Interface Neural Networks

• Interface neural networks are systems built around explicit interfaces that capture discontinuities, dynamic boundaries, or user interactions in various domains.
• They employ specialized architectures to enforce conditions in PDEs, rapidly adapt in brain–computer interfaces, and structure human–model knowledge exchange.
• Their design strategically addresses challenges like loss balancing, adaptive sampling, and control complexity to improve performance across multi-domain applications.

Interface neural networks are neural systems organized around an interface that must be represented, traversed, or controlled. In the current literature, the term is used in several technically distinct senses. A substantial part of the cited work concerns PDE interface problems, where a network approximates piecewise solutions across an internal interface $\Gamma$ and enforces jump or flux conditions. Other uses treat the network as the learned decoder at the brain–computer interface, as the computational core of a deformable touch interface, or as the bidirectional medium through which humans and neural models exchange structured knowledge (He et al., 2020, Roy et al., 2024, Wu et al., 2022, Sharma et al., 2024, He et al., 2024).

1. Terminological scope

Across the cited literature, “interface neural networks” does not denote a single architecture class. Rather, it names neural constructions whose defining object is an interface: a material interface, a moving phase boundary, a user-facing sensing surface, a human–model knowledge interface, or a scheduling interface between modules and environments.

Context	Interface object	Representative formulation
Elliptic interface PDEs	Internal interface $\Gamma$ between subdomains	Piecewise subnetworks, jump losses, flux continuity (He et al., 2020)
Moving-interface and multiphase problems	Evolving $\Gamma(t)$, level set $\phi$, or VOF field $\alpha$	Level-set PINNs, EVT-based XI-PINN, PINN-LBM coupling (Mullins et al., 4 Feb 2025)
EEG-based BCI	Decoder between EEG time series and task labels	Fast-adapting CNN classifier with $A(s)$ over fine-tuning steps (Sharma et al., 2024)
Human–NN knowledge exchange	Structured concept graphs used as a common “language”	C-SCG, I-SCG, GRN, distillation (He et al., 2024)
Soft and modular interfaces	Deformable touch surface or memory-mediated module interface	OrbTouch CNN; CU–PU–Node–Environment orchestration (Larson et al., 2017)

This breadth is not merely terminological. It indicates that the interface itself is often the source of nonstationarity, discontinuity, sparse supervision, or control complexity. In PDE settings the interface induces jump conditions and coefficient discontinuities. In BCI it induces cross-subject and cross-task domain shift. In human-facing systems it creates a requirement for interpretable mediation rather than one-way prediction.

2. Interface problems for PDEs: representations and architectures

The canonical mathematical setting is the elliptic interface problem on $\Omega=\Omega_1\cup\Omega_2$, with interface $\Gamma=\partial\Omega_1\cap\partial\Omega_2$, piecewise coefficient $\beta(x)$ or $\alpha(x)$, and conditions such as

$\Gamma$0

together with

$\Gamma$1

or, in alternative notation,

$\Gamma$2

The central design question is how to represent the piecewise field while preserving sufficient flexibility near $\Gamma$3.

A direct construction is the piecewise mesh-free least-squares network. In this formulation, the approximation is

$\Gamma$4

with routing by an indicator derived from an analytic interface or a level-set function. Training minimizes

$\Gamma$5

where interior residuals, interface jump terms, and boundary terms are all sampled pointwise, with $\Gamma$6 and $\Gamma$7 in the reported experiments (He et al., 2020).

A different parameterization is used by Adaptive Interface-PINNs. AdaI-PINNs retain a shared parameter set $\Gamma$8 across all subdomains, but allow a trainable activation slope $\Gamma$9 for each subdomain:

$\Gamma(t)$0

This removes the manual activation assignment required by I-PINNs while preserving domain decomposition. On one-, two-, and three-dimensional elliptic benchmarks, AdaI-PINNs reduce computational cost by $\Gamma(t)$1–$\Gamma(t)$2 times while producing similar or better accuracy (Roy et al., 2024).

AE-PINNs make the interface explicit by decomposing the solution into a continuous component and a discontinuous component,

$\Gamma(t)$3

with a fully connected network for $\Gamma(t)$4 and an interface-attention neural network in each subdomain for $\Gamma(t)$5. The IA-NN injects level-set information through a transmitter

$\Gamma(t)$6

and updates hidden states by

$\Gamma(t)$7

This architecture is designed to focus representation capacity near $\Gamma(t)$8, and the reported experiments show higher accuracy than PINNs, I-PINNs, and M-PINNs across 1D, 2D, and 3D benchmarks (Zheng et al., 23 Jun 2025).

Dual KANs replace MLP subnetworks by Kolmogorov–Arnold Networks, one per subdomain, with learnable univariate edge functions

$\Gamma(t)$9

In the reported elliptic interface problems, dual KANs use much smaller network widths than dual MLP PINNs while achieving higher accuracy and faster convergence (Xin et al., 2 Feb 2026).

A further branch dispenses with nonlinear training almost entirely. Local Randomized Neural Networks assign one randomized network to each subdomain,

$\phi$0

and determine only the linear coefficients by a single weighted least-squares solve. Boundary and interface conditions couple the local approximations through one global linear system, yielding a mesh-free, optimization-free solver for elliptic and parabolic interface problems (Li et al., 2023).

3. Enforcing interface conditions: penalties, hard constraints, adaptation, and theory

The major algorithmic divide is between soft and hard interface enforcement. In soft-constrained PINNs, the total loss combines PDE residuals, interface conditions, and boundary conditions,

$\phi$1

with interface terms typically penalizing both $\phi$2 and $\phi$3 at sampled points. This baseline is theoretically sharpened in the gradient-enhanced, Lipschitz-regularized analysis of interface PINNs, where tangential derivatives on $\phi$4 and $\phi$5 are added to the loss and the resulting minimizing network sequence is shown to converge to the unique solution in $\phi$6 as the number of samples increases (Wu et al., 2022).

METALIC treats interface enforcement itself as a meta-learning problem. It considers nine candidate interface penalties, including solution continuity, average solution continuity, residual penalties, residual continuity, gradient-enhanced residual, flux continuity, and first- and second-order derivative continuity. A contextual multi-armed bandit with Gaussian-process rewards selects which interface conditions to activate. On held-out PDEs, the reported offline errors are markedly below random baselines; for example, on Poisson problems, METALIC-Seq-UCB achieves $\phi$7 versus $\phi$8 for Random-Single (Li et al., 2022).

Hard-constrained methods encode the interface in the ansatz rather than in a penalty. One formulation is the windowing approach, which multiplies subnetworks by compactly supported window functions so that continuity and flux balance are satisfied by design. Another is the buffer approach, which adds auxiliary correction functions and solves small linear systems for their coefficients at each iteration. In one-dimensional problems, the windowing approach attains errors as low as $\phi$9 on simple structured cases, while the buffer approach remains around $\alpha$0 across a wider range of source terms and interface configurations; in two dimensions, the buffer formulation is reported as more robust (Chung et al., 9 Apr 2026).

A related finite-element-style strategy is enforced interface constraints for discrete PINNs. EIC-dPINN partitions the mesh into subdomains and enforces displacement continuity by projection,

$\alpha$1

after node–element pairing and inverse isoparametric mapping. This removes interface penalty terms, supports independent meshing, and yields multi-GPU speedups up to $\alpha$2 on the large problem while preserving identical error curves relative to single-GPU training (Yin et al., 16 May 2025).

Adaptive point placement is another recurring mechanism. In the deep least-squares method, candidate points are ranked by residual indicators and the top $\alpha$3 are added at each refinement level, which is especially effective when errors concentrate near $\alpha$4 under high-contrast coefficients (He et al., 2020). In PINN-based KANs, RAR-D uses

$\alpha$5

with $\alpha$6 and example-dependent $\alpha$7 to resample collocation points toward high-residual regions, reducing errors and making them more uniform (Xin et al., 2 Feb 2026).

Time-decomposed XPINNs supply yet another interface mechanism. The interface line between adjacent time slabs is replaced by a finite interface zone

$\alpha$8

with pseudo-initial, residual-matching, and temporal gradient-matching losses. In the bright two-soliton example, the interface-zone formulation reduces whole-domain relative $\alpha$9 errors from $A(s)$0 to $A(s)$1 for $A(s)$2 and from $A(s)$3 to $A(s)$4 for $A(s)$5 (Chen et al., 2023).

4. Moving interfaces, level sets, and multiphase flows

When the interface itself evolves, the network must represent both the field and the geometry of $A(s)$6. One route is level-set transport,

$A(s)$7

with $A(s)$8. In moving-interface PINNs based on PirateNet, Fourier feature embeddings, random weight factorization, adaptive residual blocks, causal training, and sequence-to-sequence learning are combined to solve level-set advection without upwind stabilization, geometric reinitialization, or mass conservation schemes. The reported relative $A(s)$9 errors are $\Omega=\Omega_1\cup\Omega_2$0 for Zalesak’s disk and $\Omega=\Omega_1\cup\Omega_2$1 for the time-reversed vortex flow problem (Mullins et al., 4 Feb 2025).

XI-PINN extends interface PINNs to parabolic moving-interface problems by introducing an extended variable. A single network approximates

$\Omega=\Omega_1\cup\Omega_2$2

with

$\Omega=\Omega_1\cup\Omega_2$3

The level set may be given analytically or learned through an inverse flow map

$\Omega=\Omega_1\cup\Omega_2$4

The paper derives approximation, statistical, and optimization error terms, gives an NTK-based convergence comparison to conventional PINNs, and reports, for a rotating circular interface example, $\Omega=\Omega_1\cup\Omega_2$5 and $\Omega=\Omega_1\cup\Omega_2$6 (Bi et al., 2 Aug 2025).

For multiphase lattice Boltzmann simulations, the interface can be preserved by sequential correction rather than by direct solution. In PINN-LBM coupling, the PINN outputs $\Omega=\Omega_1\cup\Omega_2$7, is trained on data consistency, physics residuals, and an interface-sharpness term, and then updates the LBM state primarily in an interface band identified by $\Omega=\Omega_1\cup\Omega_2$8 thresholds. In the reported droplet test, the effective interface width decreases from $\Omega=\Omega_1\cup\Omega_2$9 for pure LBM to $\Gamma=\partial\Omega_1\cap\partial\Omega_2$0 for PINN-LBM, and the interface energy drops from $\Gamma=\partial\Omega_1\cap\partial\Omega_2$1 to $\Gamma=\partial\Omega_1\cap\partial\Omega_2$2 (Li et al., 13 Apr 2025).

A related inverse formulation infers incompressible two-phase flow fields from interface motion using a VOF representation. The auxiliary field is the volume fraction $\Gamma=\partial\Omega_1\cap\partial\Omega_2$3, with

$\Gamma=\partial\Omega_1\cap\partial\Omega_2$4

and curvature

$\Gamma=\partial\Omega_1\cap\partial\Omega_2$5

The paper emphasizes that appropriate weighting of losses associated with the residual of the partial differential equations is crucial for successful training, especially when inferring velocity and pressure solely from scattered-time interface position data (Buhendwa et al., 2021).

5. Operational and human-facing interface neural networks

Outside PDEs, interface neural networks often literally sit at the boundary between signal and action. In EEG-based BCI, the interface neural network is the learned decoder between brain signals and machine commands. Fast adaptability is defined by the adaptation curve $\Gamma=\partial\Omega_1\cap\partial\Omega_2$6 measured over fine-tuning steps $\Gamma=\partial\Omega_1\cap\partial\Omega_2$7 using only $\Gamma=\partial\Omega_1\cap\partial\Omega_2$8 labeled samples per class and at most $\Gamma=\partial\Omega_1\cap\partial\Omega_2$9 gradient steps. On the PhysioNet EEG Motor Movement/Imagery dataset, a compact EEGNet-style CNN with layer normalization and transfer learning adapts faster than the same CNN with batch normalization or a MAML-based pipeline, with the following before/after results across held-out subjects (Sharma et al., 2024).

Activity	Before adaptation	After adaptation
A1	$\beta(x)$0	$\beta(x)$1
A2	$\beta(x)$2	$\beta(x)$3
A3	$\beta(x)$4	$\beta(x)$5
A4	$\beta(x)$6	$\beta(x)$7

In soft human–computer interaction, OrbTouch integrates a deformable dielectric elastomer sensing skin with a spatiotemporal convolutional neural network that maps local deformations to gesture labels and touch locations. The sensing relation is modeled by

$\beta(x)$8

and the system is reported as responsive enough for real-time interaction, including control of Tetris (Larson et al., 2017).

A different sense of interface appears in two-way human–NN knowledge interaction. Here the interface is a structured graph language: the Class-Specific Structural Concepts Graph (C-SCG) and the image-level I-SCG. A graph reasoning network is trained to simulate the original classifier through

$\beta(x)$9

with

$\alpha(x)$0

Humans then edit concepts and edges in the C-SCGs, and the corrected logic is distilled back into the original network through a two-teacher loss

$\alpha(x)$1

On the three modified classes, the reported test accuracies improve from $\alpha(x)$2 to $\alpha(x)$3, from $\alpha(x)$4 to $\alpha(x)$5, and from $\alpha(x)$6 to $\alpha(x)$7, while the average accuracy on other classes changes from $\alpha(x)$8 to $\alpha(x)$9 (He et al., 2024).

6. Modular interaction interfaces and common design pressures

A more abstract use of the term appears in Interaction Networks, where the interface is the controller layer linking modules, shared memory, and environments. An Interaction Network consists of conventional neural modules, Nodes as memory locations, Environments as IO interfaces, and a DQN-like Control Unit that decides which Processing Unit to execute and when to interact with the environment. The Control Unit estimates

$\Gamma$00

and the paper states that any existing type of neural network can be reproduced in an Interaction Network in its entirety, with only a constant computational overhead. The same paper also reports that these systems are very hard to train because of double-optimization interference, long-horizon credit assignment, and unstable reward propagation (Dietz, 2020).

Taken together, these literatures show several recurring design pressures. Known interface geometry is often required, whether as an analytic $\Gamma$01, a level-set function $\Gamma$02, or a node–element pairing across meshes (Roy et al., 2024, Bi et al., 2 Aug 2025, Yin et al., 16 May 2025). Soft formulations remain sensitive to loss balancing, particularly when interface residuals compete with bulk physics or sparse data (Wu et al., 2022, Buhendwa et al., 2021). Hard constraints, adaptive sampling, and explicit interface variables reduce that sensitivity but add architectural or preprocessing complexity (Chung et al., 9 Apr 2026, Xin et al., 2 Feb 2026). Human-facing systems face a different but analogous constraint: the interface must be both operationally effective and interpretable enough to support calibration, editing, or direct interaction (Sharma et al., 2024, He et al., 2024).

This suggests a unifying characterization. Interface neural networks are not defined by a single layer type or training algorithm, but by a design principle: the network is built so that the interface itself becomes a first-class computational object. In PDEs this object is a geometric discontinuity or moving boundary. In BCI it is a rapidly calibrating decoder under severe domain shift. In human–model systems it is a structured communication channel. In modular control it is the policy that selects which computation to execute next. Across all of these settings, performance depends on whether the interface is merely penalized after the fact or explicitly represented in the model’s state, architecture, and training protocol.