pith. sign in

arxiv: 2510.03248 · v3 · submitted 2025-09-26 · 💻 cs.LG · cs.AI· cs.CV· physics.med-ph

Multimodal Neural Operators for Real-Time Biomechanical Modelling of Traumatic Brain Injury

Anusha Agarwal , Dibakar Roy Sarkar , Somdatta Goswami This is my paper

Pith reviewed 2026-05-18 13:12 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CVphysics.med-ph
keywords neural operatorstraumatic brain injurymultimodal fusionbrain displacementmagnetic resonance elastographyreal-time modelingfinite element alternatives
0
0 comments X

The pith

Multimodal neural operators predict full-field brain displacement from imaging and patient data much faster than finite element solvers.

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

The paper investigates whether neural operators can combine three-dimensional brain scans with demographic details and scan settings to forecast how brain tissue displaces under mechanical stress. Standard physics simulations deliver precise results but run too slowly for routine medical decisions after head injury. The authors compare four operator architectures on 249 real vibration measurements taken at frequencies from 20 to 90 Hz, using two different ways to merge the volumetric and scalar inputs. They report that one architecture reaches low error rates with very few parameters and rapid inference, while others trade accuracy for lower memory use. If the findings hold, injury modeling could move from slow offline calculations to on-the-spot estimates during clinical evaluation.

Core claim

Neural operators augmented with multimodal fusion can accurately predict full-field brain displacement from heterogeneous inputs of volumetric neuroimaging, demographic parameters, and acquisition metadata. On 249 in vivo MRE datasets across 20-90 Hz, DeepONet reached the lowest error on real displacement components (MSE 0.0039, 90 percent accuracy) with the fastest inference rate and smallest parameter count, while Multi-Grid FNO performed best on imaginary components and used the least GPU memory among FNO variants; no single model led on every performance measure.

What carries the argument

Multimodal fusion inside neural operators, implemented via field projection for Fourier Neural Operator variants and branch decomposition for Deep Operator Networks, to merge volumetric imaging with scalar features and learn mappings onto displacement fields.

If this is right

  • Biomechanical evaluation of traumatic brain injury can shift from offline finite element runs to real-time inference in clinical workflows.
  • Clinicians can select among operator architectures according to whether they prioritize accuracy on real displacement, memory footprint, or overall speed.
  • The same fusion approach supplies a template for adding other patient-specific data types to soft-tissue mechanics predictions without rebuilding the underlying solver.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Retraining on data from actual high-strain injury events rather than controlled low-amplitude vibrations could reveal whether the models capture the nonlinear regimes that matter most for trauma.
  • Linking the predicted displacement fields to long-term patient recovery records would test whether the outputs carry useful information for treatment planning.
  • The observed accuracy-speed trade-offs suggest that ensembles or hybrid operator designs might deliver better performance across all criteria than any single architecture alone.

Load-bearing premise

The 249 vibration measurements from living subjects at frequencies 20 to 90 Hz capture enough of the mechanical behavior that appears in real traumatic brain injuries for the trained models to work on new patients without large drops in accuracy.

What would settle it

Apply the trained models to an independent collection of magnetic resonance elastography scans acquired on different equipment or from patients with documented traumatic brain injury and check whether mean squared error on displacement fields rises above 0.01.

Figures

Figures reproduced from arXiv: 2510.03248 by Anusha Agarwal, Dibakar Roy Sarkar, Somdatta Goswami.

Figure 1
Figure 1. Figure 1: Representative visualization of the dataset, showing T1-weighted anatomical MRI (left), brain mask (middle), [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Fourier Neural Operator (FNO) Architecture for Brain Displacement Prediction. The network takes T1 MRI [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Multi-Grid Fourier Neural Operator (FNO) Architecture for Hierarchical Brain Displacement Prediction. The [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Deep Operator Network (DeepONet) Architecture for Brain Displacement Prediction. The architecture [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Training and Validation Losses across Models. Both the FNO and F-FNO were able to converge much [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FNO Prediction Artifacts in Brain Displacement Fields. Orthogonal views showing ground truth (top), FNO [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of Model Errors and Predictions across Real and Imaginary Displacement Fields. The FNO [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
read the original abstract

Background: Traumatic brain injury modeling requires integrating volumetric neuroimaging, demographic parameters, and acquisition metadata. Finite element solvers are too computationally expensive for clinical settings. Neural operators offer much faster inference. Their ability to integrate volumetric imaging with scalar metadata remains underexplored for biomechanical predictions. Objective: This study evaluates multimodal neural operator architectures for brain biomechanics. We test strategies fusing volumetric anatomical imaging, demographic features, and acquisition parameters to predict full-field brain displacement from MRE data. Methods: We framed TBI modeling as a multimodal operator learning problem. Two fusion strategies were tested. Field projection was applied for Fourier Neural Operator (FNO) architectures. Branch decomposition was used for Deep Operator Networks (DeepONet). Four models (FNO, Factorized FNO, Multi-Grid FNO, DeepONet) were evaluated on 249 in vivo MRE datasets across frequencies from 20 to 90 Hz. Results: DeepONet achieved the highest accuracy on real displacement fields (MSE = 0.0039, 90.0% accuracy) with the fastest inference (3.83 it/s) and fewest parameters (2.09M). MG-FNO performed best on imaginary fields (MSE = 0.0058, 88.3% accuracy) requiring the lowest GPU memory among FNO variants (7.12 GB). No single architecture dominated all criteria. This reveals distinct trade-offs between accuracy, spatial fidelity, and computational cost. Conclusion: Neural operators augmented with multimodal fusion can accurately predict full-field brain displacement from heterogeneous inputs. They offer inference times orders of magnitude faster than finite element solvers. This comparison provides guidance for selecting operator learning approaches in biomedical settings.

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.

Referee Report

3 major / 2 minor

Summary. The manuscript evaluates multimodal neural operator architectures (FNO variants and DeepONet) for predicting full-field brain displacement from heterogeneous inputs including volumetric neuroimaging, demographic parameters, and acquisition metadata. It frames TBI modeling as a multimodal operator learning problem, tests field projection and branch decomposition fusion strategies, and reports performance on 249 in vivo MRE datasets across 20-90 Hz frequencies, with DeepONet achieving MSE=0.0039 and 90% accuracy on real fields alongside fast inference times.

Significance. If validated rigorously, the results would support neural operators as practical surrogates for real-time biomechanical modeling by delivering orders-of-magnitude faster inference than finite element methods while integrating multimodal data. The reported trade-offs across architectures offer concrete selection guidance for biomedical applications of operator learning.

major comments (3)
  1. Results section: The reported MSE values (0.0039 for DeepONet on real fields, 0.0058 for MG-FNO on imaginary fields) and accuracy percentages lack train-test split details, error bars, statistical tests, or baseline comparisons to finite element solvers, which are required to substantiate the central accuracy and speedup claims on the 249 datasets.
  2. Methods and Conclusion: The applicability to real-time TBI biomechanical modeling rests on the untested assumption that steady-state linear MRE responses at 20-90 Hz represent transient high-strain-rate TBI conditions; no domain-shift experiments or TBI-specific validation are described, making this assumption load-bearing for the title and conclusion claims.
  3. §3 (Fusion strategies): The specific implementation details and hyperparameter choices for multimodal fusion (field projection in FNO, branch decomposition in DeepONet) are not ablated, so the contribution of the multimodal component to the reported performance cannot be isolated from the base operator architectures.
minor comments (2)
  1. Abstract: The term '90.0% accuracy' is used without defining the underlying threshold or metric (e.g., relative error cutoff), reducing clarity of the quantitative claims.
  2. Overall: The manuscript would benefit from explicit discussion of limitations, including risks of overfitting to MRE acquisition conditions and generalization to new clinical cases.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. We address each major comment below and indicate where revisions will be made to the manuscript.

read point-by-point responses

  1. Referee: Results section: The reported MSE values (0.0039 for DeepONet on real fields, 0.0058 for MG-FNO on imaginary fields) and accuracy percentages lack train-test split details, error bars, statistical tests, or baseline comparisons to finite element solvers, which are required to substantiate the central accuracy and speedup claims on the 249 datasets.

    Authors: We agree these details are necessary. The 249 datasets were partitioned into 70% training, 15% validation, and 15% test sets using subject-wise splits to avoid data leakage; all metrics are reported on the test set. We will add error bars from five independent runs with different random seeds and include Wilcoxon signed-rank tests for pairwise architecture comparisons. For FE baselines, we will cite representative runtimes from the TBI biomechanics literature (typically minutes to hours per subject on comparable hardware) to support the reported inference speedup. These clarifications and any supporting tables will be added to the Results section. revision: yes

  2. Referee: Methods and Conclusion: The applicability to real-time TBI biomechanical modeling rests on the untested assumption that steady-state linear MRE responses at 20-90 Hz represent transient high-strain-rate TBI conditions; no domain-shift experiments or TBI-specific validation are described, making this assumption load-bearing for the title and conclusion claims.

    Authors: We acknowledge that MRE provides steady-state harmonic data and that direct validation on transient, high-strain-rate TBI scenarios is not performed in this study. MRE is a widely accepted non-invasive proxy for in vivo brain mechanics in the literature. To address the concern, we will revise the title, abstract, introduction, and conclusion to explicitly state that the models are trained and evaluated on MRE displacement fields and to discuss the modeling assumptions and scope limitations regarding extrapolation to TBI. A new limitations subsection will be added. New domain-shift experiments are outside the scope of the current dataset and would require additional data collection. revision: partial

  3. Referee: §3 (Fusion strategies): The specific implementation details and hyperparameter choices for multimodal fusion (field projection in FNO, branch decomposition in DeepONet) are not ablated, so the contribution of the multimodal component to the reported performance cannot be isolated from the base operator architectures.

    Authors: We agree that ablation experiments are needed to isolate the multimodal contribution. In the revised manuscript we will expand §3 with full implementation details (projection layer dimensions, branch network widths, and all hyperparameters) and add an ablation study comparing each multimodal architecture against its unimodal counterpart (volumetric input only). Performance deltas will be reported to quantify the benefit of incorporating demographic and acquisition metadata. These results will be presented in a new table or figure. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical evaluation on held-out MRE data is independent of fitted inputs

full rationale

The paper frames the task as supervised operator learning on 249 in vivo MRE displacement fields (20-90 Hz), trains multimodal architectures (FNO variants and DeepONet), and reports test-set MSE and accuracy metrics. No load-bearing step reduces a claimed prediction to a quantity defined solely by the model's own fitted parameters or by a self-citation chain; the central results are standard held-out empirical performance numbers rather than tautological re-statements of training data. The TBI applicability claim rests on an untested generalization assumption, but that is a validity issue, not a circularity in the derivation.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard machine-learning assumptions about operator approximability and on the representativeness of the MRE dataset for TBI biomechanics; no new physical entities are introduced.

free parameters (1)
  • Fusion hyperparameters
    Specific implementation details of field projection and branch decomposition likely involve tunable parameters whose values are not reported in the abstract.
axioms (1)
  • domain assumption Neural operators can learn accurate mappings from combined volumetric and scalar inputs to displacement fields
    Invoked in the objective and methods description when framing TBI modeling as a multimodal operator learning problem.

pith-pipeline@v0.9.0 · 5856 in / 1414 out tokens · 48207 ms · 2026-05-18T13:12:32.716823+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages · 2 internal anchors

  1. [1]

    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

    work page 2023

  2. [2]

    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

    work page 2024

  3. [3]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhattacharya, A. Stuart, and A. Anandkumar. Fourier neural operator for parametric partial differential equations, 2021

    work page 2021

  4. [4]

    L. Lu, P. Jin, and G. E. Karniadakis. Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators, 2019

    work page 2019

  5. [5]

    FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators

    Jaideep Pathak, Shashank Subramanian, Peter Harrington, Sanjeev Raja, Ashesh Chattopadhyay, Morteza Mardani, Thorsten Kurth, David Hall, Zongyi Li, Kamyar Azizzadenesheli, et al. Fourcastnet: A global data-driven high- resolution weather model using adaptive fourier neural operators, arxiv.arXiv preprint arXiv:2202.11214, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  6. [6]

    Real-time high-resolution co 2 geological storage prediction using nested fourier neural operators.Energy & Environmental Science, 16(4):1732–1741, 2023

    Gege Wen, Zongyi Li, Qirui Long, Kamyar Azizzadenesheli, Anima Anandkumar, and Sally M Benson. Real-time high-resolution co 2 geological storage prediction using nested fourier neural operators.Energy & Environmental Science, 16(4):1732–1741, 2023

    work page 2023

  7. [7]

    Deep transfer operator learning for partial differential equations under conditional shift.Nature Machine Intelligence, 4(12):1155–1164, 2022

    Somdatta Goswami, Katiana Kontolati, Michael D Shields, and George Em Karniadakis. Deep transfer operator learning for partial differential equations under conditional shift.Nature Machine Intelligence, 4(12):1155–1164, 2022

    work page 2022

  8. [8]

    TBI Data | Traumatic Brain Injury and Concussion

    Centers for Disease Control and Prevention. TBI Data | Traumatic Brain Injury and Concussion. https: //www.cdc.gov/traumatic-brain-injury/data-research/index.html , 2024. Published October 29,

    work page 2024

  9. [9]

    Accessed September 20, 2025

    work page 2025

  10. [10]

    Traumatic Brain Injury: A Roadmap for Accelerating Progress

    Chanel Matney, Katherine Bowman, and Donald Berwick. Traumatic Brain Injury: A Roadmap for Accelerating Progress. 2022

    work page 2022

  11. [11]

    L. Hu, S. Yang, B. Jin, and C. Wang. Advanced neuroimaging role in traumatic brain injury: A narrative review. Frontiers in Neuroscience, 16:872609, 2022

    work page 2022

  12. [12]

    Bahreinizad, S

    H. Bahreinizad, S. K. Chowdhury, G. Paulon, L. Wei, and F. Z. Santos. Development and validation of an mri-derived head-neck finite element model.bioRxiv, 2023

    work page 2023

  13. [13]

    A. D. Gomez, A. K. Knutsen, F. Xing, Y . C. Lu, D. Chan, D. L. Pham, P. V . Bayly, and J. L. Prince. 3-d measurements of acceleration-induced brain deformation via harmonic phase analysis and finite-element models. IEEE Transactions on Biomedical Engineering, 66(5):1456–1467, 2019

    work page 2019

  14. [14]

    Griffiths and S

    E. Griffiths and S. Budday. Finite element modeling of traumatic brain injury: Areas of future interest.Current Opinion in Biomedical Engineering, 24:100421, 2022

    work page 2022

  15. [15]

    S. Wu, W. Zhao, and S. Ji. Real-time dynamic simulation for highly accurate spatiotemporal brain deformation from impact.Computer Methods in Applied Mechanics and Engineering, 394:114913, 2022

    work page 2022

  16. [16]

    Vectorized conditional neu- ral fields: A framework for solving time-dependent parametric partial differential equations.arXiv preprint arXiv:2406.03919, 2024

    Jan Hagnberger, Marimuthu Kalimuthu, Daniel Musekamp, and Mathias Niepert. Vectorized conditional neu- ral fields: A framework for solving time-dependent parametric partial differential equations.arXiv preprint arXiv:2406.03919, 2024

    work page arXiv 2024

  17. [17]

    Fusion deeponet: A data-efficient neural operator for geometry-dependent hypersonic flows on arbitrary grids.arXiv e-prints, pages arXiv–2501, 2025

    Ahmad Peyvan and Varun Kumar. Fusion deeponet: A data-efficient neural operator for geometry-dependent hypersonic flows on arbitrary grids.arXiv e-prints, pages arXiv–2501, 2025

    work page 2025

  18. [18]

    Rethinking materials simulations: Blending direct numerical simulations with neural operators.npj Computational Materials, 10(1):145, 2024

    Vivek Oommen, Khemraj Shukla, Saaketh Desai, Rémi Dingreville, and George Em Karniadakis. Rethinking materials simulations: Blending direct numerical simulations with neural operators.npj Computational Materials, 10(1):145, 2024

    work page 2024

  19. [19]

    Multifidelity deep operator networks for data-driven and physics-informed problems.Journal of Computational Physics, 493:112462, 2023

    Amanda A Howard, Mauro Perego, George Em Karniadakis, and Panos Stinis. Multifidelity deep operator networks for data-driven and physics-informed problems.Journal of Computational Physics, 493:112462, 2023

    work page 2023

  20. [20]

    A composite neural network that learns from multi-fidelity data: Application to function approximation and inverse pde problems.Journal of Computational Physics, 401:109020, 2020

    Xuhui Meng and George Em Karniadakis. A composite neural network that learns from multi-fidelity data: Application to function approximation and inverse pde problems.Journal of Computational Physics, 401:109020, 2020. 16 APREPRINT- DECEMBER23, 2025

    work page 2020

  21. [21]

    Philip V Bayly, Ahmed Alshareef, Andrew K Knutsen, Kshitiz Upadhyay, Ruth J Okamoto, Aaron Carass, John A Butman, Dzung L Pham, Jerry L Prince, KT Ramesh, et al. Mr imaging of human brain mechanics in vivo: new measurements to facilitate the development of computational models of brain injury.Annals of biomedical engineering, 49(10):2677–2692, 2021

    work page 2021

  22. [22]

    Chen and H

    T. Chen and H. Chen. Universal approximation to nonlinear operators by neural networks with application to dynamical systems.IEEE Transactions on Neural Networks, 6(4):911–917, 1995

    work page 1995

  23. [23]

    Factorized fourier neural operators

    Alasdair Tran, Alexander Mathews, Lexing Xie, and Cheng Soon Ong. Factorized fourier neural operators. In International Conference on Learning Representations (ICLR), pages 417–422. IEEE, 2023

    work page 2023

  24. [24]

    Multi-grid tensorized fourier neural operator for high-resolution pdes, 2023

    Jean Kossaifi, Nikola Kovachki, Kamyar Azizzadenesheli, and Anima Anandkumar. Multi-grid tensorized fourier neural operator for high-resolution pdes, 2023

    work page 2023

  25. [25]

    Learning Hidden Physics and System Parameters with Deep Operator Networks

    Vijay Kag, Dibakar Roy Sarkar, Birupaksha Pal, and Somdatta Goswami. Learning hidden physics and system parameters with deep operator networks.arXiv preprint arXiv:2412.05133, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [26]

    Neural operators for stochastic modeling of nonlinear structural system response to natural hazards.arXiv preprint arXiv:2502.11279, 2025

    Somdatta Goswami, Dimitris G Giovanis, Bowei Li, Seymour MJ Spence, and Michael D Shields. Neural operators for stochastic modeling of nonlinear structural system response to natural hazards.arXiv preprint arXiv:2502.11279, 2025

    work page arXiv 2025

  27. [27]

    Fourcastnet: Accelerating global high-resolution weather forecasting using adaptive fourier neural operators

    Thorsten Kurth, Shashank Subramanian, Peter Harrington, Jaideep Pathak, Morteza Mardani, David Hall, Andrea Miele, Karthik Kashinath, and Anima Anandkumar. Fourcastnet: Accelerating global high-resolution weather forecasting using adaptive fourier neural operators. InProceedings of the platform for advanced scientific computing conference, pages 1–11, 2023

    work page 2023

  28. [28]

    Physics-informed deep neural operator networks

    Somdatta Goswami, Aniruddha Bora, Yue Yu, and George Em Karniadakis. Physics-informed deep neural operator networks. InMachine learning in modeling and simulation: methods and applications, pages 219–254. Springer, 2023

    work page 2023

  29. [29]

    Known operator learning and hybrid machine learning in medical imaging—a review of the past, the present, and the future.Progress in Biomedical Engineering, 4(2):022002, 2022

    Andreas Maier, Harald Köstler, Marco Heisig, Patrick Krauss, and Seung Hee Yang. Known operator learning and hybrid machine learning in medical imaging—a review of the past, the present, and the future.Progress in Biomedical Engineering, 4(2):022002, 2022

    work page 2022

  30. [30]

    Laplace neural operator for solving differential equations.Nature Machine Intelligence, 6(6):631–640, 2024

    Qianying Cao, Somdatta Goswami, and George Em Karniadakis. Laplace neural operator for solving differential equations.Nature Machine Intelligence, 6(6):631–640, 2024

    work page 2024

  31. [31]

    Learning nonlinear oper- ators in latent spaces for real-time predictions of complex dynamics in physical systems.Nature Communications, 15(1):5101, 2024

    Katiana Kontolati, Somdatta Goswami, George Em Karniadakis, and Michael D Shields. Learning nonlinear oper- ators in latent spaces for real-time predictions of complex dynamics in physical systems.Nature Communications, 15(1):5101, 2024

    work page 2024

  32. [32]

    Karumuri, L

    Sharmila Karumuri, Lori Graham-Brady, and Somdatta Goswami. Physics-informed latent neural operator for real-time predictions of complex physical systems.arXiv preprint arXiv:2501.08428, 2025

    work page arXiv 2025

  33. [33]

    Time-marching neural operator– fe coupling: Ai-accelerated physics modeling.Computer Methods in Applied Mechanics and Engineering, 446:118319, 2025

    Wei Wang, Maryam Hakimzadeh, Haihui Ruan, and Somdatta Goswami. Time-marching neural operator– fe coupling: Ai-accelerated physics modeling.Computer Methods in Applied Mechanics and Engineering, 446:118319, 2025

    work page 2025

  34. [34]

    Adaptive interface-pinns (adai-pinns) for inverse problems: Determining material properties for heterogeneous systems.Finite Elements in Analysis and Design, 249:104373, 2025

    Dibakar Roy Sarkar, Chandrasekhar Annavarapu, and Pratanu Roy. Adaptive interface-pinns (adai-pinns) for inverse problems: Determining material properties for heterogeneous systems.Finite Elements in Analysis and Design, 249:104373, 2025

    work page 2025

  35. [35]

    Interface physics-informed neural networks (i-pinns) to solve inverse problems in heterogeneous materials

    Dibakar Roy Sarkar, Chandrasekhar Annavarapu, and Pratanu Roy. Interface physics-informed neural networks (i-pinns) to solve inverse problems in heterogeneous materials. 2025

    work page 2025

  36. [36]

    Nayak, S

    Dibyajyoti Nayak and Somdatta Goswami. Ti-deeponet: Learnable time integration for stable long-term extrapo- lation.arXiv preprint arXiv:2505.17341, 2025

    work page arXiv 2025

  37. [37]

    Physics-informed time-integrated deeponet: Temporal tangent space operator learning for high-accuracy inference.arXiv preprint arXiv:2508.05190, 2025

    Luis Mandl, Dibyajyoti Nayak, Tim Ricken, and Somdatta Goswami. Physics-informed time-integrated deeponet: Temporal tangent space operator learning for high-accuracy inference.arXiv preprint arXiv:2508.05190, 2025

    work page arXiv 2025

  38. [38]

    A resolution independent neural operator.Computer Methods in Applied Mechanics and Engineering, 444:118113, 2025

    Bahador Bahmani, Somdatta Goswami, Ioannis G Kevrekidis, and Michael D Shields. A resolution independent neural operator.Computer Methods in Applied Mechanics and Engineering, 444:118113, 2025

    work page 2025

  39. [39]

    Luis Mandl, Somdatta Goswami, Lena Lambers, and Tim Ricken. Separable physics-informed deeponet: Breaking the curse of dimensionality in physics-informed machine learning.Computer Methods in Applied Mechanics and Engineering, 434:117586, 2025

    work page 2025

  40. [40]

    Synergistic learning with multi-task deeponet for efficient pde problem solving.Neural Networks, 184:107113, 2025

    Varun Kumar, Somdatta Goswami, Katiana Kontolati, Michael D Shields, and George Em Karniadakis. Synergistic learning with multi-task deeponet for efficient pde problem solving.Neural Networks, 184:107113, 2025. 17 APREPRINT- DECEMBER23, 2025

    work page 2025

  41. [41]

    Accelerating hamiltonian monte carlo for bayesian inference in neural networks and neural operators.Computer Methods in Applied Mechanics and Engineering, 447:118401, 2025

    Ponkrshnan Thiagarajan, Tamer A Zaki, and Michael D Shields. Accelerating hamiltonian monte carlo for bayesian inference in neural networks and neural operators.Computer Methods in Applied Mechanics and Engineering, 447:118401, 2025

    work page 2025

  42. [42]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. Pytorch: An imperative style, high-performance deep learning library, 2019

    work page 2019

  43. [43]

    Hendrycks and K

    D. Hendrycks and K. Gimpel. Gaussian error linear units (gelus), 2016

    work page 2016

  44. [44]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization, 2017. 18 APREPRINT- DECEMBER23, 2025 (a) Model Error Comparison (Real Displacement) (b) Model Prediction Comparison (Real Displacement) (c) Model Error Comparison (Imaginary Displacement) (d) Model Prediction Comparison (Imaginary Displace- ment) Figure 7: Comparison of Model Errors and P...

    work page 2017