Why do we build models in two dimensions? The answer is simple: 2D environments are computationally cheap, visually intuitive, and mathematically elegant. Whether simulating fluid dynamics, optimizing structural designs, or training neural networks to recognize shapes, reducing a problem to a flat plane allows researchers to prototype algorithms rapidly and establish fundamental baselines.
However, the physical world is stubbornly three-dimensional. When we attempt to extend successful 2D frameworks to 3D spaces, we don't just face longer runtimes; we encounter a fundamental dimensionality gap. Moving from 2D to 3D causes mathematical guarantees to dissolve, computational costs to explode exponentially, and entirely new physical phenomena to emerge. Understanding why this gap exists—and why 2D solutions cannot simply be "extruded" into 3D—is critical across computational physics, robotics, and machine learning.
The Computational Explosion: The Curse of Dimensionality
The most immediate barrier to 3D extension is numerical scaling. In grid-based simulations or spatial discretizations, a mesh of size $N \times N$ scales cubically to $N \times N \times N$ when a third dimension is added. This polynomial increase in degrees of freedom quickly cripples traditional solvers.
For example, Yang et al. (2023) demonstrate the efficacy of finite element methods (FEM) for calculating the band structures of 2D photonic crystals. However, the authors note that extending these numerical methods to 3D introduces massive computational complexity because the volume of light-matter interactions expands exponentially.
This scaling wall also impacts optimization and path planning. The generative design framework developed by Starodubcev (2022) optimizes 2D physical objects efficiently by combining evolutionary algorithms with physics simulators, yet scaling this to 3D structural optimization multiplies the design variables beyond the reach of simple search algorithms.
In autonomous navigation, Ma et al. (2016) simplify marine path planning by treating the ocean surface as a 2D grid. However, real-world oceanographic phenomena are inherently 3D; accounting for depth and vertical water column dynamics introduces a level of complexity that standard planning frameworks cannot easily absorb. Even in discrete systems like programmable matter, where Almethen (2019) shows that parallel "line-pushing" mechanisms can drastically accelerate 2D shape transformations, translating these movement mechanisms and complexity bounds to 3D spatial grids remains an open algorithmic challenge.
The Geometric and Topological Shift
In a 2D plane, boundaries are lines and interfaces are curves. In a 3D volume, boundaries become complex surfaces with out-of-plane degrees of freedom, fundamentally altering the geometry of the problem.
Consider shape representation and contact mechanics. In self-supervised learning, Foix Romero et al. (2025) developed ShapeEmbed to quantify object outlines using 2D distance matrices. Extending this framework to 3D requires handling complex surface topologies, self-occlusions, and non-trivial surface metrics that do not exist in flat outlines.
In mechanical modeling, Zhao et al. (2022) introduced a smooth barrier method for frictional contact on embedded interfaces. While highly effective for planar problems, the authors point out that adapting these energy potentials for 3D spatial domains requires entirely new integration schemes to handle complex surface-to-surface interactions.
Fluid interfaces face a similar hurdle. Lu et al. (2021) successfully modeled three-phase flows on curved 2D surfaces using a characteristic line model for wetting. However, simulating realistic porous media requires migrating this boundary condition to fully 3D pore structures, where fluid pathways twist, split, and merge in ways that a 2D plane cannot capture.
2D Interface: A simple continuous curve separating two phases.
[ Phase A ] -------- (1D Curve) -------- [ Phase B ]
3D Interface: A complex 2D manifold with local curvature, twisting, and saddle points.
Moving to 3D introduces out-of-plane degrees of freedom, requiring surface tensors rather than scalar curves.
The Emergence of Inherently 3D Physics
Approximating a system in 2D often means omitting critical physics. Some physical phenomena are mathematically impossible in two dimensions because they require three independent spatial axes to manifest.
In material mechanics, Shenoy et al. (1999) developed the Quasicontinuum method to bridge atomistic and continuum scales. They noted, however, that 2D models cannot simulate phenomena like dislocation junction formation or complex crack propagation, both of which rely on out-of-plane atomic movements and twisting forces.
The discrepancy is even more pronounced in fluid dynamics, where turbulence behaves fundamentally differently depending on dimensionality. While He et al. (2018) can map out entropy generation rates in 2D Rayleigh-Taylor turbulence, real-world 3D turbulence involves vortex stretching—a mechanism where fluid vortices are elongated by velocity gradients, amplifying vorticity and driving energy down to microscopic scales. This mechanism is completely absent in 2D. Consequently, reduced-order models like those developed by Rozza (2019) and Rozza (2017) for turbulent flows must be fundamentally re-engineered when transitioning to 3D to ensure they can stabilize and capture these volumetric energy cascades.
The Theoretical and Mathematical Wall
Perhaps the most daunting aspect of the 2D-to-3D gap is that mathematical proofs of stability, uniqueness, and convergence frequently collapse when moving to higher dimensions.
The most famous manifestation of this wall occurs in the Navier-Stokes equations. Blömker et al. (2013) proved the accuracy and stability of the continuous-time 3DVAR filter for 2D fluid dynamics. However, the authors note that extending this proof to 3D is blocked by one of the greatest unsolved problems in mathematics: proving the existence and uniqueness of smooth, long-term solutions for the 3D Navier-Stokes equations. Without this guarantee, the filter's stability cannot be mathematically certified.
Inverse problems suffer from a similar dimensional barrier. Bukhgeĭm et al. (2000) established conditional stability for identifying non-smooth material boundaries using electrostatic measurements, and Cheng et al. (2005) proved uniqueness in inverse scattering obstacles. However, both frameworks are strictly confined to 2D domains. Moving these proofs to 3D requires addressing higher-dimensional geometric complexities, where wave scattering and boundary integrals can no longer rely on simple planar curves and conformal mappings.
Ultimately, bridging the 2D-to-3D gap is not a matter of scaling up hardware or waiting for faster computers. It requires designing algorithms, neural networks, and mathematical frameworks that inherently understand the topological, geometric, and physical realities of three-dimensional space.