pith. sign in

arxiv: 2606.00892 · v1 · pith:2QQIXHYMnew · submitted 2026-05-30 · 💻 cs.LG · cs.CE· physics.comp-ph

An Exploratory Study into using Machine-Learning for Fast Step-by-step Emulation of Numerical Mechanical Thrombectomy Simulations for Ischemic Stroke

Thijs Stessen (University of Amsterdam) This is my paper

Pith reviewed 2026-06-28 18:49 UTC · model grok-4.3

classification 💻 cs.LG cs.CEphysics.comp-ph
keywords machine learningsurrogate modelingmechanical thrombectomyischemic strokenumerical simulationautoregressive stabilitydata augmentation
0
0 comments X

The pith

Machine learning surrogates accurately predict single steps of mechanical thrombectomy simulations with large speedups but become unstable over longer autoregressive runs in complex geometries.

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

The paper tests whether current machine learning models can replace slow numerical physics simulations of mechanical thrombectomy, the procedure used to remove clots in ischemic stroke. Three surrogate models were trained on two simplified aspiration simulations that differed in geometric complexity. Two of the models matched individual simulation steps well and delivered substantial speedups, especially after data augmentation, yet all models drifted when rolled out step by step over many time steps in the more complex geometry.

Core claim

Surrogate models trained on simulation data can accurately predict individual time steps of mechanical thrombectomy simulations and deliver large computational speedups, but they lose stability when used autoregressively to emulate full sequences in geometries of realistic complexity.

What carries the argument

Step-by-step machine learning surrogate models trained to replace the numerical solver at each time increment of the aspiration procedure.

If this is right

  • Single-step surrogates can replace the solver for short segments or simple geometries where stability is not required.
  • Specific data augmentations measurably raise single-step accuracy and speedup.
  • The current approach cannot yet replace full simulations for realistic patient geometries over clinically relevant time scales.

Where Pith is reading between the lines

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

  • Adding physics-based loss terms or constraint layers during training could reduce drift in long rollouts.
  • Hybrid schemes that periodically correct the surrogate with a short full-physics segment might extend usable horizon without losing all speed gains.
  • If stability is achieved, the same training pipeline could be applied to other interventional radiology simulations that currently run too slowly for intra-procedural use.

Load-bearing premise

Single-step accuracy will carry over to stable, long-horizon autoregressive predictions without external correction or retraining.

What would settle it

Compare the final clot position and flow field after 500 autoregressive surrogate steps against a full numerical run on the same complex geometry; large divergence would falsify stability.

Figures

Figures reproduced from arXiv: 2606.00892 by Thijs Stessen (University of Amsterdam).

Figure 1.1
Figure 1.1. Figure 1.1: The step-by-step approach using a machine learning model as a [PITH_FULL_IMAGE:figures/full_fig_p007_1_1.png] view at source ↗
Figure 1.2
Figure 1.2. Figure 1.2: Overview of a possible pipeline of the use of digital twins in ischemic stroke treat [PITH_FULL_IMAGE:figures/full_fig_p008_1_2.png] view at source ↗
Figure 2.1
Figure 2.1. Figure 2.1: Schematic view of a typical example of the circle of Willis. The Middle Cerebral [PITH_FULL_IMAGE:figures/full_fig_p011_2_1.png] view at source ↗
Figure 2.2
Figure 2.2. Figure 2.2: Different types of blood clots (thrombi). Calcium binds to the components of the [PITH_FULL_IMAGE:figures/full_fig_p011_2_2.png] view at source ↗
Figure 2.3
Figure 2.3. Figure 2.3: First picture: used stent retriever. Second picture: Example stent retriever and [PITH_FULL_IMAGE:figures/full_fig_p013_2_3.png] view at source ↗
Figure 2.4
Figure 2.4. Figure 2.4: Basic Transformer model. Image shows two inputs. Taken from [49]. [PITH_FULL_IMAGE:figures/full_fig_p018_2_4.png] view at source ↗
Figure 2.5
Figure 2.5. Figure 2.5: Overview of the Transolver architecture. On the left the Transolver blocks can be [PITH_FULL_IMAGE:figures/full_fig_p020_2_5.png] view at source ↗
Figure 2.6
Figure 2.6. Figure 2.6: Overview of the Erwin Architecture. On the left the construction of the nodes into [PITH_FULL_IMAGE:figures/full_fig_p021_2_6.png] view at source ↗
Figure 2.7
Figure 2.7. Figure 2.7: Example balltree formation. P is the set of nodes and T is the collection of layers of the tree. The permutation π() could for example be a rotation as described below. Taken directly from [21]. information exchange. Lastly, it uses a learned embedding in the form of a Message Passing Neural Network (MPNN), see section 2.3. In detail, Erwin works as follows. First, we will explain the BallProcessing bloc… view at source ↗
Figure 2.8
Figure 2.8. Figure 2.8: Overview of the architecture of MeshGraphNet. As can be seen it uses an encoder, [PITH_FULL_IMAGE:figures/full_fig_p023_2_8.png] view at source ↗
Figure 3.1
Figure 3.1. Figure 3.1: Frame from BendVessel (left) and ClotEntry (right). The blood clot is shown in [PITH_FULL_IMAGE:figures/full_fig_p027_3_1.png] view at source ↗
Figure 4.1
Figure 4.1. Figure 4.1: Effects on performance of data augmentations for the validation set of the ClotEntry [PITH_FULL_IMAGE:figures/full_fig_p030_4_1.png] view at source ↗
Figure 4.2
Figure 4.2. Figure 4.2: The effect on performance of adding generated datapoints using BaryCentric sam [PITH_FULL_IMAGE:figures/full_fig_p031_4_2.png] view at source ↗
Figure 4.3
Figure 4.3. Figure 4.3: Rollouts for Erwin, Transolver and MeshGraphNet on BendVessel and ClotEntry. [PITH_FULL_IMAGE:figures/full_fig_p032_4_3.png] view at source ↗
Figure 4.4
Figure 4.4. Figure 4.4: Effect of including different rotations in the training data on performance on the [PITH_FULL_IMAGE:figures/full_fig_p033_4_4.png] view at source ↗
read the original abstract

The treatment of ischemic stroke using mechanical thrombectomy involves difficult decisions under intense time constraints. Numerical physics simulations can in theory inform operators to make better decisions regarding treatment approaches and device selection, but are too slow to do so in practice. In this thesis, we investigate if current machine learning based surrogates can accurately emulate these simulations in a step-by-step manner while making them significantly faster. To do this we train three surrogate models on two simulations that involve a simplified aspiration procedure, with varying levels of geometric complexity. Our results show that two of our models accurately predict singular simulation steps and provide substantial speedups, especially when combined with specific data augmentations. However, the models showed a lack of stability when emulating simulations with complex geometries over longer time periods. Overall, this work provides a foundation for future studies to develop stable methods that scale to realistic numerical physics simulations of mechanical thrombectomy.

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

2 major / 2 minor

Summary. The manuscript is an exploratory study examining whether machine-learning surrogates can emulate step-by-step numerical simulations of mechanical thrombectomy for ischemic stroke. Three surrogate models are trained on data from two simplified aspiration simulations that differ in geometric complexity. The central claims are that two of the models produce accurate single-step predictions and yield substantial speedups (especially with data augmentations), while autoregressive rollouts become unstable on complex geometries over longer time horizons. The work positions itself as providing a foundation for future stable, scalable emulators rather than delivering a production-ready solution.

Significance. If the single-step accuracy and speedup claims are substantiated with quantitative evidence, the study would usefully document the practical difficulties of applying current ML surrogate techniques to even modestly complex hemodynamics problems. The explicit acknowledgment that multi-step stability fails on realistic geometries is a strength, as it correctly bounds the scope and identifies the central open challenge. However, the reliance on highly simplified geometries and the absence of reported metrics limit the immediate impact on either the ML-for-physics or computational-medicine communities.

major comments (2)
  1. [Abstract] Abstract: the assertion that 'two of our models accurately predict singular simulation steps' is presented without any quantitative metrics, error measures, baseline comparisons, training-set sizes, or architecture details. Because this is the primary positive result, the lack of supporting numbers makes the claim impossible to evaluate.
  2. [Abstract] The manuscript reports instability under autoregressive rollout but provides no quantitative characterization (e.g., divergence time, error growth rates, or dependence on geometry complexity). This omission is load-bearing for the paper's own stated boundary condition.
minor comments (2)
  1. The abstract and introduction should explicitly name the three surrogate model families and the two simulation geometries so readers can immediately understand the experimental scope.
  2. Clarify whether the reported speedups include data-generation overhead or only inference time; this distinction is important for assessing practical utility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our exploratory study. We agree that the abstract would benefit from including key quantitative metrics to better support the central claims and will revise it accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that 'two of our models accurately predict singular simulation steps' is presented without any quantitative metrics, error measures, baseline comparisons, training-set sizes, or architecture details. Because this is the primary positive result, the lack of supporting numbers makes the claim impossible to evaluate.

    Authors: The full manuscript details these elements in the Methods (model architectures, training-set sizes from the two simulations) and Results (error measures such as field-wise L2 norms, baseline comparisons to the numerical solver, and speedup factors) sections. We agree the abstract should be self-contained and will revise it to include summary quantitative values for single-step accuracy and speedups. revision: yes

  2. Referee: [Abstract] The manuscript reports instability under autoregressive rollout but provides no quantitative characterization (e.g., divergence time, error growth rates, or dependence on geometry complexity). This omission is load-bearing for the paper's own stated boundary condition.

    Authors: We agree that quantitative characterization strengthens the stated limitations. The manuscript qualitatively describes the instability for complex geometries over longer horizons; we will revise the abstract and results to add metrics such as divergence time steps and error growth rates as a function of geometry complexity. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

This paper is an empirical ML training study that generates simulation data from two simplified aspiration procedures, trains three surrogate models, and reports measured single-step prediction accuracy plus wall-clock speedups on held-out steps. No derivations, equations, or self-referential predictions appear; the central claims rest on direct comparison to external simulation outputs rather than any reduction to fitted inputs or self-citations. The acknowledged instability under autoregressive rollout on complex geometries is explicitly reported as a boundary condition, not an unexamined assumption.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only abstract available; no access to model architectures, loss functions, data generation details, or any fitted hyperparameters.

pith-pipeline@v0.9.1-grok · 5694 in / 1027 out tokens · 81551 ms · 2026-06-28T18:49:42.691051+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

61 extracted references · 5 canonical work pages · 3 internal anchors

  1. [1]

    Global burden, risk factor analysis, and prediction study of ischemic stroke, 1990–2030

    Jiahui Fan, Xiaoguang Li, Xueying Yu, Zhenqiu Liu, Yanfeng Jiang, Yibin Fang, Ming Zong, Chen Suo, Qiuhong Man, and Lize Xiong. Global burden, risk factor analysis, and prediction study of ischemic stroke, 1990–2030. Neurology, 101(2):e137–e150, 2023

    1990

  2. [2]

    Endovascular treat- ment for acute ischaemic stroke in routine clinical practice: prospective, observational cohort study (mr clean registry)

    Ivo GH Jansen, Maxim JHL Mulder, and Robert-Jan B Goldhoorn. Endovascular treat- ment for acute ischaemic stroke in routine clinical practice: prospective, observational cohort study (mr clean registry). bmj, 360, 2018

    2018

  3. [3]

    Trends in acute ischemic stroke treatments and mortality in the united states from 2012 to 2018

    Sharath Kumar Anand, William J Benjamin, Arjun Rohit Adapa, Jiwon V Park, D Andrew Wilkinson, Badih J Daou, James F Burke, and Aditya S Pandey. Trends in acute ischemic stroke treatments and mortality in the united states from 2012 to 2018. Neurosurgical focus, 51(1):E2, 2021

    2012

  4. [4]

    Treatments in ischemic stroke: current and future

    Maria Giulia Mosconi and Maurizio Paciaroni. Treatments in ischemic stroke: current and future. European Neurology, 85(5):349–366, 2022

    2022

  5. [5]

    Overview of mechanical thrombec- tomy techniques

    Stephan A Munich, Kunal Vakharia, and Elad I Levy. Overview of mechanical thrombec- tomy techniques. Neurosurgery, 85(suppl_1):S60–S67, 2019

    2019

  6. [6]

    First pass effect: a new measure for stroke thrombectomy devices

    Osama O Zaidat, Alicia C Castonguay, Italo Linfante, Rishi Gupta, Coleman O Martin, William E Holloway, Nils Mueller-Kronast, Joey D English, Guilherme Dabus, Tim W Malisch, et al. First pass effect: a new measure for stroke thrombectomy devices. Stroke, 49(3):660–666, 2018

    2018

  7. [7]

    Risk of thrombus fragmentation during endovascular stroke treatment

    J Kaesmacher, T Boeckh-Behrens, S Simon, C Maegerlein, JF Kleine, C Zimmer, L Schirmer, H Poppert, and T Huber. Risk of thrombus fragmentation during endovascular stroke treatment. American Journal of Neuroradiology , 38(5):991–998, 2017

    2017

  8. [8]

    Why does mechanical thrombectomy in large vessel occlusion some- times fail? a review of the literature

    Leonard LL Yeo, Pervinder Bhogal, Anil Gopinathan, Yang Cunli, Benjamin Tan, and Tommy Andersson. Why does mechanical thrombectomy in large vessel occlusion some- times fail? a review of the literature. Clinical Neuroradiology, 29(3):401–414, 2019

    2019

  9. [9]

    Lapergue, R

    B. Lapergue, R. Blanc, B. Gory, et al. Effect of endovascular contact aspiration vs stent retriever on revascularization in patients with acute ischemic stroke and large vessel oc- clusion: The aster randomized clinical trial. JAMA, 318(5):443–452, 2017

    2017

  10. [10]

    Computational modeling of interventional and implantable endovascular devices

    Mahdi Daei Daei. Computational modeling of interventional and implantable endovascular devices. PhD thesis, Institut Polytechnique de Paris, 2023

    2023

  11. [11]

    The first virtual patient-specific thrombectomy procedure

    Giulia Luraghi, Sara Bridio, Jose Felix Rodriguez Matas, Gabriele Dubini, Nikki Boodt, Frank JH Gijsen, Aad van der Lugt, Behrooz Fereidoonnezhad, Kevin M Moerman, Patrick McGarry, et al. The first virtual patient-specific thrombectomy procedure. Journal of Biomechanics, 126:110622, 2021. 36

    2021

  12. [12]

    Simulation of stent retriever thrombectomy in acute ischemic stroke by finite element analysis

    Ronghui Liu, Chang Jin, Lizhen Wang, Yisong Yang, Yubo Fan, and Weidong Wang. Simulation of stent retriever thrombectomy in acute ischemic stroke by finite element analysis. Computer Methods in Biomechanics and Biomedical Engineering , 25(7):740–749, 2022

    2022

  13. [13]

    Time-based decision making for reperfusion in acute ischemic stroke

    Mathias Grøan, Johanna Ospel, Soffien Ajmi, Else Charlotte Sandset, Martin W Kurz, Mona Skjelland, and Rajiv Advani. Time-based decision making for reperfusion in acute ischemic stroke. Frontiers in Neurology, 12:728012, 2021

    2021

  14. [14]

    Diag- nosis and management of acute ischemic stroke: speed is critical

    Tapuwa D Musuka, Stephen B Wilton, Mouhieddin Traboulsi, and Michael D Hill. Diag- nosis and management of acute ischemic stroke: speed is critical. Cmaj, 187(12):887–893, 2015

    2015

  15. [15]

    Neural operator: Learning maps between function spaces

    Zongyi Li. Neural operator: Learning maps between function spaces. In 2021 Fall Western Sectional Meeting. AMS, 2021

    2021

  16. [16]

    Multi-scale time-stepping of partial differential equations with transformers

    AmirPouya Hemmasian and Amir Barati Farimani. Multi-scale time-stepping of partial differential equations with transformers. Computer Methods in Applied Mechanics and Engineering, 426:116983, 2024

    2024

  17. [17]

    Predicting clinical outcomes of large vessel occlusion before mechanical thrombectomy using machine learning

    Hidehisa Nishi, Naoya Oishi, Akira Ishii, Isao Ono, Takenori Ogura, Tadashi Sunohara, Hideo Chihara, Ryu Fukumitsu, Masakazu Okawa, Norikazu Yamana, et al. Predicting clinical outcomes of large vessel occlusion before mechanical thrombectomy using machine learning. Stroke, 50(9):2379–2388, 2019

    2019

  18. [18]

    A machine learning approach to investigate the relationship between shape features and numerically predicted risk of ascending aortic aneurysm

    Liang Liang, Minliang Liu, Caitlin Martin, John A Elefteriades, and Wei Sun. A machine learning approach to investigate the relationship between shape features and numerically predicted risk of ascending aortic aneurysm. Biomechanics and modeling in mechanobiol- ogy, 16(5):1519–1533, 2017

    2017

  19. [19]

    Learn- ing mesh-based simulation with graph networks

    Tobias Pfaff, Meire Fortunato, Alvaro Sanchez-Gonzalez, and Peter Battaglia. Learn- ing mesh-based simulation with graph networks. In International conference on learning representations, 2020

    2020

  20. [20]

    Transolver: A Fast Transformer Solver for PDEs on General Geometries

    Haixu Wu, Huakun Luo, Haowen Wang, Jianmin Wang, and Mingsheng Long. Tran- solver: A fast transformer solver for pdes on general geometries (2024). URL https://arxiv. org/abs/2402.02366, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  21. [21]

    Erwin: A tree-based hierarchical transformer for large-scale phys- ical systems.arXiv preprint arXiv:2502.17019,

    Maksim Zhdanov, Max Welling, and Jan-Willem van de Meent. Erwin: A tree-based hierarchical transformer for large-scale physical systems. arXiv preprint arXiv:2502.17019 , 2025

    work page arXiv 2025

  22. [22]

    Occurrence of intracranial large vessel occlusion in consecutive, non-referred pa- tients with acute ischemic stroke

    Debbie Beumer, Maxim JHL Mulder, Ghesrouw Saiedie, Susanne Fonville, Robert J van Oostenbrugge, Wim H van Zwam, Philip J Homburg, Aad van der Lugt, and Diederik WJ Dippel. Occurrence of intracranial large vessel occlusion in consecutive, non-referred pa- tients with acute ischemic stroke. Neurovascular Imaging, 2(1):11, 2016

    2016

  23. [23]

    Variations in the configuration of the circle of willis

    Kanchan Kapoor, Balbir Singh, and Inder Jit Dewan. Variations in the configuration of the circle of willis. Anatomical science international , 83(2):96–106, 2008

    2008

  24. [24]

    Geometry and symmetry of willis’ circle and middle cerebral artery aneurysms development

    Carmelo Lucio Sturiale, Alba Scerrati, Luca Ricciardi, Oriela Rustemi, Anna Maria Au- ricchio, Nicolò Norri, Amedeo Piazza, Fabio Raneri, Alberto Benato, Alessio Albanese, et al. Geometry and symmetry of willis’ circle and middle cerebral artery aneurysms development. Journal of Clinical Medicine , 13(10):2808, 2024. 37

    2024

  25. [25]

    Indications for mechani- cal thrombectomy for acute ischemic stroke: current guidelines and beyond

    Ashutosh P Jadhav, Shashvat M Desai, and Tudor G Jovin. Indications for mechani- cal thrombectomy for acute ischemic stroke: current guidelines and beyond. Neurology, 97(20_Supplement_2):S126–S136, 2021

    2021

  26. [26]

    Fea- sibility of distal mechanical thrombectomy in m3, a3 and p3 segments via a 0.013-inch delivery system: preliminary experience

    Yilmaz Onal, Murat Velioglu, Ugur Demir, Erhan Celikoglu, and Hakki M Karakas. Fea- sibility of distal mechanical thrombectomy in m3, a3 and p3 segments via a 0.013-inch delivery system: preliminary experience. Turk Neurosurg, 30(4):614–620, 2020

    2020

  27. [27]

    The anatomy of the posterior communicating artery as a risk factor for ischemic cerebral infarction

    Don F Schomer, Michael P Marks, Gary K Steinberg, Iain M Johnstone, Derek B Boothroyd, Michael R Ross, Norbert J Pelc, and Dieter R Enzmann. The anatomy of the posterior communicating artery as a risk factor for ischemic cerebral infarction. New England Journal of Medicine , 330(22):1565–1570, 1994

    1994

  28. [28]

    Development of synthetic thrombus models to simulate stroke treatment in a physical neurointerventional training model

    Nadine Wortmann, Thomas Andersek, Helena Guerreiro, Anna A Kyselyova, Andreas M Frölich, Jens Fiehler, and Dieter Krause. Development of synthetic thrombus models to simulate stroke treatment in a physical neurointerventional training model. All Life , 15(1):283–301, 2022

    2022

  29. [29]

    Thrombus composition and efficacy of throm- bolysis and thrombectomy in acute ischemic stroke

    Precious Jolugbo and Robert AS Ariens. Thrombus composition and efficacy of throm- bolysis and thrombectomy in acute ischemic stroke. Stroke, 52(3):1131–1142, 2021

    2021

  30. [30]

    Acute ischemic stroke thrombi have an outer shell that impairs fibrinolysis

    Lucas Di Meglio, Jean-Philippe Desilles, Véronique Ollivier, Mialitiana Solo Nomenjana- hary, Sara Di Meglio, Catherine Deschildre, Stéphane Loyau, Jean-Marc Olivot, Raphaël Blanc, Michel Piotin, et al. Acute ischemic stroke thrombi have an outer shell that impairs fibrinolysis. Neurology, 93(18):e1686–e1698, 2019

    2019

  31. [31]

    Acute ischemic stroke: a review of imag- ing, patient selection, and management in the endovascular era

    Sharath Kumar GG and Chinmay P Nagesh. Acute ischemic stroke: a review of imag- ing, patient selection, and management in the endovascular era. part ii: Patient selection, endovascular thrombectomy, and postprocedure management. Journal of Clinical Inter- ventional Radiology ISVIR , 2(03):169–183, 2018

    2018

  32. [32]

    Recent advances in devices for mechanical thrombectomy

    Raphaël Blanc, Simon Escalard, Humain Baharvadhat, Jean Philippe Desilles, William Boisseau, Robert Fahed, Hocine Redjem, Gabriele Ciccio, Stanislas Smajda, Benjamin Maier, et al. Recent advances in devices for mechanical thrombectomy. Expert review of medical devices, 17(7):697–706, 2020

    2020

  33. [33]

    Penumbra ace

    Penumbra Inc. Penumbra ace. https://www.penumbrainc.com/products/ penumbra-ace/. Accessed: 2025-11-28

    2025

  34. [34]

    React aspiration catheter

    Medtronic. React aspiration catheter. https://www.medtronic.com/en-us/ healthcare-professionals/products/neurological/neurovascular/catheters/ aspiration-neurovascular-catheters/react-catheter.html . Accessed: 2025-11-28

    2025

  35. [35]

    Trevo xp provue retriever

    Stryker. Trevo xp provue retriever. https://www.stryker.com/us/en/neurovascular/ products/trevo--xp-provue-retriever-.html . Accessed: 2025-11-28

    2025

  36. [36]

    Eric device

    Terumo Neurovascular. Eric device. https://www.terumoneuro.com/emea/products/ eric. Accessed: 2025-11-28

    2025

  37. [37]

    An integral equation approach to boundary value problems of classical elastostatics

    Frank J Rizzo. An integral equation approach to boundary value problems of classical elastostatics. Quarterly of applied mathematics , 25(1):83–95, 1967

    1967

  38. [38]

    Explicit algorithms for the nonlinear dynamics of shells

    Ted Belytschko, Jerry I Lin, and Tsay Chen-Shyh. Explicit algorithms for the nonlinear dynamics of shells. Computer methods in applied mechanics and engineering , 42(2):225– 251, 1984. 38

    1984

  39. [39]

    Multiphysics simulations of rocket engine combustion

    Yen-Sen Chen, TH Chou, BR Gu, Jong-Shinn Wu, Bill Wu, YY Lian, and Luke Yang. Multiphysics simulations of rocket engine combustion. Computers & Fluids , 45(1):29–36, 2011

    2011

  40. [40]

    Computational nuclear physics

    Karlheinz Langanke, Joachim A Maruhn, and Steven E Koonin. Computational nuclear physics. Springer, 1991

    1991

  41. [41]

    The role of computational methods in cardiovascular medicine: a narrative review

    Ivan Fumagalli, Stefano Pagani, Christian Vergara, Dilachew A Adebo, Maurizio Del Greco, Antonio Frontera, Giovanni Battista Luciani, Gianluca Pontone, Roberto Scro- fani, Alfio Quarteroni, et al. The role of computational methods in cardiovascular medicine: a narrative review. Translational Pediatrics, 13(1):146, 2024

    2024

  42. [42]

    A review of low-cost ultrasound compatible phantoms

    Leah A Groves, Mohamed Keita, Saidou Talla, Ron Kikinis, Gabor Fichtinger, Parvin Mousavi, and Mamadou Camara. A review of low-cost ultrasound compatible phantoms. IEEE Transactions on Biomedical Engineering , 70(12):3436–3448, 2023

    2023

  43. [43]

    Mechanical characterization of thrombi retrieved with endovascular thrombectomy in patients with acute ischemic stroke

    Nikki Boodt, Philip R W Snouckaert van Schauburg, Hajo M Hund, Behrooz Fereidoon- nezhad, J Patrick McGarry, Ali C Akyildiz, Adriaan CGM Van Es, Simon F De Meyer, Diederik WJ Dippel, Hester F Lingsma, et al. Mechanical characterization of thrombi retrieved with endovascular thrombectomy in patients with acute ischemic stroke. Stroke, 52(8):2510–2517, 2021

    2021

  44. [44]

    Discover- ing symmetry breaking in physical systems with relaxed group convolution, 2024

    Rui Wang, Elyssa Hofgard, Han Gao, Robin Walters, and Tess E Smidt. Discover- ing symmetry breaking in physical systems with relaxed group convolution, 2024. URL https://arxiv. org/abs/2310.02299 , 2023

    work page arXiv 2024

  45. [45]

    Applications of deep learning in molecule gen- eration and molecular property prediction

    W Patrick Walters and Regina Barzilay. Applications of deep learning in molecule gen- eration and molecular property prediction. Accounts of chemical research , 54(2):263–270, 2020

    2020

  46. [46]

    A deep learning approach to estimate stress distribution: a fast and accurate surrogate of finite-element analysis

    Liang Liang, Minliang Liu, Caitlin Martin, and Wei Sun. A deep learning approach to estimate stress distribution: a fast and accurate surrogate of finite-element analysis. Journal of The Royal Society Interface , 15(138):20170844, 2018

    2018

  47. [47]

    Learning the intrinsic dynamics of spatio-temporal processes through latent dynamics net- works

    Francesco Regazzoni, Stefano Pagani, Matteo Salvador, Luca Dede’, and Alfio Quarteroni. Learning the intrinsic dynamics of spatio-temporal processes through latent dynamics net- works. Nature Communications, 15(1):1834, 2024

    2024

  48. [48]

    Learning to accelerate partial differ- ential equations via latent global evolution

    Tailin Wu, Takashi Maruyama, and Jure Leskovec. Learning to accelerate partial differ- ential equations via latent global evolution. Advances in Neural Information Processing Systems, 35:2240–2253, 2022

    2022

  49. [49]

    The illustrated transformer

    Jay Alammar. The illustrated transformer. https://jalammar.github.io/ illustrated-transformer/, 2018. Accessed: 2026-01-31; originally published June 27, 2018

    2018

  50. [50]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems , 30, 2017

    2017

  51. [51]

    Neural message passing for quantum chemistry

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning, pages 1263–1272. Pmlr, 2017. 39

    2017

  52. [52]

    The thrombolysis in myocardial infarction (timi) trial

    TIMI Study Group. The thrombolysis in myocardial infarction (timi) trial. phase i findings. The New England Journal of Medicine , 312(14):932–936, April 4 1985

    1985

  53. [53]

    Language modeling with gated convolutional networks

    Yann N Dauphin, Angela Fan, Michael Auli, and David Grangier. Language modeling with gated convolutional networks. In International conference on machine learning, pages 933–

  54. [54]

    GLU Variants Improve Transformer

    Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202 , 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2002

  55. [55]

    Root mean square layer normalization

    Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in neural information processing systems , 32, 2019

    2019

  56. [56]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention , pages 234–241. Springer, 2015

    2015

  57. [57]

    A new model for learning in graph domains

    Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In Proceedings. 2005 IEEE international joint conference on neural networks, 2005., volume 2, pages 729–734. IEEE, 2005

    2005

  58. [58]

    Generating cerebral vessel trees of acute ischemic stroke patients using conditional set-diffusion

    Thijs P Kuipers, Praneeta R Konduri, Henk Marquering, and Erik J Bekkers. Generating cerebral vessel trees of acute ischemic stroke patients using conditional set-diffusion. In Medical Imaging with Deep Learning , 2024

    2024

  59. [59]

    Nvidia physicsnemo: An open-source framework for physics- based deep learning in science and engineering, February 2023

    PhysicsNeMo Contributors. Nvidia physicsnemo: An open-source framework for physics- based deep learning in science and engineering, February 2023. Accessed: 2026-01-21

    2023

  60. [60]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  61. [61]

    Physics-informed neural net- works: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural net- works: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics , 378:686–707, 2019. 40

    2019