REVIEW 4 major objections 5 minor 3 references
Fully Differentiable Boundary Element Solver for Hydrodynamic Sensitivity Analysis of Wave-Structure Interactions
T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read MarineHydro.jl is a boundary element solver whose wave-force answers come with exact, automatically differentiated gradients.
desk verdict A genuinely useful differentiable BEM solver, but the 'exact gradient' headline overreaches where geometry passes through finite-differenced mesh Jacobians. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing mechanism is the free-surface Green's function rendered AD-differentiable. MarineHydro.jl implements both a global polynomial approximation and an exact expression for the Green's function, and because both are expressed through ordinary polynomial, Bessel, Struve, and elementary operations, reverse-mode automatic differentiation propagates derivatives through their evaluation and through the panel integration that assembles the influence matrices. Around the linear solve, the paper uses implicit differentiation rather than unrolling the iterative solver, so the adjoint equation reuses the same factorization and only needs the transpose matrix. A custom reverse-mode chain rule, whose Jacobian for mesh generation and hydrostatics is computed by finite differences, connects external geometry routines into the otherwise AD pipeline.
What would settle it
Run the same radius-gradient calculation with several different finite-difference step sizes in the mesh-generation Jacobian, and through a meshing map with a kink such as a panel collocation point crossing a seam as the radius varies; if the resulting AD gradient changes with step size or fails to match a complex-step or analytic gradient, the fully differentiable claim holds only for the hydrodynamic core, not for geometry design variables. A second test is to evaluate gradients at an irregular frequency for a surface-piercing body and check whether the reported $10^{-7}$ agreement with finite differences degrades.
Extended reading notes
Core claim
The core discovery is that a BEM solver for wave-structure interactions can be built so that differentiation is a core feature rather than a post-processing add-on. MarineHydro.jl assembles the influence matrices $S$, $D$, and $K$ through AD-differentiable evaluation and integration of the free-surface Green's function, then obtains the sensitivity of any scalar output by solving a discrete adjoint system $\lambda^T D = \partial\mathcal{J}/\partial\phi$ in addition to the forward system $D\phi - Sb = 0$. The total gradient $d\mathcal{J}/d\theta$ is assembled from $\lambda$, the solution $\phi$, and AD-computed partials $\partial S/\partial\theta$, $\partial D/\partial\theta$, and $\partial b/\partial\theta$, costing about one extra linear solve regardless of the number of design variables. The paper demonstrates this for two interacting heaving spheres and for wave-energy-converter power optimization, reporting absolute agreement with finite differences at or below $10^{-7}$ for added mass and damping gradients and below $10^{-9}$ for the power objective. A notable feature is that the gradients are exact with respect to the discretized numerical problem, which the authors describe as a deliberate property of their discretize-then-optimize approach.
Load-bearing premise
The load-bearing premise is that the finite-difference Jacobian used inside the custom reverse rule for mesh generation and hydrostatics is accurate enough that the reported gradients with respect to body radius and separation can be called exact; if the finite-difference step is poorly chosen or the meshing map is nonsmooth, those geometry gradients are no longer exact AD gradients even though the BEM core itself is.
Editorial extensions
If this is right
- For a fixed BEM discretization, gradients of hydrodynamic coefficients with respect to any differentiable input can be obtained at the cost of one extra linear solve, independent of the number of design variables.
- The same pipeline works for both direct and indirect boundary integral formulations and for both the approximate and exact Green's functions, with approximate-Green's-function gradients agreeing with exact-Green's-function gradients to within about 2%.
- Sensitivity maps of coupled added mass and damping for two heaving spheres show that off-diagonal interaction terms remain strongly oscillatory with separation, so plane-wave-approximation cutoffs should be used with care for mutual interaction terms.
- Gradient-based optimization of WEC mechanical power becomes feasible with exact BEM gradients; the demonstration converged in three L-BFGS iterations to a design favoring minimal radius and maximal separation.
- Because gradients scale with outputs rather than inputs, the method becomes increasingly advantageous relative to finite differences as the number of design variables grows, as shown by the dummy-variable scaling study.
Reading between the lines
- Extension beyond the paper: if the finite-difference Jacobian in the mesh chain rule were replaced by a natively AD mesh generator, the exact-differentiation claim would extend to arbitrary CAD parameters rather than just radius and separation.
- Extension beyond the paper: the same adjoint engine could serve gradient-based uncertainty quantification or gradient-enhanced surrogate training, neither of which the paper demonstrates.
- Extension beyond the paper: the paper's own caution about irregular frequencies points to a differentiable irregular-frequency removal scheme as the natural next stress test for the architecture.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper presents MarineHydro.jl, a Julia boundary element solver for linear potential-flow wave-structure interactions that supports direct and indirect boundary integral formulations, two free-surface Green's function options, and reverse-mode automatic differentiation for computing sensitivities. The main claimed novelty is the ability to compute hydrodynamic coefficients together with 'exact gradients' with respect to inputs such as wave frequency, body radius, and inter-body separation, using one extra linear solve via a discrete adjoint. The paper verifies the hydrodynamic coefficients against Hulme's analytical hemisphere results and Capytaine, and verifies gradients against analytical and finite-difference results. Two applications are presented: a sensitivity study of two identical floating spheres and a gradient-based optimization of the mechanical power of a pair of wave energy converters. The central claim, however, is qualified by the fact that mesh generation and hydrostatics are differentiated through custom finite-difference Jacobians, which conflicts with the abstract's 'fully differentiable / exact gradient' language for geometry variables.
Significance. The contribution is potentially useful: an open-source BEM solver with AD through the Green's function evaluation, influence-matrix assembly, and implicit linear solve would enable adjoint-based design optimization for wave-energy and offshore-structure problems. The manuscript includes concrete verification cases, including the exact analytical added-mass gradient for the Rankine-only sphere (Fig. 6), AD-vs-finitedifference agreement at the 1e-7 level for heave coefficients, and two worked case studies. These are genuine strengths. The significance is reduced, however, by the finite-difference treatment of mesh generation and hydrostatics: the exactness claim is only established for the hydrodynamic core, not for geometry design variables that pass through external meshing routines. The core idea remains valuable, and the issue is fixable by re-scoping claims or by implementing differentiable mesh/hydrostatics routines.
major comments (4)
- [Section 3.5 and Algorithm 1] The central claim of 'exact gradients' and 'fully differentiable' is not supported for geometry design variables. The custom reverse-mode rule for mesh generation computes its vector-Jacobian product using a finite-difference Jacobian ('The Jacobian of this function is then computed using finite differences'), and Section 4.5 states that hydrostatic gradients are also 'estimated using finite differences'. Consequently, gradients with respect to radius r and separation x in the case studies are not exact derivatives of the discretized BEM problem; they inherit truncation error and step-size sensitivity from those finite-difference Jacobians. The authors themselves acknowledge the limitation in Section 3.5 and in the conclusion, so the abstract and Section 5 should either qualify the exactness claim or the missing differentiability should be supplied.
- [Table 3 and Figs. 7-8] The AD-versus-finite-difference agreement reported for radius and separation is not an independent verification of geometry gradients. Both the 'AD gradient' (via the custom finite-difference mesh/hydrostatics rules) and the 'finite difference gradient' perturb the same external Capytaine routines, so the agreement to about 1e-9 in Table 3 primarily confirms consistency of the custom rule with a finite-difference perturbation, not exactness of the underlying mesh Jacobian. Please add an independent validation of the mesh and hydrostatics Jacobians, for example complex-step differentiation, analytic differentiation of the meshing map, or a Richardson extrapolation over finite-difference step sizes, and report the chosen step and its effect on the reported gradients.
- [Section 4.6 and Fig. 16] The scaling demonstration using 'dummy' design variables does not establish that reverse-mode AD keeps gradient cost constant for geometric design variables. The dummy variables apparently do not enter mesh generation or hydrostatics, while the actual mesh and hydrostatics Jacobians are computed by finite differences and therefore cost additional perturbations per geometric input. The constant-time result in Fig. 16 is expected for dummy variables that are not used in the computation; it does not demonstrate scalability with respect to body dimensions, shape, or layout variables. Please demonstrate scaling on actual geometric design variables or restrict the scalability claim to hydrodynamic and frequency variables.
- [Section 3.1 and Eq. (14)] The adjoint derivation itself is sound for the BEM core, but the paper should be explicit that Eq. (14) only gives exact gradients if all partial derivatives of S, D, and b are exact. Since the mesh and hydrostatics contributions use finite differences, the 'one extra linear solve yields exact gradients' statement in the introduction and abstract should be carefully scoped to the hydrodynamic part of the pipeline. The current wording overstates the result for geometry-coupled optimization.
minor comments (5)
- [Abstract and Section 5] The terms 'fully differentiable' and 'exact gradients' are used without the qualifications stated in Section 3.5; the abstract and conclusion should be aligned with the actual implementation, which is exact for the hydrodynamic core but finite-difference-based for mesh generation and hydrostatics.
- [Fig. 4 caption] There are typos in the caption: 'hydrodyamic' and 'with and analytical results' should read 'hydrodynamic' and 'with analytical results'.
- [Section 3.2] The regularization constant epsilon in the '(d + epsilon)' distance is a free parameter, but its value and its influence on gradient accuracy are not reported; please state the value used and include a brief sensitivity check.
- [Eq. (16) and Fig. 6] The analytical added-mass expression in Eq. (16) corresponds to a full sphere in unbounded fluid, while the surrounding text and Fig. 6 mention a 'hemisphere sphere'; please clarify which geometry is being benchmarked and make the notation consistent.
- [Section 4.2] The statement that the sensitivities of added mass and damping have a 'pi/2 phase difference' appears to be a general claim drawn from a specific numerical case; please add a supporting derivation or citation, or phrase it as an observation for the computed configuration.
Circularity Check
No circular derivation: the adjoint gradients are derived from the discretized BEM system, and the disclosed finite-difference mesh Jacobians are a scope limitation rather than a circular input.
full rationale
MarineHydro.jl's differentiation claim is built from the discretized BEM residual in Eq. (9). Equations (10)-(14) derive the adjoint by implicit differentiation of that residual; the adjoint variable is defined by λᵀD = ∂J/∂φ, not by the output gradients being predicted. This is the standard discretize-then-optimize adjoint and does not assume the target result. Forward verification is anchored externally: the hemisphere coefficients are compared with Hulme (1982), the Rankine-only sphere gradient is checked against the analytical expression in Eq. (16), and the Delhommeau Green's function is an independent exact formulation verified through the external Capytaine solver. The co-authored citations (Ancellin 2024 for BIE formulations; Ancellin and Dias 2019 for Capytaine) are not load-bearing: the BIE formulations are also cited to standard monographs, and Capytaine is an independent open-source solver with its own verification. The paper itself discloses the only real caveat in Sections 3.5 and 4.5: mesh and hydrostatic Jacobians are computed by finite differences, so the 'exact' and 'fully differentiable' labels overreach for radius and separation gradients, and the AD-versus-FD agreement in Table 3 for those variables is not a fully independent check. That is a correctness/scope limitation, not circularity: no equation defines the geometry-gradient prediction as the finite-difference gradient by construction, and the hydrodynamic core differentiation does not assume the gradients it reports. Hence no circular step is present.
Assumptions & free parameters
free parameters (3)
- Regularization constant epsilon in Green's function distance (d + epsilon) =
not reported
- Finite-difference step size for mesh/hydrostatics Jacobian =
not reported
- PTO controller settings (k_i = 0, d_i = diag(B)) =
k=0, d=diag(B)
assumptions (5)
- domain assumption Linear potential flow theory: incompressible, inviscid, irrotational fluid, linearized free-surface and body boundary conditions, infinite depth.
- domain assumption The Wu et al. (2017) global approximation of the free-surface Green's function is accurate enough for coefficients and for AD gradients (with agreement within about 2% of the Delhommeau exact Green's function reported in Appendix B).
- standard math Reverse-mode AD and ImplicitAD correctly differentiate the dense complex linear solve (Eq. (9)) and the Green's function integrals.
- domain assumption The reported sensitivities are not corrupted by irregular frequencies, which the paper acknowledges have not been removed.
- domain assumption The linear frequency-domain WEC dynamics (Eq. (20)) and the power formula (Eq. (19)) with a resistive PTO controller are adequate for the optimization demonstration.
Cite this review
Pith. "Pith review of Fully Differentiable Boundary Element Solver for Hydrodynamic Sensitivity Analysis of Wave-Structure Interactions." pith.science (2026). https://pith.science/paper/E7XEJYWI
@misc{pith2026250106988,
author = {Pith},
title = {Pith review of: Fully Differentiable Boundary Element Solver for Hydrodynamic Sensitivity Analysis of Wave-Structure Interactions},
year = {2026},
howpublished = {\url{https://pith.science/paper/E7XEJYWI}},
note = {Machine review of arXiv:2501.06988}
}
read the original abstract
Accurately predicting wave-structure interactions is critical for the effective design and analysis of marine structures. This is typically achieved using solvers that employ the boundary element method (BEM), which relies on linear potential flow theory. Precise estimation of the sensitivity of these interactions is equally important for system-level applications such as design optimization. Current BEM solvers are unable to provide these sensitivities as they do not support automatic differentiation (AD). To address these challenges, we have developed a fully differentiable BEM solver
Reference graph
Works this paper leans on
-
[2000]
Automatic differentiation of algorithms. J. Comput. Appl. Math. 124 (1), 171–190. Bezanson, Jeff, Edelman, Alan, Karpinski, Stefan, Shah, Viral B, 2017. Julia: A fresh approach to numerical computing. SIAM Review 59 (1), 65–98. Bigoni, Davide, 2015. Uncertainty Quantification with Applications to Engineering Problems (Ph.D. thesis). Technical University o...
arXiv 2017
-
[2008]
Adjoint: An approach for the rapid development of discrete adjoint solvers. AIAA J. 46 (4), 863–873. Martins, J.R.R.A., Kennedy, G.J., 2021. Enabling large-scale multidisciplinary design optimization through adjoint sensitivity analysis. Struct. Multidiscip. Optim. 64 (5), 2959–2974. Martins, Joaquim R.R.A., Ning, Andrew, 2022. Engineering Design Optimiza...
work page Pith review arXiv 2025
-
[2014]
Multidisciplinary design optimization of offshore wind turbines for minimum levelized cost of energy. Renew. Energy 68, 893–905. Babarit, Aurélien, Delhommeau, Gérard, 2015. Theoretical and numerical aspects of the open source BEM solver NEMOH. In: Proceedings of the 11th European Wave and Tidal Energy Conference. EWTEC2015, Nantes, France. Bartholomew-Bi...
work page 2015
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.