)=1N∑i=1N[(MiRx,iEi)2+(MiRy,iEi)2],(18)J(\boldsymbol{\theta})=\frac{1}{N}\sum_{i=1}^{N}\left[\left(\frac{M_iR_{x,i}}{E_i}\right)^2+\left(\frac{M_iR_{y,i}}{E_i}\right)^2\right], \tag*{(18)}

with the inherited upper-corner mask

M=[1−e−[x2+(1−y)2]/rm2][1−e−[(x−W)2+(1−y)2]/rm2],rm=0.01W.(19)M=\left[1-e^{-[x^2+(1-y)^2]/r_m^2}\right]\left[1-e^{-[(x-W)^2+(1-y)^2]/r_m^2}\right],\qquad r_m=0.01W. \tag*{(19)}

The reciprocal EE compensates for the attenuation of physical residuals at depth. This is residual scaling, not a modification of Eqs. (2)–(3). The mask tends to zero at the geometric lid junctions and approaches unity away from them. The hard constraints enforce the prescribed wall velocities, while the collocation exclusions described below omit the immediate corner neighborhoods from the sampled momentum objective. The mask specifies the residual weighting at the retained points. The weighted loss alone therefore does not establish local momentum accuracy there. The field comparisons below use the unweighted physical fields and retain the assessed near-lid discrepancies rather than weighting them by MM.

The objective contains no labeled interior velocity, pressure or vorticity data. Transferred parameters come from earlier PINN training, not from fitting the CFD fields. The independent CFD solution is used for the field comparisons reported below, rather than as a supervised term in Eq. (18). This separates the present data-free workflow from cavity methods that incorporate interior observations.

The “soft residual” variant replaces MR/EMR/E by MR/EMR/\sqrt{E} while keeping the output envelope unchanged. The “half envelope” variant uses κ=2.106/W\kappa= 2.106/W in both the output and reciprocal residual scale. These variants separate two mechanisms often conflated as a single weighting choice: changing the field representation and changing the residual emphasis. The half-envelope case undergoes additional Re⁡=500\operatorname{Re} = 500 reconditioning before the Re⁡=1000\operatorname{Re} = 1000 phase, so it is a workflow comparison, not a one-factor test with identical training history.

Regional collocation

Uniform sampling is a natural baseline, but allocates points according to area rather than local difficulty. The implemented balanced sampler draws 50% of its points uniformly over the cavity, 30% from the upper one-width strip, 15% from the lower one-width strip, and 5% from logarithmically distributed neighborhoods of the lower corners. Upper points within a radius 0.02W0.02W of the lid corners are excluded. Here rr is distance from the selected lower corner, and the angle is measured into its interior quadrant. The logarithmic component uses r/W∈[10−4,0.25]r/W \in[10^{-4},0.25], angles in [0.01,π/2−0.01][0.01,\pi/2 - 0.01], and equally likely left and right corners.

All points have equal weight in Eq. (18); consequently, the mixture intentionally changes regional weighting. It is neither an unbiased estimator of a uniform area integral nor a residual-adaptive sampler. Nonuniform and residual-adaptive sampling are established approaches [30]; the present design specifies their regional allocation for this deep cavity. Initialization histories are given separately below, because they differ even when the final sampler is the same.

Architecture, optimization and transfer

The network uses six hidden tanh layers of width 64, with 21,122 trainable parameters. Its weights are initialized with Xavier-normal draws and its biases with zeros before any specified transfer. Double precision is used for field evaluation and optimization.

Figure 1 connects the inherited constrained representation to regional residual evaluation, the changing-batch SOAP phase and the fixed-point self-scaled Broyden phase. Its separate assessment stage compares the resulting physical fields with CFD; the CFD values are not supervised targets in the training objective. The displayed final-phase sequence must be interpreted with the workflow-specific initialization histories.

!Physics-preserving PINN training and independent assessment. 1 Specify the problem and initialization: Set Re and D/WD/W; prescribe the matched lid (Eq. (4)), residual variant and source PINN initialization. 2 Construct the physics-preserving neural fields: (x,y)↦(Nψ,Np)↦(ψ,p)(x,y)\mapsto(N_\psi,N_p)\mapsto(\psi,p) (Eq. (9)); recover u=ψyu=\psi_y and v=−ψxv=-\psi_x (Eq. (10)). Boundary factor BB, lid lifting ψlid\psi_{\mathrm{lid}} and depth envelope EE: Eqs. (11), (12) and (13). 3 Sample points and assemble the physics loss: Draw the regional mixture (Eq. (20)). Differentiate the complete fields and evaluate Rx,RyR_x,R_y (Eqs. (2) and (3)). Apply corner mask MM (Eq. (19)) and envelope scaling; accumulate globally normalized loss and gradient: J(θ)=1N∑i=1N[(MiRx,iEi)2+(MiRy,iEi)2]J(\boldsymbol{\theta})=\frac{1}{N}\sum_{i=1}^{N}\left[\left(\frac{M_iR_{x,i}}{E_i}\right)^2+\left(\frac{M_iR_{y,i}}{E_i}\right)^2\right] (Eq. (18)). 4 SOAP optimization: 10,000 updates. Minimize JJ in Eq. (18); learning rate 0.002. Re-draw the regional batch every 50 steps. Recompute the loss and gradient at each update. 5 Freeze the collocation set: Fix N=131072N=131072 points and retain the same objective (Eq. (18)) throughout the quasi-Newton phase. 6 Self-scaled Broyden: 20,000 updates. Compute the search direction and parameter update (Eq. (21)); perform line-search evaluations and curvature updates while minimizing the same JJ. 7 Independent field assessment: Evaluate u,v,p,ψ,ωu,v,p,\psi,\omega on common physical coordinates; use Eqs. (10) and (7). Compare with matched CFD fields after pressure-gauge alignment. Report bulk and regional discrepancies (velocity metric: Eq. (27)). CFD fields are used for assessment only, not for training. bbox=0.12,0.16,0.88,0.78

Figure 1. Physics-preserving PINN training and independent assessment. The two-output representation enforces incompressibility and wall velocities by construction. SOAP uses periodically redrawn regional batches, followed by self-scaled Broyden optimization on a fixed point set. Equation references identify the definitions used at each stage. The shown iteration budgets describe the final training phase; source initialization and any earlier reconditioning are workflow-specific.

Optimization on a regional objective

The iteration index is kk, gk=∇θJ(θk)\boldsymbol{g}_k = \nabla_{\boldsymbol{\theta}}J(\boldsymbol{\theta}_k) is the loss gradient, HkH_k approximates the inverse Hessian, dk\boldsymbol{d}_k is the search direction, and αk\alpha_k is its line-search step length. The constants c1c_1 and c2c_2 control sufficient decrease and curvature. The update vectors are sk=θk+1−θk\boldsymbol{s}_k = \boldsymbol{\theta}_{k+1} - \boldsymbol{\theta}_k and yk=gk+1−gk\boldsymbol{y}_k = \boldsymbol{g}_{k+1} - \boldsymbol{g}_k; this bold yk\boldsymbol{y}_k is not the spatial coordinate yy. The symbol qq denotes the sampling probability density. Its components qΩq_{\Omega}, qupperq_{\mathrm{upper}}, qfloorq_{\mathrm{floor}} and qcornerq_{\mathrm{corner}} correspond to the full domain Ω\Omega, upper strip, floor strip and lower-corner neighborhoods, respectively. SOAP denotes the Shampoo-based optimizer with Adam-like updates [37]; SSB denotes self-scaled Broyden optimization in the retained state labels and tables. Optimizer studies motivate examining such staged training separately from physical accuracy [35, 36]. The regional sampler can be

written as the mixture

q=0.50qΩ+0.30qupper+0.15qfloor+0.05qcorner.(20)q = 0.50q_{\Omega} + 0.30q_{\mathrm{upper}} + 0.15q_{\mathrm{floor}} + 0.05q_{\mathrm{corner}}. \tag*{(20)}

The first component covers the full cavity, the next two cover one-width upper and lower strips, and the last samples the bottom corners logarithmically in radius. Component regions overlap: points in a floor strip can also be drawn from the full-domain component. Their proportions describe allocation of training effort, not a partition of the physical domain. After rejecting the prescribed upper-corner disks, each draw has equal weight in the empirical objective. Increasing the floor and corner sampling probability therefore increases their contribution to optimization. We do not apply importance weights to convert the mixture back to an area-uniform integral.

All four detailed Re=1000\mathrm{Re} = 1000 workflows use the balanced mixture in their final phase. The refined initialization comes from an earlier Re=1000\mathrm{Re} = 1000 model trained with 50% full-domain, 30% floor-strip and 20% lower-corner points. Control and soft-residual cases instead start from the same specified Re=500\mathrm{Re} = 500 field. The half-envelope workflow includes additional Re=500\mathrm{Re} = 500 reconditioning because changing EE changes the physical initial velocity. The four retained final histories each record completion of 10,000 SOAP and 20,000 SSB steps. The refined initializer is identified by its contract hash, but its link to the earlier 20,000- versus 30,000-step precursor has not been recovered in the inspected archive. The half-envelope wrapper configures an additional 10,000-SOAP/20,000-SSB reconditioning phase; its matching warm history is not present in the inspected folders, so that configuration is not a verified cumulative budget. Thus control versus soft residual is the closest initialization-matched comparison; control versus refined measures a combined workflow, not the isolated benefit of regional sampling.

Training first applies SOAP for 10,000 steps with learning rate 0.002 and a new regional batch every 50 steps. The retained control, soft-residual, refined and half-envelope histories use SOAP batches of 131,072 points; the original Re=250\mathrm{Re} = 250 and 500 histories use 262,144 points. The subsequent fixed-point phase uses N=131072N = 131072 in both groups. SOAP performs adaptive updates in a preconditioned coordinate system, providing an initial reduction of the weighted momentum residual. The subsequent 20,000-step self-scaled Broyden phase uses a fixed point set. For gradient gkg_k and inverse-Hessian representation HkH_k, its trial direction and step are

dk=−Hkgk,θk+1=θk+αkdk.(21)d_k = -H_k g_k,\qquad\theta_{k+1} = \theta_k + \alpha_k d_k. \tag*{(21)}

The line search uses sufficient-decrease and curvature tests with c1=10−4c_1 = 10^{-4} and c2=0.9c_2 = 0.9. The parameter and gradient differences sks_k and yky_k update the self-scaled curvature representation. The supplied implementation stores vector generators and small coefficient blocks rather than a dense parameter-by-parameter matrix. The intended phase preserves its accumulated generators rather than using a short rolling history. Fixed collocation points make consecutive gradients refer to the same objective during this refinement.

The full N=131072N = 131072 point loss is evaluated in chunks of 65,536 to control differentiation memory. Each chunk contributes its residual sum divided by the same global NN, so chunking preserves the full-batch objective and gradient. Training uses seed 12345, with a separate held-out stream at seed 20264. A change in Reynolds number or residual definition begins a new optimizer phase; continuation within a phase preserves its network and optimizer state. This separates physical transfer of a learned field from continuation of the same minimization problem.

Figure 2 defines the common domain and the height measured upward from the floor. The moving upper wall supplies the forcing, while the stationary sidewalls and floor confine the return flow. The lid arrow denotes the direction of the regularized boundary motion, not a spatially uniform velocity at the wall junctions.

Figure 2. Geometry shared by the Re = 100, 250, 500 and 1000 cases: width WW, depth D=5WD=5W and height hh measured from the floor. The upper wall moves with the regularized profile Ulid(x)U_{\mathrm{lid}}(x); both sidewalls and the floor are stationary. The lid arrow indicates its direction, not a uniform corner velocity.

Independent CFD and evaluation protocol

Spectral/hp reference and sensitivity cases

In Eq. (22), fhf_h is a discrete scalar field on element Ωe\Omega_e, ee is the element index, f^ije\hat{f}_{ij}^{e} are its expansion coefficients, ϕi\phi_i and ϕj\phi_j are basis functions, and PP is polynomial degree. The indices i,ji,j enumerate modes from zero to PP. Here the subscript hh denotes a discrete field, not height above the floor.

Spectral/hp discretization and velocity correction

Superscripts nn and n+1n+1 identify successive time levels separated by time step Δt\Delta t. The vector A\mathbf{A} is the explicit multistep combination defined below and N\mathcal{N} is the convective operator. In Eq. (26), uhu_h denotes one discrete velocity component, FhF_h its assembled right-hand-side forcing, and ww a test function. The inner product is (a,b)=∫Ωab dΩ(a,b)=\int_{\Omega}ab\,\mathrm{d}\Omega, with the dot product understood for vector arguments. In the velocity solve, FhF_h represents the corresponding component of A/Δt−∇pn+1\mathbf{A}/\Delta t-\nabla p^{n+1}, including the boundary lifting. The reference calculations use the incompressible Navier–Stokes solver in Nektar++ 5.9.0, with a continuous Galerkin projection on structured quadrilateral elements. Within an element, each field is expanded in a tensor product of modified polynomial basis functions [45],

fh∣Ωe=∑i=0P∑j=0Pf^ijeϕi(ζ1)ϕj(ζ2),(22)f_h\big|_{\Omega_e}=\sum_{i=0}^{P}\sum_{j=0}^{P}\widehat{f}_{ij}^{e}\phi_i(\zeta_1)\phi_j(\zeta_2), \tag*{(22)}

where (ζ1,ζ2)(\zeta_1,\zeta_2) are element coordinates. The session uses P+1P+1 modes in each direction for u,v,pu,v,p: seven modes for the order-six calculation and nine for order eight. Element-interface assembly enforces the continuous approximation. The base mesh has 32×16032 \times160 elements; the mesh-refined calculation has 48×24048 \times240. For D/W=5D/W=5, both maintain the same horizontal-to-vertical element-size ratio. Spectral/hp dealiasing is enabled for the nonlinear terms.

The session explicitly selects the nonlinear unsteady equations, the velocity-correction scheme [46] and second-order implicit–explicit time integration. This treats advection explicitly and viscous diffusion implicitly. In a second-order multistep notation, let

A=2un−12un−1−Δt[2N(un)−N(un−1)],N(u)=u⋅∇u.(23)\mathbf{A}=2\mathbf{u}^{n}-\frac{1}{2}\mathbf{u}^{n-1}-\Delta t\left[2\mathcal{N}\left(\mathbf{u}^{n}\right)-\mathcal{N}\left(\mathbf{u}^{n-1}\right)\right],\qquad\mathcal{N}(\mathbf{u})=\mathbf{u}\cdot\nabla\mathbf{u}. \tag*{(23)}

After startup, the split advance solves a pressure Poisson problem and one Helmholtz problem for each velocity component:

∇2pn+1=∇⋅AΔt,(24)\nabla^2p^{n+1}=\frac{\nabla\cdot\mathbf{A}}{\Delta t}, \tag*{(24)}
(32Δt−Re⁡−1∇2)un+1=AΔt−∇pn+1.(25)\left(\frac{3}{2\Delta t}-\operatorname{Re}^{-1}\nabla^2\right)\mathbf{u}^{n+1}=\frac{\mathbf{A}}{\Delta t}-\nabla p^{n+1}. \tag*{(25)}

Thus pressure and velocity are computed through separate elliptic systems instead of a monolithic coupled solve. The weak Helmholtz operator has the form

32Δt(w,uh)+Re⁡−1(∇w,∇uh)=(w,Fh),(26)\frac{3}{2\Delta t}(w,u_h)+\operatorname{Re}^{-1}(\nabla w,\nabla u_h)=(w,F_h), \tag*{(26)}

for test functions ww vanishing on velocity Dirichlet boundaries, with boundary lifting included in assembly. The pressure Poisson operator is similarly integrated against test functions and assembled over the elements. These discrete elliptic solves, rather than neural residual minimization, determine the reference fields.

Both velocity components have prescribed Dirichlet conditions: zero at stationary walls and the same regularized horizontal speed at the lid. Pressure boundaries carry Nektar++’s high-order pressure condition, identified by the Extensible Markup Language (XML) session’s USERDEFINEDTYPE="H". Although its XML entry contains VALUE="0", this is not a claim that the physical normal pressure derivative vanishes everywhere. The solver constructs the compatible high-order condition for velocity correction. The production session does not select a coupled Stokes solver, substepping or selective-frequency damping. The CFD calculation evolves the flow in time; the main PINN calculation minimizes the steady equations. Their comparison therefore retains the physical time of each CFD snapshot.

The reference uses Δt=2×10−4\Delta t=2\times10^{-4}; the time-refined case uses 10−410^{-4}. Order, mesh and time step are varied separately, with the same lid and viscosity. The labels “order eight”, “mesh refined” and “half time step” describe these three numerical controls; they are not different physical cases. Field export and comparison are performed at matched physical times, allowing discretization effects to be separated from temporal evolution. The sensitivity table and the near-lid and floor profiles examine both the dominant circulation and the derivative-sensitive pressure and vorticity fields.

We compare the sensitivity fields at t≃170t\simeq170; the later base field at t≃317t\simeq317 is used for the PINN comparisons. The numerical framework and splitting method follow [45, 46, 47, 48]. For the sensitivity calculation, exported fields are sampled on a common 201×1001201\times1001 physical grid with Δx=Δh=0.005W\Delta x=\Delta h=0.005W.

Duplicate element-boundary values are averaged; the velocity spread at duplicate points was zero. Bilinear resampling accommodates the different visualization-node locations and introduces an interpolation component. We therefore report a sampled vector norm rather than a spectral-element quadrature norm or a Richardson-extrapolated error estimate.

The polynomial and mesh refinements differ from the base field by approximately 0.054% and 0.065%, respectively (Table 4). Halving the time step gives a difference of 0.000312%. The fine-time-step solution changes by 0.002886% between t≃160t \simeq160 and 170, consistent with slowly evolving flow during this interval. These checks describe sensitivity at t≃170t \simeq170; the later PINN comparison at t≃317t \simeq317 is reported separately.

CaseOrderElementsΔt\Delta tInterior (%)Floor (%)
Base632×16032 \times1602×10−42 \times10^{-4}——
Polynomial refinement832×16032 \times1602×10−42 \times10^{-4}0.0537690.004485
Mesh refinement648×24048 \times2402×10−42 \times10^{-4}0.0651000.073659
Time-step halving632×16032 \times16010−410^{-4}0.0003120.000096924
t=160→170t = 160 \to170, same Δt/2\Delta t/2 mesh632×16032 \times16010−410^{-4}0.002886—

Table 4. Common-time CFD sensitivity at Re⁡=1000\operatorname{Re} = 1000, D/W=5D/W = 5, t≃170t \simeq170. Interior excludes a 0.02W0.02W wall band. Differences are relative sampled vector norms against the base field, except the last row, on the 201×1001201 \times1001 comparison grid. Floor uses 0.02<x/W<0.980.02 < x/W < 0.98 and 0.02<h/W<0.40.02 < h/W < 0.4; its normalization uses the base velocity RMS on those same samples. The last regional entry is unavailable because the t≃160t \simeq160 source field was not recovered.

The floor reference velocity RMS in these t≃170t \simeq170 samples is 3.125289×10−8U3.125289 \times10^{-8}U. The companion regional-sensitivity data report both floor cuts, interior and upper-left pressure offsets, pressure-shape discrepancies and sampled-vorticity differences. The profile overlays use a separate 321×1601321 \times1601 display grid; they do not define the norms in Table 4. These sampled common-time checks are not an uncertainty bound for the later t≃317t \simeq317 reference.

Figure 3 shows reference pressure and sampled-vorticity sensitivity at Re⁡=1000\operatorname{Re} = 1000 and t=170t = 170. The lines are almost indistinguishable. These common-time comparisons are distinct from the later base field at t≃317t \simeq317 used in the detailed PINN assessment.

Figure 3. Discretization sensitivity of pressure (top) and vorticity (bottom) at Re=1000\mathrm{Re}=1000, D/W=5D/W=5 and the matched physical time t=170t=170. Left: upper cavity; right: floor region. The base calculation (P=6P=6, 32×16032\times160 elements, Δt=2×10−4\Delta t=2\times10^{-4}) is compared with polynomial refinement (P=8P=8), mesh refinement (48×24048\times240 elements) and time-step refinement (Δt=10−4\Delta t=10^{-4}), varied separately. Identical contour levels are used for all four calculations in each panel. Pressure is referenced to the mean in each displayed region. All fields are sampled on a common physical grid; vorticity is evaluated using the same differentiation operator.

Field and regional comparison

Table 5 identifies the states behind the two neural assessments and distinguishes them from the CFD sensitivity study. The tabulated histories describe the retained preparation, not matched cumulative computational cost.

AssessmentRetained neural state and preparationReference and evaluation
Common seriesReynoldsOriginal envelope fields at Re⁡=100\operatorname{Re} = 100, 250 and 500; balanced control at 1000 initialized from 500121×605121 \times605 grid; times in Table 7; floor 0.02<h/W<0.650.02 < h/W < 0.65
Detailed control and soft residualSame specified Re⁡=500\operatorname{Re} = 500 source; same final balanced sampler; residual definitions differ161×801161 \times801 grid; CFD t≃317t \simeq317; floor 0.02<h/W<0.40.02 < h/W < 0.4
Detailedrefined workflowPrior floor-refined Re⁡=1000\operatorname{Re} = 1000 source, then balanced final samplingSame detailed grid, reference, gauge and masks
Detailedhalf envelopeChanged envelope and additional Re⁡=500\operatorname{Re} = 500 reconditioning before final phaseSame detailed grid, reference, gauge and masks
CFD sensitivityNo neural training; order, mesh and time-step variantsMatched t≃170t \simeq170; 201×1001201 \times1001 sampled grid

Table 5. Assessment provenance. The two neural assessments use different sampling grids and floor masks. The Reynolds-number series uses the control, not the refined field, at Re⁡=1000\operatorname{Re} = 1000.

The detailed Re=1000\mathrm{Re}=1000 workflow comparisons use a 161×801161\times801 grid and the CFD field at t≃317t\simeq317. The separate Reynolds-number assessment uses a 121×605121\times605 grid and the states and reference times in Table 7. These protocols are not interchangeable. Let R\mathcal{R} be a set of NRN_{\mathcal{R}} comparison-grid points, with ii indexing its points. The variables u,v,p,ωu,v,p,\omega are nondimensional as defined above; the superscript CFD identifies the reference and the unsuperscripted field is the PINN prediction. We report dimensional root-mean-square (RMS) discrepancies,

Re⁡\operatorname{Re}CFD timeNeural field
100120Original depth-envelope comparison arrays
250160Original formulation, completed 20,000-step SSB weights
500160Original formulation, completed 20,000-step SSB weights
1000317Balanced control, initialized from the completed Re⁡=500\operatorname{Re} = 500 field

Table 7. Retained sources for the common D/W=5D/W = 5 Reynolds-number assessment. Time is in units of W/UW/U. The saved CFD times differ across the series.

Eu(R)=U[1NR∑i∈R((ui−uiCFD)2+(vi−viCFD)2)]1/2,Ep(R)=ρU2[1NR∑i∈R(pg,i−pg,iCFD)2]1/2,Eω(R)=UW[1NR∑i∈R(ωi−ωiCFD)2]1/2.(27)\begin{aligned} E_u(\mathcal{R}) &= U\left[\frac{1}{N_{\mathcal{R}}}\sum_{i\in\mathcal{R}}\left((u_i-u_i^{\mathrm{CFD}})^2+(v_i-v_i^{\mathrm{CFD}})^2\right)\right]^{1/2},\\ E_p(\mathcal{R}) &= \rho U^2\left[\frac{1}{N_{\mathcal{R}}}\sum_{i\in\mathcal{R}}\left(p_{g,i}-p_{g,i}^{\mathrm{CFD}}\right)^2\right]^{1/2},\\ E_\omega(\mathcal{R}) &= \frac{U}{W}\left[\frac{1}{N_{\mathcal{R}}}\sum_{i\in\mathcal{R}}\left(\omega_i-\omega_i^{\mathrm{CFD}}\right)^2\right]^{1/2}. \tag*{(27)} \end{aligned}

The arithmetic mean uses the stated sampled grid and region; the vector velocity RMS includes both components without a factor of one half. Tables list Eu/UE_u/U, Ep/(ρU2)E_p/(\rho U^2) and EωW/UE_\omega W/U, so every entry has an explicit physical scale. These RMS quantities are absolute differences on the lid-based scales. The additional NRMSE diagnostics below divide by a fixed reference RMS on the same region, never by a local point value. This reporting convention does not change the predicted fields or their agreement with CFD. Small lid-scaled differences in the weak lower flow must be read together with the physical profiles, which retain its local amplitudes and shapes, and the matched-reference ratios in Table 11 and Appendix § 11.

WorkflowInteriorUpper-leftFloor h/W<0.4h/W < 0.4
Control0.056240.325530.03317
Soft residual0.039620.196380.17638
Refined0.026580.133630.02176
Half envelope0.038440.182660.09175

Table 11. Matched-region velocity NRMSE on the detailed 161×801161 \times801 grid, reported as dimensionless ratios. Each denominator is the CFD vector RMS on that same region with the original uniform sample weights. The original absolute RMS values remain in Table 10. Single retained workflows are compared; this is not an initialization-matched ablation.

The interior is the sampled region 0.02<x/W<0.980.02 < x/W < 0.98, 0.02<h/W<4.980.02 < h/W < 4.98; it excludes the 0.02W0.02W band adjacent to every wall and therefore excludes the geometric corners. This evaluation region is distinct from the corner mask used in the training residual: the reported field discrepancies are unweighted comparisons on the stated evaluation points, not masked training losses. Within this interior, the upper-left subregion selects x/W<0.25x/W < 0.25, h/W>4.75h/W > 4.75, and the floor subregion selects h/W<0.4h/W < 0.4. The upper-left metric therefore characterizes the near-lid interior, not the excluded geometric corner or wall band. Pressure comparisons subtract each field’s spatial mean over the common interior, denoted by pgp_g. The pressure difference then has zero interior mean; the same gauge is retained for upper-left and floor errors, with no additional regional centering in the global-gauge RMS calculations. Separately labeled shape diagnostics subtract the regional difference mean.

For a region R\mathcal{R}, write ⟨f⟩R=NR−1∑i∈Rfi\langle f\rangle_{\mathcal{R}} = N_{\mathcal{R}}^{-1}\sum_{i\in\mathcal{R}} f_i. The matched-reference diagnostics are

NRMSE⁡u=Eu/U⟨(uCFD)2+(vCFD)2⟩R1/2,NRMSE⁡ω=EωW/U⟨(ωCFD)2⟩R1/2.(28)\operatorname{NRMSE}_u = \frac{E_u/U}{\left\langle(u^{\mathrm{CFD}})^2 + (v^{\mathrm{CFD}})^2 \right\rangle_{\mathcal{R}}^{1/2}}, \qquad\operatorname{NRMSE}_{\omega} = \frac{E_\omega W/U}{\left\langle(\omega^{\mathrm{CFD}})^2 \right\rangle_{\mathcal{R}}^{1/2}}. \tag*{(28)}

Both numerator and denominator use identical points, masks and uniform sample weights. These dimensionless ratios complement Eq. (27); they are not pointwise relative errors and do not divide by a near-zero local velocity. Pressure NRMSE uses the RMS of the same globally aligned reference pressure; a separately identified shape NRMSE uses the regional mean-subtracted reference. Tiny shape denominators must be reported alongside their ratios.

Training residuals, held-out physical residuals and CFD discrepancies measure different quantities. The comparison does not rank cases by masked training loss. Streamlines are integrated from sampled physical velocities with shared seeds; they are supplemented by horizontal velocity, pressure and vorticity profiles at h/W=4.90h/W = 4.90, 4.954.95 and 4.984.98. To establish an elliptic or hyperbolic stagnation point quantitatively, one must solve for zeros of the differentiable velocity and classify the local Jacobian. A positive determinant is an elliptic candidate, and a negative determinant is a saddle candidate in this incompressible setting. A streamline plot alone supplies neither circulation nor a separatrix-area measurement.

Comparison-grid and interpolation sensitivity

The same retained Re = 1000 control source and CFD snapshot were sampled with the original bilinear procedure on the two comparison grids. Changing from 121×605121 \times605 to 161×801161 \times801 changes interior velocity, pressure and sampled-vorticity RMS by 9.78%, 4.80% and 20.46%, respectively (Table 6). The masks and gauge rule are unchanged, but the sampled locations and finite-difference operator change with the grid. The pressure value here uses the native CFD snapshot resampled directly, rather than treating the previously retained common-grid CFD interpolant as a new native field; it does not overwrite the Reynolds-series table.

GridEu/UE_u/UEp/(ρU2)E_p/(\rho U^2)EωW/UE_\omega W/U
121 × 6056.542×10−36.542 \times10^{-3}2.665×10−32.665 \times10^{-3}6.649×10−16.649 \times10^{-1}
161 × 8017.181×10−37.181 \times10^{-3}2.793×10−32.793 \times10^{-3}8.009×10−18.009 \times10^{-1}
321 × 16017.053×10−37.053 \times10^{-3}2.855×10−32.855 \times10^{-3}8.660×10−18.660 \times10^{-1}

Table 6. Comparison-grid sensitivity using the same retained control arrays and the same t≃317t \simeq317 CFD source, original bilinear sampling, interior mask, gauge, and uniform sample weights. Vorticity is recomputed from each sampled velocity grid. The 321×1601321 \times1601 row samples the retained 161×801161 \times801 neural-field interpolant; it is not direct network refinement or evidence of CFD/neural convergence.

The 321 × 1601 probe samples the retained 161 × 801 neural-field interpolant. It measures postprocessing sensitivity only: no new network inference or CFD solve was performed. Direct network inference on finer grids remains pending, and these three rows establish neither neural-field nor CFD convergence.

Results

The common-grid series establishes the regional and observable dependence across Reynolds numbers. The subsequent detailed assessment compares the retained training workflows at Re⁡=1000\operatorname{Re} = 1000 under its separate protocol.

A common assessment at Reynolds numbers 100, 250, 500 and 1000

We next compare the four retained D/W=5D/W = 5 cases with one postprocessing procedure. The Reynolds-number series uses the original envelope networks at Re⁡=100\operatorname{Re} = 100, 250 and 500 and the balanced control at Re⁡=1000\operatorname{Re} = 1000. The latter starts from the completed Re⁡=500\operatorname{Re} = 500 field and uses the regional training procedure described above. Thus, the series assesses the available workflows across Reynolds numbers; it does not isolate Reynolds number from the change in training procedure. Table 7 identifies the reference snapshots. All four reference fields use Nektar++ with the regularized lid in Eq. (4).

We sample every velocity and pressure field on the same 121×605121 \times605 grid, with 0.002≤x/W≤0.9980.002 \le x/W \le0.998 and 0.002≤h/W≤4.9980.002 \le h/W \le4.998. The retained Re⁡=100\operatorname{Re} = 100 arrays already use this grid. For Re⁡=250\operatorname{Re} = 250 and 500, we evaluate the completed neural weights in double precision on that grid; for Re⁡=1000\operatorname{Re} = 1000, we bilinearly sample the retained neural arrays. The CFD fields use the same bilinear sampling procedure. Each pressure field has one mean removed over 0.02<x/W<0.980.02 < x/W < 0.98, 0.02<h/W<4.980.02 < h/W < 4.98. We then compute ω=∂xv−∂hu\omega= \partial_x v - \partial_h u from every sampled velocity with the same second-order difference operator. This choice fixes the postprocessing operator, although interpolation and finite-difference sensitivity remain part of the sampled derivative discrepancy. The detailed workflow assessment instead uses 161×801161 \times801 points and a floor region 0.02<h/W<0.40.02 < h/W < 0.4; this series uses 121×605121 \times605 points and 0.02<h/W<0.650.02 < h/W < 0.65. Consequently, its control interior RMS discrepancies (6.542×10−3U6.542 \times10^{-3}U in velocity and 6.649×10−1U/W6.649 \times10^{-1}U/W in vorticity) differ from the detailed values (7.181×10−3U7.181 \times10^{-3}U and 8.009×10−1U/W8.009 \times10^{-1}U/W) without representing a different training result.

Figure 4 and Figure 5 compare the streamline geometry across the four Reynolds numbers. Four prominent alternating recirculating vortices are visible in the retained full-depth fields. With increasing Reynolds number, the intermediate and lower vortex centers shift upward, whereas the main lid-driven vortex becomes more confined to the upper cavity. Its shape changes from a vertically elongated circulation at Re⁡=100\operatorname{Re} = 100 toward a rounder core at Re⁡=500\operatorname{Re} = 500 and 1000. The CFD and PINN integral curves recover the same broad stack, with localized differences near the core and dividing streamlines. This comparison concerns geometry: equally visible lower vortices do not have equally strong velocities, as the depth-amplitude profiles show.

Figure 4. Full-depth streamline comparison at D/W=5D/W = 5, Re=100Re = 100, 250250, 500500 and 10001000, from left to right. Black solid integral curves are CFD, and orange dashed curves are PINN. Both use identical integration and automatic-seeding parameters on the frozen 121×605121 \times605 velocity grid. CFD times are tU/W=120tU/W = 120, 160160, 160160 and 317317, respectively. Streamline density is not a measure of velocity magnitude.

Figure 5. Upper-cavity streamline enlargement, 3.8≤h/W≤53.8 \le h/W \le5, for Re⁡=100\operatorname{Re} = 100, 250, 500 and 1000 at D/W=5D/W = 5. Colors, line styles and reference times are those of Figure 4. The enlarged view resolves changes in the primary circulation and differences between the CFD and neural fields.

These changes can be interpreted through the vorticity-transport balance in Eq. (8). Increasing ReRe reduces the coefficient of viscous diffusion relative to advection on the lid-based scale. Near the lid, rotation is then redistributed more strongly by the circulation, and sharper wall and turning-flow gradients can coexist with a rounder recirculating core. The lower vortices experience a much smaller local Reynolds number and respond to the upper flow through weak viscous coupling. Their displacement in Figure 4 therefore reflects a reorganization of the stack, not evidence that all vortices become equally inertia-dominated.

Figure 6 and Figure 7 compare upper-cavity and floor speed contours, respectively. The upper view begins at h/W=3.8h/W = 3.8 to omit the featureless lower portion, whereas the floor enlargement resolves the weak lower motion on its own physical amplitude scale. CFD and PINN contours use the same levels in each panel, and the floor range is shared across the four Reynolds numbers. A visible lower vortex can carry very little kinetic energy, so these amplitude comparisons complement the preceding streamline geometry. Contour settings and cropped bounds are recorded in figure_revision_20261011.json.

Figure 6. Speed contours over 3.8≤h/W≤4.9983.8 \le h/W \le4.998 at D/W=5D/W = 5, with Re⁡=100\operatorname{Re} = 100, 250, 500 and 1000 from left to right. Black solid lines are Nektar++ and orange dashed lines are PINN. The contour index jj enumerates the stated levels. Shared levels are 0.085573j0.085573j, j=1,…,9j = 1,\ldots,9, in units of UU. The blank lower portion is omitted.

Figure 7. Floor speed enlargement at D/W=5D/W = 5, Re⁡=100\operatorname{Re} = 100, 250, 500 and 1000. The upper row is omitted because it repeats Figure 6. Line styles are those of that figure. Levels are 1.4844×10−12105j/61.4844 \times10^{-12} 10^{5j/6}, j=0,…,6j = 0,\ldots,6, in units of UU.

Figures 8 and 9 show upper-cavity and floor pressure contours. In the upper cavity, pressure gradients redirect the faster fluid around the main vortex and near the sidewalls. The upper view begins at h/W=4h/W = 4; the floor view removes the mean of each displayed strip to expose its weak spatial variation. This local centering is for visualization only: pressure errors and horizontal profiles retain the single interior gauge. Shared contour levels permit comparison of the spatial patterns without rescaling one model to match another.

Figure 8. Pressure contours over 4≤h/W≤4.9984 \le h/W \le4.998 at D/W=5D/W = 5, Re⁡=100\operatorname{Re} = 100, 250, 500 and 1000, in units of ρU2\rho U^2 after interior mean alignment. Black solid lines are CFD; orange dashed lines are PINN. Shared levels are 0.027731j0.027731j, j=−5,…,5j = -5,\ldots,5.

Figure 9. Floor pressure-contour enlargement at D/W=5D/W = 5, Re⁡=100\operatorname{Re} = 100, 250, 500 and 1000 with the mean of each displayed strip removed. The repeated upper row is omitted. Levels are ±1.1137×10−14105j/6\pm1.1137 \times10^{-14} 10^{5j/6}, j=0,…,6j = 0,\ldots,6, shared across the four cases in units of ρU2\rho U^2.

Figures 10 and 11 compare full-depth sampled vorticity and its enlarged upper and floor views. The full-depth view is retained because sign changes occur throughout the cavity; the enlarged views distinguish the strong near-lid gradients from weak lower structure. Both sampled velocity fields use the same differentiation operator, and each panel uses contour levels shared across Reynolds numbers.

Figure 10. Full-depth sampled vorticity at D/W=5D/W=5, Re⁡=100\operatorname{Re}=100, 250, 500and$1000and \$1000 in units of U/WU/W. The same velocity-difference operator is used for all cases. Black solid contours denote CFD and orange dashed contours PINN. Shared levels are 1.24935j1.24935j, j=−5,…,5j=-5,\ldots,5.

Earlier CFD studies already investigated deep cavities across Reynolds numbers. Cheng and Hung considered depth-to-width ratios from 0.1 to 7 and Reynolds numbers from 0.01 to 5000 [10]. Their reported changes in near-lid vortex size and position, together with weaker inertial effects and more nearly symmetric vortices at greater depth, provide a useful qualitative comparison. Zhuo et al. considered ratios from 1.5 to 7 and Reynolds numbers from 100 to 5000, examining vortex positions, strengths and sizes [11]. The stacked weak circulation and stronger deformation near the lid in the present fields are consistent with that broad picture. Different lid treatments and reference times preclude a quantitative identification of vortex centers or circulation with those studies. The present contribution is the neural workflow and regional multifield assessment, not the discovery of deep-cavity recirculation.

Figure 11. Upper and floor vorticity-contour enlargements for Re = 100, 250, 500 and 1000. Upper levels are 3.93427j3.93427j, j=−5,…,5j=-5,\ldots,5; floor levels are ±5.3412×10−12105j/6\pm5.3412 \times10^{-12} 10^{5j/6}, j=0,…,6j=0,\ldots,6, in units of U/WU/W. Shared levels permit comparison of shape and amplitude across Reynolds numbers.

Figure 12–Figure 14 resolve both upper corners on the retained common 121×605121 \times605 grid. With the nondimensional fields and the same pressure gauge as Eq. (27), define the dimensionless pointwise discrepancies directly as

Dimensionless velocity-vector discrepancy

Figure 12. Dimensionless velocity-vector discrepancy eue_u from Eq. (29) at D/W=5D/W = 5, Re⁡=100\operatorname{Re} = 100, 250, 500 and 1000 from left to right. Upper-left windows (0≤x/W≤0.20 \le x/W \le0.2, top row) and upper-right windows (0.8≤x/W≤10.8 \le x/W \le1, bottom row) both cover 4.8≤h/W≤54.8 \le h/W \le5 with equal spatial aspect ratios. The shared linear scale retains the full interior maximum 0.5746; gray is the original excluded wall band. Interior RMS discrepancies remain 8.049×10−3U8.049 \times10^{-3}U, 7.674×10−3U7.674 \times10^{-3}U, 9.541×10−3U9.541 \times10^{-3}U and 6.542×10−3U6.542 \times10^{-3}U.

Dimensionless sampled-vorticity discrepancy in upper-left and upper-right lid-corner windows

Figure 14. Dimensionless sampled-vorticity discrepancy eωe_\omega, with the same upper-left (top) and upper-right (bottom) windows and Reynolds-number columns. Both sampled velocity fields retain the same second-order differentiation operator. The shared linear scale retains the full interior maximum 72.27. Interior RMS discrepancies remain 0.5196U/W0.5196U/W, 0.5101U/W0.5101U/W, 0.5624U/W0.5624U/W and 0.6649U/W0.6649U/W. Gray denotes the original excluded wall band; the original samples are shown without smoothing.

eu=∥uPINN−uCFD∥2,ep=∣pgPINN−pgCFD∣,eω=∣ωPINN−ωCFD∣.(29)e_u = \left\lVert\mathbf{u}^{\mathrm{PINN}}-\mathbf{u}^{\mathrm{CFD}}\right\rVert_2,\qquad e_p = \left|p_g^{\mathrm{PINN}}-p_g^{\mathrm{CFD}}\right|,\qquad e_\omega= \left|\omega^{\mathrm{PINN}}-\omega^{\mathrm{CFD}}\right|. \tag*{(29)}

The maps show eue_u, epe_p and eωe_\omega. The fields use the fixed physical scales UU, ρU2\rho U^2 and U/WU/W, respectively; the map values therefore involve no division by a local speed or other local reference value and are not percentages. Their dimensional RMS discrepancies are Eu(R)=U⟨eu2⟩R1/2E_u(\mathcal{R})=U\langle e_u^2\rangle_{\mathcal{R}}^{1/2}, Ep(R)=ρU2⟨ep2⟩R1/2E_p(\mathcal{R})=\rho U^2\langle e_p^2\rangle_{\mathcal{R}}^{1/2} and Eω(R)=(U/W)⟨eω2⟩R1/2E_\omega(\mathcal{R})=(U/W)\langle e_\omega^2\rangle_{\mathcal{R}}^{1/2}, equivalent to Eq. (27); the brackets denote the arithmetic mean over the stated sampled region.

Each figure has upper-left and upper-right corner windows, 0≤x/W≤0.20 \le x/W \le0.2 and 0.8≤x/W≤10.8 \le x/W \le1, both with 4.8≤h/W≤54.8 \le h/W \le5. The two rows use equal spatial aspect ratios. Every observable retains one shared linear scale across all eight panels, from zero to its full observed interior maximum over the four Reynolds numbers; values are not clipped or rescaled to the local window. Cells represent the original samples without smoothing or interpolation. Gray retains the original excluded 0.02W0.02W wall strips rather than denoting zero error. These local windows resolve the sampled near-junction behavior; they do not establish zero discrepancy elsewhere. Table 8, the full-depth fields and the lower-flow profiles retain the separate bulk and weak-flow assessment.

ReEu/UE_u/UEp/(ρU2)E_p/(\rho U^2)EωW/UE_\omega W/U
1001.255×10−101.255 \times10^{-10}1.033×10−41.033 \times10^{-4}4.259×10−104.259 \times10^{-10}
2509.133×10−109.133 \times10^{-10}2.444×10−42.444 \times10^{-4}2.915×10−92.915 \times10^{-9}
5002.332×10−92.332 \times10^{-9}5.348×10−45.348 \times10^{-4}7.867×10−97.867 \times10^{-9}
10001.256×10−91.256 \times10^{-9}7.074×10−57.074 \times10^{-5}4.027×10−94.027 \times10^{-9}

Table 8. Floor RMS discrepancies, with 0.02<h/W<0.650.02 < h/W < 0.65 and 0.02<x/W<0.980.02 < x/W < 0.98. Entries use the physical scales specified in Eq. (27). Pressure retains the single interior mean alignment.

The velocity discrepancy is concentrated near the lid (Figure 12). The strip h/W>4.35h/W > 4.35 contains 12.69% of the interior samples and contributes 92.41–99.91% of the total squared velocity discrepancy. The corner windows resolve local structure within this upper flow; the full-depth streamlines, floor RMS discrepancies and lower-flow profiles test the circulation below it.

The pressure differences also cluster in the upper flow (Figure 13), with 91.02–99.97% of the squared discrepancy in the upper 0.65W0.65W. Their concentration is broader at Re⁡=500\operatorname{Re}=500: the narrower strip h/W>4.90h/W > 4.90 contributes 43.60%, compared with 88.93–98.48% in the other cases. The two corner windows show only part of this broader distribution. The locally centered floor pressure contours describe weak spatial shape, while these discrepancy maps retain the single interior gauge.

Dimensionless gauge-aligned pressure discrepancy

Figure 13. Dimensionless gauge-aligned pressure discrepancy epe_p, with the same upper-left (top) and upper-right (bottom) windows and Reynolds-number columns as Figure 12. Pressure retains the single full-interior mean alignment; no local recentering is applied. The shared linear scale retains the full interior maximum 0.6836. Interior RMS discrepancies remain 5.838×10−3ρU25.838 \times10^{-3}\rho U^2, 2.498×10−3ρU22.498 \times10^{-3}\rho U^2, 2.082×10−3ρU22.082 \times10^{-3}\rho U^2 and 2.665×10−3ρU22.665 \times10^{-3}\rho U^2. The broader Re⁡=500\operatorname{Re} = 500 pressure discrepancy also extends outside these windows.

Sampled-vorticity differences are more sharply localized (Figure 14). The strip h/W>4.90h/W > 4.90 contains 1.50% of the interior samples but contributes 99.16–99.95% of the squared discrepancy. The two neighborhoods within radius 0.1W0.1W of the lid corners contain 0.168% of the interior samples and contribute 95.42–99.74%. Thus, the global sampled-vorticity norm is dominated by a small region of strong near-junction gradients. It measures a real local departure that the bulk contour pattern alone cannot resolve, rather than a comparable discrepancy spread throughout the cavity.

The floor comparison has a different scale (Table 8). Its velocity RMS discrepancy ranges from 1.255×10−10U1.255 \times10^{-10}U to 2.332×10−9U2.332 \times10^{-9}U for the four retained fields. The RMS entries are candidate-minus-reference discrepancies, not the reference speed. They quantify agreement of the sampled fields under this protocol; small values on a lid-based scale alone do not certify the local accuracy of weak recirculation. The lower-flow profiles provide a complementary test of its amplitudes and shape. Floor pressure RMS includes the offset left after the single interior gauge alignment. We additionally separate that offset from local pressure shape, without replacing the original gauge-aligned metric.

Horizontal profiles resolve the stronger upper gradients directly (Figure 15). At h/W=4.5h/W = 4.5, the negative horizontal-velocity trough changes shape and shifts as Reynolds number increases. At h/W=4.98h/W = 4.98, the four CFD fields have different velocity and shear profiles immediately below the regularized lid. The neural curves recover much of this broad variation, with the clearest departures near the lateral walls at the highest cut. These pressure and sampled-vorticity profiles complement the spatial error maps by exposing the narrow near-junction differences on a physical amplitude scale. The deeper cuts in Figure 16 assess the floor and intermediate flow on their own scales.

Figure 15. Horizontal profiles at h/W=4.5h/W = 4.5, 4.94.9 and 4.984.98 at D/W=5D/W = 5, Re=100\mathrm{Re} = 100, 250250, 500500 and 10001000. Rows show u/Uu/U, interior-aligned p/(ρU2)p/(\rho U^2) and sampled ωW/U\omega W/U. Color denotes Reynolds number, solid lines CFD and dashed lines PINN. Each panel has one shared amplitude axis for all curves.

Horizontal velocity, pressure and vorticity profiles for Re = 100, 250, 500 and 1000 at three h/W values

Figure 16. Horizontal velocity, pressure and vorticity profiles at D/W=5D/W = 5, Re=100\mathrm{Re} = 100, 250, 500 and 1000 at h/W=0.1h/W = 0.1, 2.5 and 3.5. Colors and line styles match Figure 15. Scientific notation retains the weak floor amplitudes; pressure retains the global interior gauge rather than a separate gauge for each cut.

The paired cuts help interpret the contour changes quantitatively. In Figure 15, the trough at h/W=4.5h/W = 4.5 samples the return branch beneath the primary circulation, whereas the highest cut samples the forward lid-driven branch. Their changing widths and extrema indicate redistribution within the vortex, not merely an overall multiplication of the velocity. The pressure row tests the associated momentum redistribution, and the vorticity row is more sensitive to differences in profile slope. At h/W=2.5h/W = 2.5 and 3.53.5 in Figure 16, the same fixed horizontal line can sample different portions of an intermediate vortex as the stack shifts. Changes in sign or curve shape there should therefore be interpreted together with Figure 4, rather than as an isolated deterioration of the neural solution. The h/W=0.1h/W = 0.1 cut supplies the complementary weak-flow test near the floor.

The evaluation grid also affects a derivative comparison. As a diagnostic, we retain every second grid point in each direction and recompute vorticity from the coarser velocities. Relative to the fine-grid vorticity at those points, the interior change is 1.49%1.49\%, 2.87%2.87\%, 4.57%4.57\% and 6.71%6.71\% for CFD, and 8.82%8.82\%, 7.99%, 11.09% and 17.54% for PINN in increasing Reynolds-number order. This is a postprocessing-grid check, not a CFD mesh-refinement result. It explains why the global map RMS values in Figure 12–Figure 14 are reported with their grid and derivative definition. They should also be distinguished from the denser Re = 1000 workflow comparisons.

The four snapshots are at tU/W=120tU/W = 120, 160, 160 and 317. Their different times are part of the source definition. The present comparison describes these retained fields and does not use their ordering to establish a matched-time or asymptotically steady Reynolds-number trend. A temporal assessment would require a sequence of late fields at each Reynolds number. Within the available series, the consistent assessment already distinguishes good agreement in weak floor motion from the more demanding velocity-gradient comparison near the lid.

Here the superscripts PINN and CFD denote neural and reference pressure. For Δp=pgPINN−pgCFD\Delta p = p_g^{\mathrm{PINN}} - p_g^{\mathrm{CFD}} in the floor region, define bf=⟨Δp⟩fb_f = \langle\Delta p \rangle_f and ep,fshape=⟨(Δp−bf)2⟩f1/2e_{p,f}^{\mathrm{shape}} = \langle(\Delta p-b_f)^2 \rangle_f^{1/2}. These quantities obey

(ep,fglobal)2=bf2+(ep,fshape)2(30)\left(e_{p,f}^{\mathrm{global}}\right)^2 = b_f^2 + \left(e_{p,f}^{\mathrm{shape}}\right)^2 \tag*{(30)}

Here ⟨⋅⟩f\langle\cdot\rangle_f is the arithmetic mean over sampled floor-region points, bfb_f is the regional mean pressure offset, and ep,fglobal=⟨(Δp)2⟩1/2e_{p,f}^{\mathrm{global}} = \langle(\Delta p)^2 \rangle^{1/2} is the RMS pressure discrepancy under the single interior gauge. The superscript shape denotes the discrepancy after subtracting that regional offset; the subscript ff identifies the floor region. This decomposition distinguishes the regional offset from the locally gauge-aligned shape. An independent check on the frozen common arrays verifies this identity and source-content stability. Table 9 shows that the global-gauge floor RMS is dominated by the regional offset. The much smaller shape discrepancy describes these sampled pressure fields; it is not a bound on native pressure gradients or CFD discretization error.

ReFloor mean offset bfb_fLocal-gauge shape RMS
1001.033×10−41.033 \times10^{-4}1.152×10−101.152 \times10^{-10}
2502.444×10−42.444 \times10^{-4}8.709×10−128.709 \times10^{-12}
5005.348×10−45.348 \times10^{-4}8.174×10−128.174 \times10^{-12}
1000−7.074×10−5-7.074 \times10^{-5}3.965×10−123.965 \times10^{-12}

Table 9. Table 9: Additional floor pressure decomposition on the same grid and region as Table 8. Both quantities use units ρU2\rho U^2. The original globally gauge-aligned RMS remains in Table 8.

Deep-cavity agreement and regional trade-offs

Table 10 compares four completed training workflows. The control begins from the Re⁡=500\operatorname{Re} = 500 model. The refined workflow uses the regional strategy described above and begins from a prior floor-refined Re⁡=1000\operatorname{Re} = 1000 model. Thus the comparison concerns accumulated workflows, rather than an initialization-matched test of regional collocation. Interior velocity RMS discrepancy falls from 7.181×10−3U7.181 \times10^{-3}U to 3.394×10−3U3.394 \times10^{-3}U, and the upper-left value falls from 6.868×10−2U6.868 \times10^{-2}U to 2.819×10−2U2.819 \times10^{-2}U. Floor velocity RMS discrepancy falls from 5.506×10−10U5.506 \times10^{-10}U to 3.612×10−10U3.612 \times10^{-10}U. The refined workflow reaches an interior velocity RMS discrepancy of 3.394×10−3U3.394 \times10^{-3}U, while its upper-left interior value remains 2.819×10−2U2.819 \times10^{-2}U. The optimization histories in Appendix § 12 document loss reduction for the retained objectives; the field comparisons, rather than the loss ordering, establish these regional discrepancies.

WorkflowInitialization / residual choiceInteriorUpper-leftFloor
ControlRe⁡=500\operatorname{Re} = 500; MR/EMR/E7.181×10−37.181 \times10^{-3}6.868×10−26.868 \times10^{-2}5.506×10−105.506 \times10^{-10}
Soft residualRe⁡=500\operatorname{Re} = 500; MR/EMR/\sqrt{E}5.059×10−35.059 \times10^{-3}4.143×10−24.143 \times10^{-2}2.928×10−92.928 \times10^{-9}
Refined workflowPrior floor-refined Re⁡=1000\operatorname{Re} = 1000; MR/EMR/E3.394×10−33.394 \times10^{-3}2.819×10−22.819 \times10^{-2}3.612×10−103.612 \times10^{-10}
Half envelopeExtra Re⁡=500\operatorname{Re} = 500 reconditioning; half κ\kappa4.908×10−34.908 \times10^{-3}3.854×10−23.854 \times10^{-2}1.523×10−91.523 \times10^{-9}

Table 10. Table 10: Regional velocity RMS discrepancy Eu/UE_u/U against the same CFD field at t≃317t \simeq317, Re⁡=1000\operatorname{Re} = 1000, D/W=5D/W = 5, on the common 161×801161 \times801 grid. The interior excludes the 0.02W0.02W wall band; upper-left and floor regions are subsets and do not include the geometric corners.

The refined case starts from a floor-refined Re⁡=1000\operatorname{Re} = 1000 model, whereas the control starts from Re⁡=500\operatorname{Re} = 500. The reductions describe the combined effect of regional training and this initialization history. Separating the contribution of sampling alone would require repeating the comparison from identical initial parameters.

Softening the residual scale gives interior and upper-left velocity RMS discrepancies of 5.059×10−3U5.059 \times10^{-3}U and 4.143×10−2U4.143 \times10^{-2}U, respectively, but a floor value of 2.928×10−9U2.928 \times10^{-9}U. Reducing the envelope decay gives 4.908×10−3U4.908 \times10^{-3}U in the interior, 3.854×10−2U3.854 \times10^{-2}U in the upper-left region and 1.523×10−9U1.523 \times10^{-9}U at the floor. For these retained states, both variants improve the interior velocity RMS while increasing the floor discrepancy relative to the control. This observed change in ranking supports sensitivity to residual emphasis and representation; it does not establish an unavoidable trade-off or an optimal scale.

Full-depth structure and localized profiles

Figure 17 and Figure 18 show the CFD, control, soft-residual, refined-workflow and half-envelope fields with consistent plotting parameters. The full-depth view organizes the flow spatially; the two zooms expose near-lid recirculation and floor structures that would otherwise occupy only a small fraction of the figure. Streamlines are not fitted to the CFD and no vortex is inserted into the neural ansatz. These views should be read together with Table 10, because visually similar patterns can have different velocity amplitudes.

Figure 17. Full-depth streamlines for the CFD field at t≃317t \simeq317 and four completed PINN workflows at Re⁡=1000\operatorname{Re} = 1000, D/W=5D/W = 5. The refined model combines transfer initialization and regional training.

Figure 18. Upper and lower zooms at Re⁡=1000\operatorname{Re}=1000, D/W=5D/W=5, of Figure 17, with upper 4.35≤h/W≤54.35 \le h/W \le5 and lower 0≤h/W≤0.650 \le h/W \le0.65 views. Differences in streamlines complement, but do not replace, the amplitude errors in Table 10.

Figure 17 shows how confinement converts a single moving-wall forcing into a stack of recirculating vortices. The rightward lid motion drives the primary clockwise circulation: fluid turns downward near the right wall and returns toward the left beneath the lid. Since the cavity is closed, this transport must be balanced by return flow rather than by a net discharge through the domain. Motion along the lower flank of the primary vortex provides a tangential forcing for the fluid below it, where a counter-rotating vortex can form. Repetition of this coupling produces alternating rotation through the depth. The interfaces between vortices are internal dividing streamlines, not solid boundaries. Viscous stresses transmit motion across them, while the strength of the induced circulation decreases away from the moving wall. This interpretation explains why several lower vortices remain visible in the streamline plot even though their contribution to the total kinetic energy is small.

The two windows in Figure 18 separate physically different regimes. The upper window resolves the turning flow beneath the regularized lid and the near-wall branches of the main circulation. The rapid change from lid-driven motion to stationary sidewalls generates strong velocity gradients even though the prescribed lid is smooth. The floor window instead resolves the weak return flow constrained by the stationary bottom wall, including small corner recirculation where visible. Such corner eddies arise from the local viscous response to intersecting walls [3]; they should not be counted as additional full-width vortices in the deep-cavity stack. Shared streamline settings make the geometric comparison consistent, while the velocity profiles and regional norms establish whether the corresponding amplitudes agree.

The broad agreement among the panels in Figure 17 indicates that the neural fields reproduce this organization without prescribing the number or location of the vortices. For comparing workflows, the useful features are the shape of the primary circulation, the positions of the return-flow branches and the spacing of successive vortices. A local displacement of a dividing streamline redistributes motion between adjacent vortices and can produce a regional discrepancy even when the full-depth pattern remains recognizable. The refined workflow’s lower velocity discrepancies in Table 10 therefore complement its recovery of the overall streamline structure.

The near-lid profiles in Figure 19 compare velocity amplitudes at several heights across the rapidly varying upper flow. For all four workflows, the upper-left discrepancy exceeds the interior value. The refined procedure improves both measures, although the near-lid region remains the main source of disagreement with CFD.

Figure 19. Horizontal profiles at Re⁡=1000\operatorname{Re} = 1000, D/W=5D/W = 5, of velocity (top), pressure (middle) and vorticity (bottom) at three heights beneath the lid. Pressure gauges are aligned over the common interior. Vorticity uses the same second-order grid derivative for every field. Profiles are linearly sampled between grid heights.

The three rows of Figure 19 test related but distinct parts of the momentum balance. The horizontal-velocity profiles locate the positive lid-driven branch and the negative return-flow branch; changes in the positions of their extrema reveal a displacement of the circulation as well as an amplitude difference. Pressure provides the force needed to redirect the flow within the enclosure, while vorticity measures the local rotation and shear through Eq. (7). Comparing cuts at h/W=4.90h/W = 4.90, 4.954.95 and 4.984.98 shows how quickly these quantities change as the lid is approached. A field that follows the broad velocity profile can still differ in the narrow gradient regions, which become more apparent in the vorticity row. The refined curves improve the combined field comparison, especially in the upper-left region, rather than merely producing a visually similar central vortex.

Pressure and vorticity in the upper and lower cavity

Pressure and vorticity comparisons give a more demanding view of the neural fields than streamlines alone. Pressure is obtained directly from the second network output. Vorticity is evaluated from the sampled velocity derivatives, using the same operator for every model and the CFD reference. The pressure gauges are aligned over the interior before any discrepancy is calculated. For the contour plots, we also remove the mean within each displayed region so that the very small pressure variations near the floor remain visible.

Figure 20 compares upper pressure fields, while Figure 21 shows their associated vorticity. All models recover the broad pressure depression associated with the main circulation and the strong shear adjacent to the lid and right wall. The profiles in Figure 19 expose differences near the left lid corner that are difficult to judge from full-cavity streamlines. The control develops a larger local departure in velocity, pressure and vorticity. The soft-residual and refined fields reduce this departure, although the two procedures need not give the same ordering for every variable or region.

Figure 20. Pressure contours in the upper cavity, 4≤h/W≤54 \le h/W \le5, for CFD, control, soft-residual, refined and half-envelope fields at Re⁡=1000\operatorname{Re} = 1000, D/W=5D/W = 5. Pressure is mean-centered over the displayed region, with common contour levels. The featureless lower part is omitted from this view.

Figure 21. Full-depth vorticity contours at Re⁡=1000\operatorname{Re}=1000, D/W=5D/W=5, in the same model order as Figure 20. Values are in units of U/WU/W, computed with the same sampled-grid derivative. Enlarged views below resolve structures that occupy little space on this full-depth scale.

The pressure depression in Figure 20 is consistent with the inward acceleration required by the curved flow around the primary vortex. Pressure also redistributes momentum where the lid-driven branch turns into the sidewall flow; its contours therefore need not coincide with streamlines. Equation (6) connects this spatial variation to products of velocity gradients. Removing a constant pressure mean changes neither those gradients nor the momentum balance. In Figure 21, the strongest contours concentrate near the moving lid and the descending wall branch, whereas the deeper circulation occupies a much weaker vorticity range. A zero-vorticity contour marks a change in local rotation, but is not necessarily a streamline separating two vortices. Reading the pressure, vorticity and streamline panels together avoids identifying a shear feature as a distinct vortex.

Table 12 quantifies the pressure and sampled-vorticity differences. The refined workflow lowers interior pressure RMS discrepancy from 2.793×10−3ρU22.793 \times10^{-3}\rho U^2 to 1.167×10−3ρU21.167 \times10^{-3}\rho U^2. Its interior vorticity RMS discrepancy falls from 8.009×10−1U/W8.009 \times10^{-1}U/W to 3.142×10−1U/W3.142 \times10^{-1}U/W; the upper-left value falls from 7.747U/W7.747U/W to 3.025U/W3.025U/W. These different physical observables must not be ranked by comparing their numerical magnitudes across units. Vorticity involves second derivatives of the streamfunction, whereas velocity involves first derivatives, so small-amplitude spatial differences in the learned streamfunction can produce stronger curvature differences, especially near the lid. The common sampled differentiation operator makes the comparison consistent but does not remove sensitivity to field resolution. Among all four workflows, half envelope gives the smallest interior and upper-left pressure RMS, 1.064×10−3ρU21.064 \times10^{-3}\rho U^2 and 4.688×10−3ρU24.688 \times10^{-3}\rho U^2. The refined workflow has the smallest interior and upper-left sampled-vorticity RMS. The comparison supports a joint assessment of the physical fields rather than a single ranking based on speed.

Table 12: Pressure and sampled-vorticity RMS discrepancies at Re⁡=1000\operatorname{Re}=1000, D/W=5D/W=5, against CFD at t≃317t \simeq317. Entries use Eq. (27); pressure retains the single interior gauge alignment in every region.

WorkflowRegionEp/(ρU2)E_p/(\rho U^2)EωW/UE_\omega W/U
ControlInterior2.793×10−32.793 \times10^{-3}8.009×10−18.009 \times10^{-1}
Soft residualInterior1.195×10−31.195 \times10^{-3}3.669×10−13.669 \times10^{-1}
RefinedInterior1.167×10−31.167 \times10^{-3}3.142×10−13.142 \times10^{-1}
Half envelopeInterior1.064×10−31.064 \times10^{-3}3.583×10−13.583 \times10^{-1}
ControlUpper-left2.684×10−22.684 \times10^{-2}7.747×1007.747 \times10^{0}
Soft residualUpper-left7.798×10−37.798 \times10^{-3}3.536×1003.536 \times10^{0}
RefinedUpper-left9.974×10−39.974 \times10^{-3}3.025×1003.025 \times10^{0}
Half envelopeUpper-left4.688×10−34.688 \times10^{-3}3.453×1003.453 \times10^{0}
ControlFloor6.855×10−56.855 \times10^{-5}2.056×10−92.056 \times10^{-9}
Soft residualFloor2.963×10−42.963 \times10^{-4}2.131×10−82.131 \times10^{-8}
RefinedFloor1.802×10−41.802 \times10^{-4}3.025×10−93.025 \times10^{-9}
Half envelopeFloor3.131×10−43.131 \times10^{-4}1.297×10−81.297 \times10^{-8}

Table 12.

The detailed floor decomposition uses the same four workflows and original floor mask. The reference floor pressure has global-gauge RMS 8.336×10−3ρU28.336 \times10^{-3}\rho U^2, but regional shape RMS only 1.705×10−10ρU21.705 \times10^{-10}\rho U^2. Thus the offset-dominated global RMS and the tiny local-shape discrepancy answer different questions; neither replaces the other.

The lower cavity gives a different picture (Figure 22 and Figure 23). Its velocities and vorticities are many orders of magnitude smaller than those near the lid. The control and refined fields reproduce the main lower-region variations more closely than the soft-residual field. The soft-residual modification changes the horizontal velocity shape near the floor as well as the vorticity amplitude, consistent with its larger floor velocity RMS discrepancy. The control has the smaller sampled floor-vorticity RMS discrepancy, 2.056×10−9U/W2.056 \times10^{-9}U/W compared with 3.025×10−9U/W3.025 \times10^{-9}U/W for the refined workflow, despite the latter’s smaller floor velocity RMS. The ordering depends on whether field amplitudes or their spatial derivatives are examined.

Figure 22. Floor-region pressure (top) and vorticity (bottom) at Re⁡=1000\operatorname{Re} = 1000, D/W=5D/W = 5, in the same model order as Figure 20. Pressure means are removed within the displayed region to expose its weak spatial variation; quantitative pressure errors retain the interior-aligned gauge. The small vorticity amplitudes are labeled in units of U/WU/W.

Figure 23. Lower-cavity horizontal profiles at Re=1000\mathrm{Re} = 1000, D/W=5D/W = 5, of velocity (top), pressure variation about each cut’s mean (middle) and vorticity (bottom). Heights are h/W=0.05h/W = 0.05, 0.100.10, and 0.300.30. Scientific-notation axis scales retain the weak physical amplitudes instead of normalizing each curve.

Pressure is nearly constant across the lowest horizontal cuts. Figure 23 therefore shows pressure relative to the mean along each cut; the gauge-aligned RMS values remain in Table 12. A global pressure offset accumulated through the upper region can exceed the local floor variation, so these two measurements answer different questions. The line contours and centered profiles describe the shape of the weak pressure field, while the RMS table records its discrepancy under one consistent gauge. The independent regional decomposition also gives the control the smallest sampled floor pressure-shape RMS: 3.260444×10−12ρU23.260444 \times10^{-12}\rho U^2 (1.913% of the local CFD shape RMS), compared with 4.015840×10−12ρU24.015840 \times10^{-12}\rho U^2 (2.356%) for refined and larger discrepancies for soft residual and half envelope (Table 13). This concerns retained sampled fields, not a bound on native pressure gradients or robustness across seeds.

WorkflowOriginal Ep/(ρU2)E_p/(\rho U^2)bf/(ρU2)b_f/(\rho U^2)Shape RMS/(ρU2)(\rho U^2)
Control6.855×10−56.855 \times10^{-5}−6.855×10−5-6.855 \times10^{-5}3.260×10−123.260 \times10^{-12}
Soft residual2.963×10−42.963 \times10^{-4}2.963×10−42.963 \times10^{-4}4.129×10−114.129 \times10^{-11}
Refined1.802×10−41.802 \times10^{-4}1.802×10−41.802 \times10^{-4}4.016×10−124.016 \times10^{-12}
Half envelope3.131×10−43.131 \times10^{-4}3.131×10−43.131 \times10^{-4}3.293×10−113.293 \times10^{-11}

Table 13. Detailed floor pressure decomposition, on the original 0.02<x/W<0.980.02 < x/W < 0.98, 0.02<h/W<0.40.02 < h/W < 0.4 region. All values use ρU2\rho U^2. The signed offset is PINN minus CFD after the single interior gauge alignment; the shape RMS subtracts only this floor difference mean. The identity Ep2=bf2+Ep,shape2E_p^2 = b_f^2 + E_{p,\mathrm{shape}}^2 preserves the original gauge-aligned RMS.

The floor contours in Figure 22 and the cuts in Figure 23 test whether each workflow preserves weak-flow shape as well as magnitude. The stationary floor brings the velocity to zero, while the adjacent return flow produces a nonzero shear and hence vorticity above it. Pressure variations there are small because both inertial forcing and the transmitted motion are weak. Comparing the cuts at h/W=0.05h/W = 0.05, 0.100.10 and 0.300.30 follows the recovery from the near-wall layer into the lower vortex. The scientific-notation scales are essential: a visible difference on these axes can be important relative to the local circulation while remaining negligible on a lid-speed scale. The control and refined results retain this weak structure more closely than the softened-residual workflow, consistent with the regional velocity comparison.

Variation of physical amplitudes through the depth

The functions Au(h)A_u(h), Ap(h)A_p(h) and Aω(h)A_\omega(h) below are horizontal RMS amplitudes of speed, pressure variation and vorticity at height hh. The symbol ⟨p⟩x(h)\langle p\rangle_x(h) is the horizontal mean pressure at that height; the subscript xx identifies horizontal averaging, consistently with the region-average notation above. The upper pressure view and full-depth vorticity view (Figure 20 and Figure 21) complement the local comparisons. Most strong contours lie near the lid, where the pressure gradient sustains the main circulation and wall shear generates large vorticity. The pressure view omits the featureless lower portion; its weak variations appear in the separate floor enlargement. A common set of dimensionalized contour levels cannot display variations spanning many orders of magnitude with equal detail. The floor enlargement and full-depth amplitude profiles resolve this range without rescaling one model to match another.

The energy concentration explains the need for a separate weak-flow assessment. The upper 0.65W0.65W occupies 13%13\% of the cavity area, but contains 77.37%77.37\% of the sampled integral of u2+v2u^2+v^2 in the CFD field. The corresponding fractions are 77.49%77.49\%, 77.40%77.40\% and 77.40%77.40\% for the control, soft-residual and refined fields. The reference and these three workflows therefore give very similar measures of where most kinetic energy resides even when their floor-flow discrepancies differ considerably. Agreement in this integrated quantity alone would miss that difference.

At each height, define the horizontal RMS amplitudes

Au(h)=[1W∫0W(u2+v2) dx]1/2,(31)A_u(h)=\left[\frac{1}{W}\int_0^W (u^2+v^2)\,dx\right]^{1/2}, \tag*{(31)}
Ap(h)=[1W∫0W[p−⟨p⟩x(h)]2 dx]1/2,⟨p⟩x(h)=1W∫0Wp dx,(32)A_p(h)=\left[\frac{1}{W}\int_0^W [p-\langle p\rangle_x(h)]^2\,dx\right]^{1/2}, \qquad\langle p\rangle_x(h)=\frac{1}{W}\int_0^W p\,dx, \tag*{(32)}
Aω(h)=[1W∫0Wω2 dx]1/2.(33)A_\omega(h)=\left[\frac{1}{W}\int_0^W \omega^2\,dx\right]^{1/2}. \tag*{(33)}

Figure 24 evaluates these integrals by trapezoidal quadrature on the retained common grid. Its logarithmic horizontal axes preserve the physical amplitude variation with height. Only exact zero amplitudes are omitted from the logarithmic plot. The pressure amplitude is invariant under addition of a constant, so it measures horizontal pressure variation rather than the accumulated offset between the upper and lower regions. The gauge-aligned pressure discrepancy in Table 12 remains a separate measure of field agreement.

Figure 24. Horizontal RMS velocity, pressure variation and vorticity at Re⁡=1000\operatorname{Re}=1000, D/W=5D/W=5, over the complete cavity depth. Logarithmic amplitude axes expose weak lower flow on its original physical scale. All models use the same quadrature grid; pressure is centered separately along each horizontal cut for this amplitude diagnostic.

The amplitude curves connect the upper and floor views instead of treating them as unrelated windows. Local minima can accompany changes in the organization of the recirculation, but a minimum alone does not establish a new vortex or locate a separatrix. The profiles measure the magnitude of the horizontal field at that height; the streamline geometry supplies the complementary spatial structure. Both are needed when comparing a method that closely reproduces the dominant vortex but changes the weaker lower flow.

Figure 24 also explains the coexistence of an inertia-sensitive upper circulation and a weak viscous lower stack. A local Reynolds number based on a characteristic vortex speed UcU_c and vortex length LcL_c is Rec=Re(Uc/U)(Lc/W)Re_c = Re(U_c/U)(L_c/W). Thus, a large Reynolds number defined by the lid speed does not imply strong inertia in every lower vortex: as UcU_c decreases through the depth, viscous transport can dominate locally. This follows directly from the competing advection and diffusion terms in Eq. (8). The close upper-amplitude curves and the wider relative separation at low amplitudes are consequently compatible with good overall agreement and more sensitive regional differences.

The reference velocity amplitude is Au/U=0.2856A_u/U=0.2856 at h/W=4.7h/W=4.7 and 1.889×10−81.889\times10^{-8} at h/W=0.3h/W=0.3. Their ratio is approximately 1.51×1071.51\times10^7. Table 14 gives the same quantities for the neural fields. The similar upper amplitudes and visibly different lower amplitudes explain why a method can appear successful in a full-cavity image yet require a more demanding assessment near the floor. These are values at specified heights, not maximum velocities or an estimate of decay rate fitted across the depth.

Table 14: Horizontal RMS velocity amplitudes and upper-region energy concentration from the common sampled fields. The energy fraction uses the integral of u2+v2u^2+v^2 over h/W≥4.35h/W \ge4.35 divided by its full-depth integral.

FieldAu(4.7W)/UA_u(4.7W)/UAu(0.3W)/UA_u(0.3W)/UUpper energy (%)
CFD0.2856231.889×10−81.889 \times10^{-8}77.370
Control0.2841071.822×10−81.822 \times10^{-8}77.495
Soft residual0.2900451.563×10−81.563 \times10^{-8}77.404
Refined0.2882961.852×10−81.852 \times10^{-8}77.401

Table 14.

Discussion

What the workflow comparison establishes

The central result is the dependence of neural-field agreement on both region and observable. The refined workflow has the smallest detailed velocity and interior sampled-vorticity discrepancies. Half envelope gives the smallest interior and upper-left pressure RMS, whereas the control has the smallest globally gauge-aligned floor pressure RMS and the smallest floor sampled-vorticity RMS. These rankings describe one retained trajectory per workflow; they do not establish robustness across seeds or general superiority.

Attribution is essential to this interpretation. The hard-constrained representation, depth envelope, reciprocal residual scaling and staged SOAP–Broyden implementation are inherited components of the published formulation and retained starting implementation. The present evaluation concerns regional collocation, transferred fields and selected scale variants. All four detailed workflows use the same balanced sampler in their final phase. The refined initialization includes previous Re = 1000 training; the half-envelope case includes additional Re = 500 reconditioning. Equal final iteration counts therefore do not imply equal cumulative effort or isolate a sampling effect. The control/soft-residual pair is the closer initialization-matched comparison, whereas the other comparisons assess accumulated workflows. No efficiency advantage over CFD is inferred from these phase budgets.

Velocity agreement and sampled rotational structure

Velocity, gauge-aligned pressure and sampled vorticity test distinct features of the approximation. In the detailed refined field, an interior velocity RMS discrepancy of 3.394×10−3U3.394 \times10^{-3}U coexists with vorticity RMS discrepancies of 3.142×10−1U/W3.142 \times10^{-1}U/W in the interior and 3.025U/W3.025U/W in the upper-left region. The common Reynolds-number series gives interior velocity RMS discrepancies of 6.542×10−3U6.542 \times10^{-3}U to 9.541×10−3U9.541 \times10^{-3}U and vorticity RMS discrepancies of 0.5101U/W0.5101U/W to 0.6649U/W0.6649U/W. These are separate protocols, with the common series using the balanced control at Re = 1000, not the refined field. Physical-scale RMS and matched-reference NRMSE retain this distinction while exposing both absolute and relative regional differences.

The derivative diagnostic applies the same second-order finite-difference operator to both sampled velocity fields. It does not measure native spectral derivatives against neural automatic derivatives. The larger response of the PINN arrays to sampling-grid coarsening shows sensitivity of this diagnostic; this sampling sensitivity cannot be subtracted from the PINN–CFD discrepancies to obtain a corrected error. Nor does a common operator eliminate interpolation or sampling error. The excluded wall band also means that the regional norms do not assess wall shear or the geometric lid corners.

Pressure uses one interior gauge for regional error statistics. Locally centered contours and profiles instead display shape, and the floor decomposition reports the residual offset and shape separately. These displays must not be substituted for the globally aligned discrepancies. Similarly, streamlines locate the broad recirculation stack but do not quantify circulation, separatrix area or dynamical stability. The observed changes in vortex geometry are consistent with established deep-cavity flow, rather than evidence of a newly discovered mechanism.

The spatial evidence places the main accuracy limitation near the lid, especially at its junctions with the sidewalls (Figure 12–Figure 15). The prescribed omission of collocation points and the smooth mask reduce direct residual enforcement in the innermost corner neighborhoods. The mask is already close to unity at the interior points used for the field assessment, so the measured localization alone cannot isolate its role. A controlled comparison would be needed to separate the effects of point omission, residual weighting and finite-budget optimization. The reported near-junction pressure and derivative discrepancies therefore remain part of the achieved accuracy of this workflow, alongside its recovery of the full-depth circulation and its separate weak-floor velocity agreement.

Reference sensitivity and scope

The independent NekTar++ calculation provides a matched numerical reference, not an exact solution. Polynomial-order, mesh and time-step changes are compared at t≃170t \simeq170; the detailed neural assessment uses the later base field at t≃317t \simeq317. The small global velocity differences at the earlier time do not establish identical-time convergence of the later field or a regional uncertainty bound for floor velocity, pressure or vorticity. The very small floor RMS discrepancies in Table 8 are errors between sampled fields, not reference-speed measurements. The floor profiles and RMS discrepancies must be interpreted within these reference limits.

The Reynolds-number series covers one aspect ratio and four retained states with different CFD times and preparation histories. It demonstrates no monotonic error trend under this assessment, but cannot establish Reynolds-independent attainable accuracy. Some training losses were still decreasing at the retained stopping point (Appendix § 12); the comparisons therefore concern achieved finite-budget states rather than common terminal convergence or variability across independent repetitions.

Finally, solving steady equations is distinct from demonstrating temporal relaxation, uniqueness or physical stability. Deep-cavity Hopf thresholds depend on geometry and lid conditions [12, 13]; conventional-lid thresholds cannot be assigned to the Gaussian-regularized problem. The present snapshots do not locate an instability or characterize turbulence. These scope conditions delimit the reported assessment without changing the governing problem or the numerical acceptance criteria.

Conclusions

We assessed data-free PINNs for steady, two-dimensional recirculation in a cavity of depth-to-width ratio five at Re=100Re = 100, 250, 500 and 1000. The formulation and staged optimizers are inherited; the contribution is the matched evaluation of prescribed regional training, transferred initialization and selected scale variants through velocity, gauge-aligned pressure and sampled vorticity.

The common 121×605121 \times605 assessment gives interior velocity RMS discrepancies of 8.049×10−3U8.049 \times10^{-3}U, 7.674×10−3U7.674 \times10^{-3}U, 9.541×10−3U9.541 \times10^{-3}U and 6.542×10−3U6.542 \times10^{-3}U, respectively. Interior pressure RMS discrepancies range from 2.082×10−3ρU22.082 \times10^{-3}\rho U^2 to 5.838×10−3ρU25.838 \times10^{-3}\rho U^2, and sampled-vorticity RMS discrepancies range from 0.5101U/W0.5101U/W to 0.6649U/W0.6649U/W. Floor velocity RMS discrepancies are 1.255×10−10U1.255 \times10^{-10}U to 2.332×10−9U2.332 \times10^{-9}U; these are sampled-field differences, not measurements of the reference speed or bounds on its uncertainty. The lower-flow profiles retain the local amplitude and shape information that small lid-scaled RMS values alone cannot establish.

In the separate 161×801161 \times801 assessment at Re=1000Re = 1000, the refined workflow reduces interior, upper-left and floor velocity RMS discrepancies from 7.181×10−3U7.181 \times10^{-3}U, 6.868×10−2U6.868 \times10^{-2}U and 5.506×10−10U5.506 \times10^{-10}U to 3.394×10−3U3.394 \times10^{-3}U, 2.819×10−2U2.819 \times10^{-2}U and 3.612×10−10U3.612 \times10^{-10}U, respectively. Interior pressure RMS discrepancy falls from 2.793×10−3ρU22.793 \times10^{-3}\rho U^2 to 1.167×10−3ρU21.167 \times10^{-3}\rho U^2. Sampled interior vorticity RMS discrepancy decreases from 8.009×10−1U/W8.009 \times10^{-1}U/W to 3.142×10−1U/W3.142 \times10^{-1}U/W, while its upper-left value remains 3.025U/W3.025U/W. These improvements compare accumulated workflows, not isolated sampler changes. The control has the smallest globally gauge-aligned floor pressure RMS and the smallest floor sampled-vorticity RMS; half envelope has the smallest interior and upper-left pressure RMS. No configuration dominates every regional metric.

The resulting lesson is spatial as well as observable-dependent: the retained solutions recover the deep recirculation stack and weak floor motion, while a small near-lid region dominates the sampled derivative discrepancy. Accurate resolution of the lid-junction neighborhoods remains a target for improving these fields. Conclusions concern the retained finite-budget states, one geometry, the matched regularized lid and the stated reference snapshots. The reference-sensitivity calculation at t≃170t \simeq170, the sampled-grid derivative test and the training-loss histories characterize reference choice, postprocessing sensitivity and optimization progress, respectively. Together with the regional field comparisons, these checks define the scope of the assessment.

Acknowledgments

The calculations used the Unity research computing environment at the University of Massachusetts Amherst and University of Florida Research Computing resources. C.J.M. acknowledges support from the U.S. Department of Energy, Office of Fusion Energy Sciences, under Award No. DE-SC0024634.

Declaration of AI Use

The authors used OpenAI ChatGPT (Sol 6.1 and GPT-6 Astra Ultra) and Anthropic Claude Opus 5.5 to assist with code improvement and debugging, and with editing and improving the manuscript text. The authors take full responsibility for the final manuscript, code, results, and conclusions.

Reproducibility and data availability

Training uses seed 12345 and double precision; independent test points use seed 20264. We record the Reynolds number, initialization, residual mask, depth scale and sampling distribution for each case. A resumed phase retains its quasi-Newton state, while a new physical objective starts with a fresh optimizer. The evaluation array labeled “Re500” refers to initialization; its physical comparison is at Re = 1000.

The public project repository is https://github.com/Ehsan-Roohi/FlowMLLab. The separate private Cavity-PINN-CFD-Gallery repository organizes derived case figures and comparison summaries. The manuscript package records the selected source fields and figure settings in figure_manifest.json and multiphysics_analysis.json; build_figures.py and build_multiphysics_figures.py reproduce the publication graphics and build_manuscript.py assembles the LaTeX project. Training components are cavity.py, train.py and optimizer.py, with the regional distribution defined in pinn_d5_balanced_experiment.py and phase execution in pinn_d5_balanced_run.py. The Nektar++ case generator prepare_nektar.py records geometry, polynomial order, lid and time-integration settings. The public repository contains related teaching and validation infrastructure, not a complete release of the campaign training states. Repository access and the availability of individual raw fields or training states are distinct from the presence of derived figures; the package identifies the materials used for each reported comparison.

The comparison arrays use axes (x,h)(x,h); plotting routines transpose them to the (h,x)(h,x) convention. Pressure fields are evaluated neural outputs, not pressure reconstructed from plotted velocity. Vorticity is calculated as ∂xv−∂hu\partial_xv-\partial_hu with second-order central differences and second-order one-sided boundary differences on the common grid. Pressure discrepancies use the interior-aligned gauge; locally centered pressure contours remove an additional constant only for visualization. Line-contour levels are selected from the reference’s first to ninety-ninth percentiles in each displayed region and then shared across models. This exposes weak interior structure without changing field amplitudes; profiles and error calculations use all values.

The package includes build_multiphysics_figures.py, which reconstructs the pressure and vorticity comparisons, and multiphysics_analysis.json, which records source hashes, plotting settings and quantitative discrepancies. Raw fields and model states are not modified by these scripts.

For the four-Reynolds-number comparison, analyze_reynolds_series.py records source Secure Hash Algorithm 256-bit (SHA-256) hashes, array orientation, sampling, gauge alignment and derivative settings in reynolds_analysis.json. The completed Re = 250 and 500 weights are used only for inference, without loading an optimizer or changing a checkpoint. The retained common-grid arrays, reynolds_common_fields.npz, accompany the local analysis; raw weights and third-party source code are not included in the main manuscript compressed archive. The existing GitHub case provenance was checked against the downloaded accepted CFD fields. The archive contains the figures in Portable Document Format (PDF) and matching editable TikZ sources. The independent floor pressure decomposition is recorded in floor_pressure_shape_check.json; check_floor_pressure_shape.py verifies the gauge decomposition on the retained common arrays without changing them.

Supplementary Movies S1 and S2 accompany the Re = 1000, D/W=5D/W = 5 case. Movie S1 shows retained iterates of the refined PINN during self-scaled Broyden optimization, from iteration 25 to 20,000, with full-cavity, floor and upper-left views. Movie S2 shows the transient Nektar++ sequence over 1≤tU/W≤1881 \le tU/W \le188, with full-cavity and floor views; it precedes the later reference field used in the quantitative comparison.

Optimization iterations and physical time are distinct axes. Saved frames are held longer for readability without temporal interpolation. The PINN viewing copy applies a monotonic contrast adjustment inside the plot regions to make faint lines readable; contour positions and levels are unchanged. Source and viewing-copy hashes and presentation settings accompany the movies.

A separate, private author-held assessment bundle contains the four final Re=1000\mathrm{Re}=1000 weight files, final-phase histories and experiment contracts; retained common-grid and detailed fields; the t≃317t \simeq317 CFD snapshot; and a portable frozen-field postprocessing script with reference metrics and hash manifests. It supports assessment of the sampled results, not a complete campaign rerun. A retained Re=100\mathrm{Re}=100 CFD session segment is included only as an example. The matching Re=1000\mathrm{Re}=1000 session and restart inputs, full optimizer archives, and the two unresolved initialization-history links are not supplied. Inherited training and evaluator source with unresolved redistribution rights is excluded. This bundle has not been publicly released and has no public DOI; any release remains subject to author review of rights and scope. The companion regional-sensitivity CSV documents the newly checked t≃170t \simeq170 diagnostics, but does not supply the missing t≃160t \simeq160 source field.

The public viewing copies are Movie S1 (PINN optimization) and Movie S2 (transient CFD).

Matched-reference supplemental diagnostics

Table 15 uses the detailed 161×801161 \times801 fields and exactly the masks of Table 10 and Table 12. Sampled vorticity uses the common finite-difference operator, not native automatic or spectral derivatives. The CFD vorticity RMS denominators are 1.274, 3.408 and 1.064×10−71.064 \times10^{-7} in units U/WU/W for the interior, upper-left and floor regions. Pressure retains the global interior gauge except in the explicitly labeled regional-shape column. The floor shape denominator is 1.705×10−10ρU21.705 \times10^{-10} \rho U^2; its ratios should be read together with the absolute decomposition, not as a pressure-gradient accuracy claim.

WorkflowRegionNRMSEω_\omegaPressurePressure shape
ControlInterior0.628840.111780.11178
ControlUpper-left2.273462.053882.19930
ControlFloor0.019320.008220.01913
Soft residualInterior0.288050.047830.04783
Soft residualUpper-left1.037580.596820.64786
Soft residualFloor0.200300.035550.24218
RefinedInterior0.246700.046700.04670
RefinedUpper-left0.887660.763360.82864
RefinedFloor0.028430.021610.02356
Half envelopeInterior0.281330.042580.04258
Half envelopeUpper-left1.013220.358810.38424
Half envelopeFloor0.121870.037560.19314

Table 15. Matched-region NRMSE ratios for sampled vorticity, globally gauge-aligned pressure and regional pressure shape. Uniform weights and finite nonzero reference denominators are used throughout.

ReReRegionNRMSEu\mathrm{NRMSE}_{u}NRMSEω\mathrm{NRMSE}_{\omega}
100Interior0.076050.42775
100Upper-left0.282490.83999
100Floor0.013870.01109
250Interior0.067110.41104
250Upper-left0.322520.95728
250Floor0.029660.02972
500Interior0.078500.44876
500Upper-left0.391061.28010
500Floor0.049150.05299
1000Interior0.051300.53085
1000Upper-left0.297401.98359
1000Floor0.018740.01565

Table 16. Matched-reference velocity and sampled-vorticity NRMSE in the common 121×605121 \times605 Reynolds-number series. The original interior and upper-left masks are retained; here the floor uses 0.02<h/W<0.650.02 < h/W < 0.65, unlike the detailed-workflow floor mask. The Re=1000Re = 1000 field is the control. Each denominator is the CFD RMS on the identical sampled region.

Recorded optimization histories

All four detailed final histories record SOAP 10,000 and SSB 20,000 with seed 12345. These counters do not establish matching cumulative preparation costs: the refined initializer’s preceding 20,000/30,000-step ancestry and the half-envelope reconditioning history remain unlinked in the inspected archive. The retained wall_total timer starts after SOAP and measures accumulated SSB time only; it is not total training time or an efficiency comparison with CFD.

Figure 25 separates the SOAP and self-scaled Broyden histories rather than joining phases with different point distributions. It uses the saved training records for the original Re=250Re = 250 and 500 cases and for the Re=1000Re = 1000 control, soft-residual and refined workflows. During SOAP, periodic resampling produces visible fluctuations; the fixed-point Broyden phase decreases more smoothly. The plots retain the recorded losses without smoothing. The different residual scales, sampling distributions and initial fields belong to each workflow, so vertical separation between curves is not a comparison of physical field error. The smaller soft-residual loss, for example, accompanies a larger floor discrepancy than the refined workflow in Table 10.

Figure 25. Recorded weighted momentum-loss histories at D/W=5D/W = 5, Re=250\mathrm{Re} = 250, 500 and 1000, including the detailed Re=1000\mathrm{Re} = 1000 variants: changing-batch SOAP (left) and fixed-batch self-scaled Broyden (right). Curves use the retained completed histories without smoothing. The horizontal coordinate is the iteration within each phase; vertical axes are logarithmic. Different objectives and initialization histories prevent interpreting the loss ordering as a physical-accuracy ranking.

The sustained loss reduction in Figure 25 shows that the staged procedure progressively satisfies its weighted momentum objective. Resampling during SOAP changes the points at which that objective is evaluated, so short fluctuations can reflect a newly drawn regional batch as well as the parameter update. Once the points are fixed, the Broyden line search and curvature updates work on one discrete objective, which explains the smoother optimization history. The field comparisons independently quantify the achieved velocity, pressure and sampled-vorticity discrepancies; the loss curves alone do not establish reconstruction accuracy. Several histories are still decreasing at the retained stopping point. They document the achieved optimization progress, rather than a common terminal convergence criterion. Longer training may improve the fields, but the histories alone do not establish the magnitude or regional distribution of such an improvement; the reported discrepancies refer to the retained states. The regional plots locate the remaining differences, particularly in the strong near-lid gradients.

References

  1. [1]Burggraf, Odus R.. Analytical and numerical studies of the structure of steady separated flows. Journal of Fluid Mechanics 24 (1966) 113–151. doi: 10.1017/S0022112066000545.DOI
  2. [2]Pan, Frank; Acrivos, Andreas. Steady flows in rectangular cavities. Journal of Fluid Mechanics 28 (1967) 643–655. doi: 10.1017/S002211206700237X.DOI
  3. [3]Moffatt, H. K.. Viscous and resistive eddies near a sharp corner. Journal of Fluid Mechanics 18 (1964) 1–18. doi: 10.1017/S0022112064000015.DOI
  4. [4]Shankar, P. N.; Deshpande, M. D.. Fluid mechanics in the driven cavity. Annual Review of Fluid Mechanics 32 (2000) 93–136. doi: 10.1146/annurev.fluid.32.1.93.DOI
  5. [5]Ghia, U.; Ghia, K. N.; Shin, C. T.. High-Re solutions for incompressible flow using the Navier–Stokes equations and a multigrid method. Journal of Computational Physics 48 (1982) 387–411. doi: 10.1016/0021-9991(82)90058-4.DOI
  6. [6]Schreiber, R.; Keller, H. B.. Driven cavity flows by efficient numerical techniques. Journal of Computational Physics 49 (1983) 310–333. doi: 10.1016/0021-9991(83)90129-8.DOI
  7. [7]Botella, O.; Peyret, R.. Benchmark spectral results on the lid-driven cavity flow. Computers & Fluids 27 (1998) 421–433. doi: 10.1016/S0045-7930(98)00002-4.DOI
  8. [8]Erturk, E.; Corke, T. C.; Gökçöl, C.. Numerical solutions of 2-D steady incompressible driven cavity flow at high Reynolds numbers. International Journal for Numerical Methods in Fluids 48 (2005) 747–774. doi: 10.1002/fld.953.DOI
  9. [9]Bruneau, Charles-Henri; Saad, Mazen. The 2D lid-driven cavity problem revisited. Computers & Fluids 35 (2006) 326–348. doi: 10.1016/j.compfluid.2004.12.004.DOI
  10. [10]Cheng, M.; Hung, K. C.. Vortex structure of steady flow in a rectangular cavity. Computers & Fluids 35 (2006) 1046–1062. doi: 10.1016/j.compfluid.2005.08.006.DOI
  11. [11]Zhuo, C. S.; Zhong, C. W.; Cao, J.. Filter-matrix lattice Boltzmann simulation of lid-driven deep-cavity flows, Part I: Steady flows. Computers & Mathematics with Applications 65 (2013) 1863–1882. doi: 10.1016/j.camwa.2013.02.020.DOI
  12. [12]Zhuo, C. S.; Zhong, C. W.; Cao, J.. Filter-matrix lattice Boltzmann simulation of lid-driven deep-cavity flows, Part II: Flow bifurcation. Computers & Mathematics with Applications 65 (2013) 1883–1893. doi: 10.1016/j.camwa.2013.03.002.DOI
  13. [13]Lin, Li-Song; Chang, Hung-Wen; Lin, Chao-An. Multi relaxation time lattice Boltzmann simulations of transition in deep 2D lid driven cavity using GPU. Computers & Fluids 80 (2013) 381–387. doi: 10.1016/j.compfluid.2012.01.018.DOI
  14. [14]Lagaris, Isaac Elias; Likas, Aristidis; Fotiadis, Dimitrios I.. Artificial neural networks for solving ordinary and partial differential equations. IEEE Transactions on Neural Networks 9 (1998) 987–1000. doi: 10.1109/72.712178.DOI
  15. [15]Sirignano, Justin; Spiliopoulos, Konstantinos. DGM: A deep learning algorithm for solving partial differential equations. Journal of Computational Physics 375 (2018) 1339–1364. doi: 10.1016/j.jcp.2018.08.029.DOI
  16. [16]E, Weinan; Yu, Bing. The Deep Ritz method: A deep learning-based numerical algorithm for solving variational problems. (2017). https://arxiv.org/abs/1710.00211.
  17. [17]Raissi, Maziar; Perdikaris, Paris; Karniadakis, George Em. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics 378 (2019) 686–707. doi: 10.1016/j.jcp.2018.10.045.DOI
  18. [18]Karniadakis, George Em; Kevrekidis, Ioannis G.; Lu, Lu; Perdikaris, Paris; Wang, Sifan; Yang, Liu. Physics-informed machine learning. Nature Reviews Physics 3 (2021) 422–440. doi: 10.1038/s42254-021-00314-5.DOI
  19. [19]Cuomo, Salvatore; Di Cola, Vincenzo Schiano; Giampaolo, Fabio; Rozza, Gianluigi; Raissi, Maziar; Piccialli, Francesco. Scientific machine learning through physics-informed neural networks: Where we are and what’s next. Journal of Scientific Computing 92 (2022) 88. doi: 10.1007/s10915-022-01939-z.DOI
  20. [20]Jin, Xiaowei; Cai, Shengze; Li, Hui; Karniadakis, George Em. NSFnets (Navier–Stokes flow nets): Physics-informed neural networks for the incompressible Navier–Stokes equations. Journal of Computational Physics 426 (2021) 109951. doi: 10.1016/j.jcp.2020.109951.DOI
  21. [21]Wang, Sifan; Teng, Yujun; Perdikaris, Paris. Understanding and Mitigating Gradient Flow Pathologies in Physics-Informed Neural Networks. SIAM Journal on Scientific Computing 43 (2021) A3055–A3081. doi: 10.1137/20M1318043.DOI
  22. [22]Wang, Sifan; Yu, Xinling; Perdikaris, Paris. When and why PINNs fail to train: A neural tangent kernel perspective. Journal of Computational Physics 449 (2022) 110768. doi: 10.1016/j.jcp.2021.110768.DOI
  23. [23]Krishnapriyan, Aditi; Gholami, Amir; Zhe, Shandian; Kirby, Robert; Mahoney, Michael W.. Characterizing possible failure modes in physics-informed neural networks. (2021). https://arxiv.org/abs/2109.01050.
  24. [24]De Ryck, Tim; Jagtap, Ameya D.; Mishra, Siddhartha. Error estimates for physics-informed neural networks approximating the Navier–Stokes equations. IMA Journal of Numerical Analysis 44 (2024) 83–119. doi: 10.1093/imanum/drac085.
  25. [25]Sukumar, N.; Srivastava, Ankit. Exact imposition of boundary conditions with distance functions in physics-informed deep neural networks. Computer Methods in Applied Mechanics and Engineering 389 (2022) 114333. doi: 10.1016/j.cma.2021.114333.DOI
  26. [26]Jagtap, Ameya D.; Kharazmi, Ehsan; Karniadakis, George Em. Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems. Computer Methods in Applied Mechanics and Engineering 365 (2020) 113028. doi: 10.1016/j.cma.2020.113028.DOI
  27. [27]Kharazmi, Ehsan; Zhang, Zhongqiang; Karniadakis, George Em. hp-VPINNs: Variational physics-informed neural networks with domain decomposition. Computer Methods in Applied Mechanics and Engineering 374 (2021) 113547. doi: 10.1016/j.cma.2020.113547.
  28. [28]Moseley, Ben; Markham, Andrew; Nissen-Meyer, Tarje. Finite basis physics-informed neural networks (FBPINNs): A scalable domain decomposition approach for solving differential equations. Advances in Computational Mathematics 49 (2023) 62. doi: 10.1007/s10444-023-10065-9.DOI
  29. [29]Lu, Lu; Meng, Xuhui; Mao, Zhiping; Karniadakis, George Em. DeepXDE: A deep learning library for solving differential equations. SIAM Review 63 (2021) 208–228. doi: 10.1137/19M1274067.DOI
  30. [30]Wu, Chenxi; Zhu, Min; Tan, Qinyang; Kartha, Yadhu; Lu, Lu. A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering 403 (2023) 115671. doi: 10.1016/j.cma.2022.115671.DOI
  31. [31]Yu, Jeremy; Lu, Lu; Meng, Xuhui; Karniadakis, George Em. Gradient-enhanced physics-informed neural networks for forward and inverse PDE problems. Computer Methods in Applied Mechanics and Engineering 393 (2022) 114823. doi: 10.1016/j.cma.2022.114823.DOI
  32. [32]McClenny, Levi D.; Braga-Neto, Ulisses M.. Self-adaptive physics-informed neural networks. Journal of Computational Physics 474 (2023) 111722. doi: 10.1016/j.jcp.2022.111722.
  33. [33]Wang, Xili; Tang, Kejun; Zhai, Jiayu; Wan, Xiaoliang; Yang, Chao. Deep adaptive sampling for surrogate modeling without labeled data. Journal of Scientific Computing 101 (2024) 77. doi: 10.1007/s10915-024-02711-1.DOI
  34. [34]Rathore, Pratik; Lei, Weimu; Frangella, Zachary; Lu, Lu; Udell, Madeleine. Challenges in training PINNs: A loss landscape perspective. Proceedings of the 41st International Conference on Machine Learning (2024), pp. 42159–42191. https://proceedings.mlr.press/v235/rathore24a.html.arxiv.org/abs/2402.01868
  35. [35]Urbán, Jorge F.; Stefanou, Petros; Pons, José A.. Unveiling the optimization process of physics informed neural networks: How accurate and competitive can PINNs be?. Journal of Computational Physics 523 (2025) 113656. doi: 10.1016/j.jcp.2024.113656.DOI
  36. [36]Kiyani, Elham; Shukla, Khemraj; Urbán, Jorge F.; Darbon, Jérôme; Karniadakis, George Em. Optimizing the optimizer for physics-informed neural networks and Kolmogorov–Arnold networks. Computer Methods in Applied Mechanics and Engineering 446 (2025) 118308. doi: 10.1016/j.cma.2025.118308.DOI
  37. [37]Vyas, Nikhil; Morwani, Depen; Zhao, Rosie; Kwun, Mujin; Shapira, Itai; Brandfonbrener, David; Janson, Lucas; Kakade, Sham. SOAP: Improving and Stabilizing Shampoo using Adam. (2024). https://arxiv.org/abs/2409.11321.
  38. [38]Lee, Jongmok; Shin, Seungmin; Kim, Taewan; Park, Bumsoo; Choi, Ho; Lee, Anna; Choi, Minseok; Lee, Seunghul. Physics informed neural networks for fluid flow analysis with repetitive parameter initialization. Scientific Reports 15 (2025) 16740. doi: 10.1038/s41598-025-99354-5.DOI
  39. [39]McDevitt, Christopher J.; Fowler, Eric; Roy, Subrata. Physics-Constrained Deep Learning of Incompressible Cavity Flows. AIAA SCITECH 2024 Forum (2024), paper 1692. doi: 10.2514/6.2024-1692.DOI
  40. [40]Fowler, Eric; McDevitt, Christopher J.; Roy, Subrata. Physics-informed neural network simulation of thermal cavity flow. Scientific Reports 14 (2024) 15203. doi: 10.1038/s41598-024-65664-3.DOI
  41. [41]Takao, Ryuta; Ii, Satoshi. Fine-tuning physics-informed neural networks for cavity flows using coordinate transformation. Computers & Fluids 306 (2026) 106957. doi: 10.1016/j.compfluid.2025.106957.DOI
  42. [42]Pal, Ritik; Mukherjee, Soubhik; Dutta, Urmi; Choudhury, Arghya. Solving Navier–Stokes Equations Using Data-free Physics-Informed Neural Networks With Hard Boundary Conditions. (2025). doi: 10.48550/arXiv.2511.14497.
  43. [43]Sheikholeslami, Mohammad; Salehi, Saeed; Mao, Wengang; Eslamdoost, Arash; Nilsson, Håkan. State-of-the-art implementations of PINNs for the lid-driven cavity problem—a critical review with future perspectives. Results in Engineering 32 (2026) 112399. doi: 10.1016/j.rineng.2026.112399.DOI
  44. [44]Stužalec, Tomasz; Wójcik, Daria; Uriarte, Carlos; Łoś, Marcin; Paszyńska, Anna; Paszyński, Maciej. Reliable physics-informed neural networks for Navier–Stokes simulations. Can we trust AI-generated numerical simulations?. Journal of Computational Science 95 (2026) 102817. doi: 10.1016/j.jocs.2026.102817.
  45. [45]Karniadakis, George Em; Sherwin, Spencer J.. Spectral/hp element methods for computational fluid dynamics. Oxford University Press, 2 ed. (2005). doi: 10.1093/acprof:oso/9780198528692.001.0001.DOI
  46. [46]Karniadakis, George Em; Israeli, Moshe; Orszag, Steven A.. High-order splitting methods for the incompressible Navier–Stokes equations. Journal of Computational Physics 97 (1991) 414–443. doi: 10.1016/0021-9991(91)90007-8.DOI
  47. [47]Cantwell, C. D.; Moxey, D.; Comerford, A.; Bolis, A.; Rocco, G.; Mengaldo, G.; De Grazia, D.; Yakovlev, S.; Lombard, J.-E.; Ekelschot, D.; Jordi, B.; Xu, H.; Mohamied, Y.; Eskilsson, C.; Nelson, B.; Vos, P.; Biotto, C.; Kirby, R. M.; Sherwin, S. J.. Nektar++: An open-source spectral/hp element framework. Computer Physics Communications 192 (2015) 205–219. doi: 10.1016/j.cpc.2015.02.008.DOI
  48. [48]Moxey, David; Cantwell, Chris D.; Bao, Yan; Cassinelli, Andrea; Castiglioni, Giacomo; Chun, Sehun; Juda, Emilia; Kazemi, Ehsan; Lackhove, Kilian; Marcon, Julian; Mengaldo, Gianmarco; Serson, Douglas; Turner, Michael; Xu, Hui; Peiró, Joaquim; Kirby, Robert M.; Sherwin, Spencer J.. Nektar++: Enhancing the capability and application of high-fidelity spectral/hp element methods. Computer Physics Communications 249 (2020) 107110. doi: 10.1016/j.cpc.2019.107110.DOI

Paper details

vlnc:W_txx2ecbu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Sol 6.1, GPT-6 Astra Ultra, Anthropic Claude Opus 5.5

View statement
Looking for related papers…
Contents