arxiv: 2604.08076 · v1 · submitted 2026-04-09 · 💻 cs.CE · math.AP

Recognition: no theorem link

φ-DeepONet: A Discontinuity Capturing Neural Operator

Sumanta Roy , Stephen T. Castonguay , Pratanu Roy , Michael D. Shields

Authors on Pith no claims yet

Pith reviewed 2026-05-10 17:54 UTC · model grok-4.3

classification 💻 cs.CE math.AP

keywords discontinuitiesdeeponetneuraloperatorbranchcombinedembeddingfields

0 comments

The pith

φ-DeepONet learns mappings with discontinuities in inputs and outputs by combining multiple branch networks with a nonlinear interface embedding in the trunk, trained via physics- and interface-informed loss, and shows accurate results on 1D/2D benchmarks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Neural operators learn to map one function to another, for example turning a material property field into a stress field. Most versions assume everything is smooth and continuous. Real problems often have jumps at material boundaries or shocks. φ-DeepONet splits the problem. Discontinuous inputs are fed to several separate branch networks. For the output, the domain is split into regions marked by a simple one-hot code; this code is fed together with position into a modified trunk network that learns a nonlinear embedding of the interface location. The branch and trunk outputs are multiplied to give the final field. Training adds both the governing physics equations and explicit interface conditions to the loss. The authors test the method on simple one- and two-dimensional problems that contain strong jumps and report that the predictions remain accurate and stable.

Core claim

We evaluate φ-DeepONet on several one- and two-dimensional benchmark problems and demonstrate that it delivers accurate and stable predictions even in the presence of strong interface-driven discontinuities.

Load-bearing premise

That a nonlinear latent embedding built from a one-hot representation of domain decomposition, when combined with spatial coordinates in the trunk, is sufficient to capture and propagate strong and weak discontinuities in the output fields.

Figures

Figures reproduced from arXiv: 2604.08076 by Michael D. Shields, Pratanu Roy, Stephen T. Castonguay, Sumanta Roy.

**Figure 2.** Figure 2: Schematic of the architecture of ϕ−DeepONet for approximating the operator G : u 7→ s. where N , B, and I denote the PDE, boundary, and interface condition operators, respectively. Here, N is the number of input function realizations, m is the number of collocation points used to evaluate the PDE residual, and b and t are the numbers of points used to enforce the boundary and interface conditions. Hard enf… view at source ↗

**Figure 3.** Figure 3: Convergence profiles (physics loss or MSE vs epochs) for the various [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

**Figure 4.** Figure 4: 1D single-interface problem: performance of the [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

**Figure 5.** Figure 5: Variation of the test L2 errors with varying size of the input dataset (Ntrain.) 0.0 0.2 0.4 0.6 0.8 1.0 y 0.0 0.1 0.2 0.3 0.4 0.5 s(y) µ = 1.0, `s = 0.2 0.0 0.5 1.0 1 2 u(x) (a) 0.0 0.2 0.4 0.6 0.8 1.0 y 0.00 0.05 0.10 0.15 0.20 0.25 s(y) µ = 1.0, `s = 0.15 0.0 0.5 1.0 0.5 1.0 u(x) (b) 0.0 0.2 0.4 0.6 0.8 1.0 y 0.00 0.05 0.10 0.15 0.20 0.25 0.30 s(y) µ = 1.0, `s = 0.1 0.0 0.5 1.0 0 2 u(x) (c) 0.0 0.2 0.4 … view at source ↗

**Figure 6.** Figure 6: Prediction of the ϕ-DeepONet model on out-of-distribution test samples for the 1D problem with one interface (Section 4.1). The model is trained on input functions sampled from a GP with mean µ = 1.0 and length scale ls = 0.2 (Eqs. (15) and (16)), and tested on several other combinations of µ and ls. Across the rows, µ is varied, and across the columns, ls is varied. The corresponding input functions are a… view at source ↗

**Figure 7.** Figure 7: 1D multiple-interface problem: performance of the [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

**Figure 8.** Figure 8: Prediction of the ϕ-DeepONet model on out-of-distribution test samples for the 1D problem with four interfaces (Section 4.2). The model is trained on input functions sampled from a GP with mean µ = 1.0 and length scale ls = 0.2 (Eqs. (15) and (16)), and tested on several other combinations of µ and ls. Across the rows, µ is varied, and across the columns, ls is varied. The corresponding input functions fed… view at source ↗

**Figure 9.** Figure 9: Performance of the ϕ-DeepONet (non-linear CE, D = 3) framework on two random test samples for the 2D problem (subfigures (a) and (b)), along with the distribution of the test errors (sub-figure (c)). 15 [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗

**Figure 10.** Figure 10: 1D multiple-interface problem with discontinuous [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗

**Figure 11.** Figure 11: 1D problem with discontinuous inputs u(x) and outputs s(y): performance of the ϕ-DeepONet framework on two random test samples (subfigures (a) and (b)), along with the distribution of the test errors (subfigure (c)). for which the problem admits the known analytical solution: s(y) = ( y 2 1 + y 2 2 , y ∈ Ω1, 0.1(y 2 1 + y 2 2 ) 2 − 0.01 log 2 p y 2 1 + y 2 2 , y ∈ Ω2. (23) The operator G is trained on… view at source ↗

**Figure 12.** Figure 12: 2D petal-shaped interface: the predicition of the [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗

**Figure 13.** Figure 13: 2D petal-shaped interface: An ablation study showing the performance of various variants of [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗

read the original abstract

We present $\phi-$DeepONet, a physics-informed neural operator designed to learn mappings between function spaces that may contain discontinuities or exhibit non-smooth behavior. Classical neural operators are based on the universal approximation theorem which assumes that both the operator and the functions it acts on are continuous. However, many scientific and engineering problems involve naturally discontinuous input fields as well as strong and weak discontinuities in the output fields caused by material interfaces. In $\phi$-DeepONet, discontinuities in the input are handled using multiple branch networks, while discontinuities in the output are learned through a nonlinear latent embedding of the interface. This embedding is constructed from a {\it one-hot} representation of the domain decomposition that is combined with the spatial coordinates in a modified trunk network. The outputs of the branch and trunk networks are then combined through a dot product to produce the final solution, which is trained using a physics- and interface-informed loss function. We evaluate $\phi$-DeepONet on several one- and two-dimensional benchmark problems and demonstrate that it delivers accurate and stable predictions even in the presence of strong interface-driven discontinuities.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

φ-DeepONet adds multiple branches plus a one-hot-derived trunk embedding to handle input and output discontinuities in DeepONet, but it requires known domain decomposition upfront and the abstract supplies almost no quantitative evidence.

read the letter

The core idea is a targeted tweak to DeepONet: multiple branch networks to accept discontinuous input functions, and a trunk that concatenates spatial coordinates with a nonlinear embedding built from a one-hot encoding of the domain decomposition. This is then dotted with the branch outputs and trained with a physics- and interface-informed loss. The claim is that this produces stable predictions on problems with strong interface-driven jumps in one and two dimensions. That architectural choice is new relative to the standard DeepONet literature referenced in the abstract, and it directly attacks a practical limitation—most neural operators struggle when material interfaces or shocks produce non-smooth fields. The one-hot embedding is a simple way to inject subdomain information into the trunk, and when the decomposition is supplied it could plausibly let the network learn the jump conditions without explicit enforcement at every point. The benchmarks mentioned are the right test cases for this kind of work. The main limitation is exactly the one the stress-test note flags: the method presupposes that the interfaces (and therefore the one-hot labels) are known exactly at query time. Nothing in the abstract describes a mechanism to discover or tolerate uncertainty in those locations, so the operator is not yet general for cases where discontinuities must be inferred from data alone. The physics-informed loss cannot compensate for missing or incorrect subdomain information. In addition, the abstract asserts “accurate and stable predictions” without reporting any error norms, baseline comparisons, or details on how the loss balances the PDE residual against interface conditions. That leaves the central performance claim weakly supported. A reader working on operator learning for engineering PDEs with known interfaces would find the architectural pattern useful to try. The paper is coherent on its own terms and engages the relevant literature, so it is worth sending to referees even though the current evidence is thin and the generality is narrower than the title suggests.

Referee Report

2 major / 1 minor

Summary. The paper introduces φ-DeepONet, a physics-informed neural operator extending DeepONet to handle discontinuities. Input discontinuities are addressed via multiple branch networks, while output discontinuities are captured through a nonlinear latent embedding constructed from a one-hot representation of the domain decomposition, which is concatenated with spatial coordinates in a modified trunk network. Branch and trunk outputs are combined via dot product, and the model is trained with a physics- and interface-informed loss. The authors evaluate the method on several 1D and 2D benchmark problems and claim it delivers accurate and stable predictions even with strong interface-driven discontinuities.

Significance. If the quantitative claims hold, the work would address a known limitation of standard neural operators (which rely on continuity assumptions) and could be useful for interface problems in computational mechanics and engineering. The explicit incorporation of domain decomposition into the trunk and the physics-informed loss represent a concrete architectural attempt to capture discontinuities without post-processing. However, the approach's dependence on a priori known interfaces reduces its generality relative to fully data-driven alternatives.

major comments (2)

[Abstract / Method] Abstract and method description: the architecture presupposes that the domain decomposition (and thus the one-hot interface labels) is known exactly when evaluating the operator. This is a load-bearing assumption for the central claim of a general 'discontinuity capturing neural operator,' yet the manuscript provides no mechanism for learning or tolerating uncertainty in interface locations. Benchmarks may succeed only because the decomposition is supplied as part of the problem setup; without this, the nonlinear latent embedding cannot be constructed.
[Abstract] Abstract: the claim that 'benchmark tests produced accurate stable results' is stated without any quantitative metrics, baseline comparisons (e.g., standard DeepONet or PINN variants), error bars, or details on how the physics- and interface-informed loss is constructed (e.g., weighting of interface terms or enforcement of jump conditions). This leaves the performance assertions weakly supported and prevents assessment of whether the method improves upon existing approaches.

minor comments (1)

[Abstract] Notation: the symbol φ is used in the title and abstract but its precise definition (as the nonlinear embedding or the full operator) is not clarified early in the text, which may confuse readers familiar with standard DeepONet notation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and have prepared revisions to clarify assumptions and strengthen the presentation of results.

read point-by-point responses

Referee: [Abstract / Method] Abstract and method description: the architecture presupposes that the domain decomposition (and thus the one-hot interface labels) is known exactly when evaluating the operator. This is a load-bearing assumption for the central claim of a general 'discontinuity capturing neural operator,' yet the manuscript provides no mechanism for learning or tolerating uncertainty in interface locations. Benchmarks may succeed only because the decomposition is supplied as part of the problem setup; without this, the nonlinear latent embedding cannot be constructed.

Authors: We acknowledge that φ-DeepONet requires a priori knowledge of the domain decomposition to construct the one-hot representation and nonlinear latent embedding. This design choice targets interface problems in which material or geometric interfaces are known exactly, as is standard in many computational mechanics settings. The method does not incorporate mechanisms to infer or tolerate uncertainty in interface locations, which would necessitate a separate interface-detection component. We will revise the abstract and Section 2 to explicitly state this assumption and reframe the contribution as a discontinuity-capturing operator for known-interface problems rather than a fully general unknown-discontinuity solver. This clarification does not alter the reported numerical results but improves the scope statement. revision: yes
Referee: [Abstract] Abstract: the claim that 'benchmark tests produced accurate stable results' is stated without any quantitative metrics, baseline comparisons (e.g., standard DeepONet or PINN variants), error bars, or details on how the physics- and interface-informed loss is constructed (e.g., weighting of interface terms or enforcement of jump conditions). This leaves the performance assertions weakly supported and prevents assessment of whether the method improves upon existing approaches.

Authors: The abstract is intentionally concise, but the full manuscript (Section 4) reports quantitative L2 errors, direct comparisons against standard DeepONet on the same benchmarks, error bars obtained from multiple independent runs, and the explicit form of the physics- and interface-informed loss (including weighting coefficients for interface residual terms and enforcement of jump conditions). To address the concern, we will augment the abstract with representative quantitative metrics and a brief reference to the loss construction. These additions will be drawn directly from the existing numerical results without introducing new experiments. revision: yes

Circularity Check

0 steps flagged

No significant circularity; architectural claims rest on explicit design and benchmarks.

full rationale

The paper presents φ-DeepONet as an architectural extension of DeepONet that uses multiple branch networks for discontinuous inputs and a nonlinear latent embedding constructed from a one-hot domain decomposition concatenated into a modified trunk network, with training via a physics- and interface-informed loss. No equations or results in the provided description reduce the claimed predictions or performance to quantities defined by the same fitted parameters, self-citations, or ansatzes imported from prior author work. The central claims are supported by direct evaluation on one- and two-dimensional benchmark problems rather than any derivation that loops back to its inputs by construction. This is the normal case of a self-contained methodological contribution.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

Only the abstract is available, so the ledger is necessarily high-level. The method introduces a new latent embedding entity and relies on the assumption that the modified architecture can approximate discontinuous operators.

free parameters (1)

network architecture hyperparameters
Branch and trunk network widths, depths, and activation choices are chosen to fit the discontinuous benchmarks but are not enumerated in the abstract.

axioms (1)

domain assumption The modified trunk with one-hot interface embedding can represent strong and weak discontinuities when combined with the branch outputs via dot product.
Invoked in the description of how output discontinuities are learned.

invented entities (1)

nonlinear latent embedding of the interface no independent evidence
purpose: To encode and propagate interface locations into the trunk network output for discontinuous solution fields.
New component constructed from one-hot domain decomposition and spatial coordinates; no independent evidence outside the paper is provided.

pith-pipeline@v0.9.0 · 5504 in / 1378 out tokens · 31843 ms · 2026-05-10T17:54:01.493095+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Shock-Centered Low-Rank Structure and Neural-Operator Representation of Rarefied Micro-Nozzle Flows
physics.flu-dyn 2026-05 unverdicted novelty 7.0

Shock-centered scaling of DSMC fields in micro-nozzles reveals low-rank density structure, enabling DeepONet surrogates with mean errors reduced to 4.51% on hardest test cases.

Reference graph

Works this paper leans on

46 extracted references · 9 canonical work pages · cited by 1 Pith paper · 3 internal anchors

[1]

Heat conduction in layered, composite materials.Journal of applied physics, 57(5):1569–1573, 1985

James Baker-Jarvis and Ramarao Inguva. Heat conduction in layered, composite materials.Journal of applied physics, 57(5):1569–1573, 1985

1985
[2]

Heat transfer analysis in multi-layered materials with interfacial thermal resistance.Composite Structures, 293:115728, 2022

Wei-bin Yuan, Nanting Yu, Long-yuan Li, and Yuan Fang. Heat transfer analysis in multi-layered materials with interfacial thermal resistance.Composite Structures, 293:115728, 2022

2022
[3]

Evaluation of the moisture effect on the material interface using multiscale modeling.Multiscale Science and Engineering, 1(2):108–118, 2019

Renyuan Qin and Denvid Lau. Evaluation of the moisture effect on the material interface using multiscale modeling.Multiscale Science and Engineering, 1(2):108–118, 2019

2019
[4]

Multi-material modeling of sorption-desorption processes with experimental validation.Chemical Engineering Science, 253:117542, 2022

Pratanu Roy, Stephen T Castonguay, Jennifer M Knipe, Yunwei Sun, Elizabeth A Glascoe, and Hom N Sharma. Multi-material modeling of sorption-desorption processes with experimental validation.Chemical Engineering Science, 253:117542, 2022

2022
[5]

Robert J Fisher and Robert A Peattie. Controlling tissue microenvironments: biomimetics, transport phenomena, and reacting systems.Tissue Engineering II: Basics of Tissue Engineering and Tissue Applications, pages 1–73, 2006

2006
[6]

Modeling fluid flow in fractured porous media with the interfacial conditions between porous medium and fracture.Transport in Porous Media, 139(1):109–129, 2021

N Hosseini and AR Khoei. Modeling fluid flow in fractured porous media with the interfacial conditions between porous medium and fracture.Transport in Porous Media, 139(1):109–129, 2021

2021
[7]

A finite element method for interface problems in domains with smooth boundaries and interfaces.Advances in Computational Mathematics, 6(1):109–138, 1996

James H Bramble and J Thomas King. A finite element method for interface problems in domains with smooth boundaries and interfaces.Advances in Computational Mathematics, 6(1):109–138, 1996

1996
[8]

A locally modified parametric finite element method for interface problems

Stefan Frei and Thomas Richter. A locally modified parametric finite element method for interface problems. SIAM Journal on Numerical Analysis, 52(5):2315–2334, 2014

2014
[9]

Grid modification during simulated fracture propagation, March 21 2023

Dakshina M Valiveti, Chandrasekhar A Srinivas, and Vadim Dyadechko. Grid modification during simulated fracture propagation, March 21 2023. US Patent 11,608,730

2023
[10]

An efficient finite element method for embedded interface problems.International journal for numerical methods in engineering, 78(2):229–252, 2009

John Dolbow and Isaac Harari. An efficient finite element method for embedded interface problems.International journal for numerical methods in engineering, 78(2):229–252, 2009

2009
[11]

Physics- informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics- informed machine learning.Nature Reviews Physics, 3(6):422–440, 2021

2021
[12]

Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational physics, 378:686–707, 2019

2019
[13]

Multilayer feedforward networks are universal approxi- mators.Neural networks, 2(5):359–366, 1989

Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approxi- mators.Neural networks, 2(5):359–366, 1989

1989
[14]

Physics-informed neural networks (pinns) for fluid mechanics: A review.Acta Mechanica Sinica, 37(12):1727–1738, 2021

Shengze Cai, Zhiping Mao, Zhicheng Wang, Minglang Yin, and George Em Karniadakis. Physics-informed neural networks (pinns) for fluid mechanics: A review.Acta Mechanica Sinica, 37(12):1727–1738, 2021

2021
[15]

Scientific machine learning through physics–informed neural networks: Where we are and what’s next

Salvatore Cuomo, Vincenzo Schiano Di Cola, Fabio Giampaolo, Gianluigi Rozza, Maziar Raissi, and Francesco Piccialli. Scientific machine learning through physics–informed neural networks: Where we are and what’s next. Journal of Scientific Computing, 92(3):88, 2022

2022
[16]

Understanding physics-informed neural networks: Tech- niques, applications, trends, and challenges.AI, 5(3):1534–1557, 2024

Amer Farea, Olli Yli-Harja, and Frank Emmert-Streib. Understanding physics-informed neural networks: Tech- niques, applications, trends, and challenges.AI, 5(3):1534–1557, 2024

2024
[17]

Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

2023
[18]

Neural operators for accelerating scientific simulations and design.Nature Reviews Physics, 6(5):320–328, 2024

Kamyar Azizzadenesheli, Nikola Kovachki, Zongyi Li, Miguel Liu-Schiaffini, Jean Kossaifi, and Anima Anand- kumar. Neural operators for accelerating scientific simulations and design.Nature Reviews Physics, 6(5):320–328, 2024. 22 APREPRINT- APRIL10, 2026

2024
[19]

Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nature machine intelligence, 3(3): 218–229, 2021

Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nature machine intelligence, 3(3): 218–229, 2021

2021
[20]

Tianping Chen and Hong Chen. Universal approximation to nonlinear operators by neural networks with arbitrary activation functions and its application to dynamical systems.IEEE transactions on neural networks, 6(4): 911–917, 1995

1995
[21]

A comprehensive and fair comparison of two neural operators (with practical extensions) based on fair data.Computer Methods in Applied Mechanics and Engineering, 393:114778, 2022

Lu Lu, Xuhui Meng, Shengze Cai, Zhiping Mao, Somdatta Goswami, Zhongqiang Zhang, and George Em Karniadakis. A comprehensive and fair comparison of two neural operators (with practical extensions) based on fair data.Computer Methods in Applied Mechanics and Engineering, 393:114778, 2022

2022
[22]

Fourier Neural Operator for Parametric Partial Differential Equations

Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895, 2020

work page internal anchor Pith review arXiv 2010
[23]

arXiv preprint arXiv:2205.02191 , year=

Tapas Tripura and Souvik Chakraborty. Wavelet neural operator: a neural operator for parametric partial differential equations.arXiv preprint arXiv:2205.02191, 2022

work page arXiv 2022
[24]

Learning in latent spaces improves the predictive accuracy of deep neural operators.arXiv preprint arXiv:2304.07599, 2023

Katiana Kontolati, Somdatta Goswami, George Em Karniadakis, and Michael D Shields. Learning in latent spaces improves the predictive accuracy of deep neural operators.arXiv preprint arXiv:2304.07599, 2023

work page arXiv 2023
[25]

Extended physics-informed neural networks (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations

Ameya D Jagtap and George Em Karniadakis. Extended physics-informed neural networks (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations. Communications in Computational Physics, 28(5), 2020

2020
[26]

Multi-domain physics-informed neural network for solving forward and inverse problems of steady-state heat conduction in multilayer media.Physics of Fluids, 34(11), 2022

Benrong Zhang, Guozheng Wu, Yan Gu, Xiao Wang, and Fajie Wang. Multi-domain physics-informed neural network for solving forward and inverse problems of steady-state heat conduction in multilayer media.Physics of Fluids, 34(11), 2022

2022
[27]

Interface pinns (i-pinns): A physics-informed neural networks framework for interface problems.Computer Methods in Applied Mechanics and Engineering, 429:117135, 2024

Antareep Kumar Sarma, Sumanta Roy, Chandrasekhar Annavarapu, Pratanu Roy, and Shriram Jagannathan. Interface pinns (i-pinns): A physics-informed neural networks framework for interface problems.Computer Methods in Applied Mechanics and Engineering, 429:117135, 2024

2024
[28]

Adaptive interface-pinns (adai- pinns): An efficient physics-informed neural networks framework for interface problems.Communications in Computational Physics, 37(3):603–622, 2025

Sumanta Roy, Chandrasekhar Annavarapu, Antareep Kumar Sarma, et al. Adaptive interface-pinns (adai- pinns): An efficient physics-informed neural networks framework for interface problems.Communications in Computational Physics, 37(3):603–622, 2025

2025
[29]

Ae-pinns: Attention-enhanced physics-informed neural networks for solving elliptic interface problems.arXiv preprint arXiv:2506.18332, 2025

Jiachun Zheng, Yunqing Huang, and Nianyu Yi. Ae-pinns: Attention-enhanced physics-informed neural networks for solving elliptic interface problems.arXiv preprint arXiv:2506.18332, 2025

work page arXiv 2025
[30]

Hg-pinn: A residual physics-informed neural network framework for heterogeneous groundwater flow simulation.Desalination and Water Treatment, page 101577, 2025

Zhang Yuan, Yongxia Wu, and Chenyu Sun. Hg-pinn: A residual physics-informed neural network framework for heterogeneous groundwater flow simulation.Desalination and Water Treatment, page 101577, 2025

2025
[31]

wbpinn: Weight balanced physics-informed neural networks for multi-objective learning.Applied Soft Computing, 170:112632, 2025

Fujun Cao, Xiaobin Guo, Xinzheng Dong, and Dongfang Yuan. wbpinn: Weight balanced physics-informed neural networks for multi-objective learning.Applied Soft Computing, 170:112632, 2025

2025
[32]

Ig-pinns: Interface-gated physics-informed neural networks for solving elliptic interface problems.Journal of Computational Physics, page 114540, 2025

Jiachun Zheng, Yunqing Huang, and Nianyu Yi. Ig-pinns: Interface-gated physics-informed neural networks for solving elliptic interface problems.Journal of Computational Physics, page 114540, 2025

2025
[33]

A discontinuity capturing shallow neural network for elliptic interface problems.Journal of Computational Physics, 469:111576, 2022

Wei-Fan Hu, Te-Sheng Lin, and Ming-Chih Lai. A discontinuity capturing shallow neural network for elliptic interface problems.Journal of Computational Physics, 469:111576, 2022

2022
[34]

A discontinuity-capturing neural network with categorical embedding and its application to anisotropic elliptic interface problems.arXiv preprint arXiv:2503.15441, 2025

Wei-Fan Hu, Te-Sheng Lin, and Ming-Chih Lai. A discontinuity-capturing neural network with categorical embedding and its application to anisotropic elliptic interface problems.arXiv preprint arXiv:2503.15441, 2025

work page arXiv 2025
[35]

Solving parametric elliptic interface problems via interfaced operator network.Journal of Computational Physics, 514:113217, 2024

Sidi Wu, Aiqing Zhu, Yifa Tang, and Benzhuo Lu. Solving parametric elliptic interface problems via interfaced operator network.Journal of Computational Physics, 514:113217, 2024

2024
[36]

Physics-informed tailored finite point operator network for parametric interface problems.arXiv preprint arXiv:2409.10284, 2024

Ting Du, Xianliang Xu, Wang Kong, Ye Li, and Zhongyi Huang. Physics-informed tailored finite point operator network for parametric interface problems.arXiv preprint arXiv:2409.10284, 2024

work page arXiv 2024
[37]

Mionet: Learning multiple-input operators via tensor product.SIAM Journal on Scientific Computing, 44(6):A3490–A3514, 2022

Pengzhan Jin, Shuai Meng, and Lu Lu. Mionet: Learning multiple-input operators via tensor product.SIAM Journal on Scientific Computing, 44(6):A3490–A3514, 2022

2022
[38]

arXiv preprint arXiv:1604.06737 , year=

Cheng Guo and Felix Berkhahn. Entity embeddings of categorical variables.arXiv preprint arXiv:1604.06737, 2016

work page arXiv 2016
[39]

Exact imposition of boundary conditions with distance functions in physics-informed deep neural networks.Computer Methods in Applied Mechanics and Engineering, 389:114333, 2022

Natarajan Sukumar and Ankit Srivastava. Exact imposition of boundary conditions with distance functions in physics-informed deep neural networks.Computer Methods in Applied Mechanics and Engineering, 389:114333, 2022. 23 APREPRINT- APRIL10, 2026

2022
[40]

Adam: A Method for Stochastic Optimization

Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017. URLhttps://arxiv. org/abs/1412.6980

work page internal anchor Pith review Pith/arXiv arXiv 2017
[41]

SOAP: Improving and Stabilizing Shampoo using Adam

Nikhil Vyas, Depen Morwani, Rosie Zhao, Mujin Kwun, Itai Shapira, David Brandfonbrener, Lucas Janson, and Sham Kakade. Soap: Improving and stabilizing shampoo using adam.arXiv preprint arXiv:2409.11321, 2024

work page internal anchor Pith review arXiv 2024
[42]

Understanding the difficulty of training deep feedforward neural networks

Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. InProceedings of the thirteenth international conference on artificial intelligence and statistics, pages 249–256. JMLR Workshop and Conference Proceedings, 2010

2010
[43]

Simon and Schuster, 2024

Grigory Sapunov.Deep learning with JAX. Simon and Schuster, 2024

2024
[44]

Learning the solution operator of parametric partial differential equations with physics-informed deeponets.Science advances, 7(40):eabi8605, 2021

Sifan Wang, Hanwen Wang, and Paris Perdikaris. Learning the solution operator of parametric partial differential equations with physics-informed deeponets.Science advances, 7(40):eabi8605, 2021

2021
[45]

Application of the finite-element method to the diffusion and reaction of chemical species in multilayered polymeric bodies.Mathematical Modelling, 7(2-3):385–395, 1986

DT Wadiak. Application of the finite-element method to the diffusion and reaction of chemical species in multilayered polymeric bodies.Mathematical Modelling, 7(2-3):385–395, 1986

1986
[46]

Modeling diffusion and types iv sorption of water vapor in heterogeneous systems.Chemical Engineering Science, 275:118695, 2023

Stephen T Castonguay, Pratanu Roy, Yunwei Sun, Sylvie Aubry, Brandon Foley, Elizabeth A Glascoe, and Hom N Sharma. Modeling diffusion and types iv sorption of water vapor in heterogeneous systems.Chemical Engineering Science, 275:118695, 2023. 24

2023