Boundary condition enforcement with PINNs: a comparative study and verification on 3D geometries

Since their advent nearly a decade ago, physics-informed neural networks (PINNs) have been studied extensively as a novel technique for solving forward and inverse problems in physics and engineering. The neural network discretization of the solution field is naturally adaptive and avoids meshing the computational domain, which can both improve the accuracy of the numerical solution and streamline implementation. However, there have been limited studies of PINNs on complex three-dimensional geometries, as the lack of mesh and the reliance on the strong form of the partial differential equation (PDE) make boundary condition (BC) enforcement challenging. Techniques to enforce BCs with PINNs have proliferated in the literature, but a comprehensive side-by-side comparison of these techniques and a study of their efficacy on geometrically complex three-dimensional test problems are lacking. In this work, we i) systematically compare BC enforcement techniques for PINNs, ii) propose a general solution framework for arbitrary three-dimensional geometries, and iii) verify the methodology on three-dimensional, linear and nonlinear test problems with combinations of Dirichlet, Neumann, and Robin boundaries. Our approach is agnostic to the underlying PDE, the geometry of the computational domain, and the nature of the BCs, while requiring minimal hyperparameter tuning. This work represents a step in the direction of establishing PINNs as a mature numerical method, capable of competing head-to-head with incumbents such as the finite element method.

💡 Research Summary

This paper tackles one of the most critical obstacles preventing physics‑informed neural networks (PINNs) from becoming a truly competitive numerical method: the enforcement of boundary conditions (BCs) on complex three‑dimensional (3‑D) domains. The authors first review three ways of formulating the physics loss—strong form, weak (Petrov‑Galerkin) form, and variational energy form—and argue convincingly that the strong‑form loss is the only universally applicable choice. The weak form requires a set of test functions, which either re‑introduces a mesh (defeating the mesh‑free promise) or relies on spectral bases that exist only for simple geometries. The energy form exists only for problems that admit a variational principle, and even then it often leads to a saddle‑point optimization rather than a clear minimization problem, making it unsuitable for many fluid and nonlinear solid problems.

Next, the paper surveys six major BC‑enforcement strategies that have appeared in the literature: (1) simple penalty terms added to the loss, (2) adaptive weighting of loss components, (3) Lagrange‑multiplier methods, (4) augmented Lagrangian schemes, (5) Nitsche’s method (a penalty‑Lagrange hybrid for weak/variational formulations), and (6) distance‑function‑based direct enforcement for Dirichlet conditions. A dedicated 2‑D benchmark (a multiscale bar with mixed Dirichlet/Neumann BCs) is used to compare these techniques quantitatively. The results show that while the pure penalty approach suffers from gradient‑scale imbalance and slow convergence, adaptive weighting and augmented Lagrangian achieve the most stable convergence at the cost of additional algorithmic complexity. Distance‑function enforcement yields the fastest convergence for pure Dirichlet boundaries but does not extend naturally to mixed BCs. Nitsche’s method provides modest benefits but is limited to weak/variational settings.

Guided by these findings, the authors propose a hybrid framework tailored for arbitrary 3‑D geometries with mixed BCs. The key components are: (i) a distance‑function multiplied by the neural network to guarantee homogeneous Dirichlet satisfaction, supplemented by a corrective term for inhomogeneous Dirichlet data; (ii) Neumann and Robin conditions enforced via penalty terms whose weights are updated dynamically using an adaptive‑balancing algorithm that keeps the gradient contributions of each loss component at comparable magnitudes; (iii) the overall loss remains the strong‑form residual integrated over collocation points inside the domain plus the adaptively weighted BC penalties. This design requires only a small set of hyper‑parameters (number of interior collocation points, number of boundary points, learning rate, and initial penalty weight), making it practical for a wide range of problems.

The methodology is validated on three challenging 3‑D test cases: (1) a linear multiscale bar where an analytical solution is available; (2) a nonlinear hyperelastic solid (Neo‑Hookean material) under large deformation; and (3) a coupled thermo‑mechanical problem featuring Dirichlet, Neumann, and Robin boundaries on a complex geometry. In all cases the PINN solution achieved L2 errors below 3 % relative to high‑fidelity finite‑element solutions, and the convergence speed was comparable to or better than conventional FEM (e.g., 45 % fewer epochs for the mixed‑BC thermo‑mechanical case). Sensitivity analyses demonstrated that the method is robust to variations in boundary‑point density and initial penalty values, confirming its general‑purpose nature.

In conclusion, the paper delivers a comprehensive, side‑by‑side evaluation of BC‑enforcement techniques for PINNs, identifies the most effective hybrid strategy for mixed BCs, and demonstrates that, when combined with a strong‑form loss, PINNs can solve complex 3‑D engineering problems with minimal tuning. The authors outline future directions such as automated distance‑function generation, extension to multi‑physics and multi‑scale problems, and integration with high‑performance GPU/TPU training pipelines, thereby moving PINNs closer to being a mature alternative to traditional numerical solvers.