Pith. sign in

REVIEW 4 major objections 5 minor 130 references

A physics-informed U-Net evolves and super-resolves fuzzy dark matter fields by enforcing Schrödinger–Poisson dynamics, matching full simulations from sparse training data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 02:24 UTC pith:HRZ4GEC2

load-bearing objection Useful first full-SP FDM generative SR plus a data-efficient evolution U-Net, but the physics loss only regularizes thin a-windows in a 1 h⁻¹ Mpc box—abstract generalization language outruns Appendix B. the 4 major comments →

arxiv 2607.28604 v1 pith:HRZ4GEC2 submitted 2026-07-30 astro-ph.CO

Cosmo-SPINN: Fuzzy Dark Matter Simulations with Physics-Informed Generative Networks

Ashutosh Kumar Mishra , Emma Tolley , Nicolas Cerardi This is my paper
classification astro-ph.CO
keywords fuzzy dark matterSchrödinger–Poissonphysics-informed neural networksgenerative U-Netsuper-resolutioncosmological simulationswave dark matter
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Fuzzy dark matter is expensive to simulate because the wave equation must be resolved everywhere, so boxes stay small and high-resolution runs are scarce. This paper shows that a generative U-Net can do two jobs at once—evolve fields from initial conditions to a later scale factor, and super-resolve low-resolution cubes—if training adds an explicit loss that penalizes violations of the Schrödinger–Poisson equations. With that physics loss, the evolution model matches target 1 h⁻¹ Mpc simulations using only about 20 percent of the temporal snapshots and still generalizes to new initial conditions. For super-resolution, the same idea yields the first generative models trained on full Schrödinger–Poisson fuzzy-dark-matter runs, recovering missing small-scale power while cutting the usual generative artifacts. The practical claim is that modern generative modeling can stay faithful to wave dark-matter physics and therefore reduce both storage and compute for the simulations cosmologists need.

Core claim

Including an explicit Schrödinger–Poisson residual (or a first-order kick–drift–kick consistency) loss inside a generative U-Net substantially improves both the evolution and the super-resolution of fuzzy-dark-matter fields relative to data-only training, so that a 1 h⁻¹ Mpc box can be reproduced accurately from only ~20 percent of the temporal snapshots and, for the first time, generative super-resolution can be trained directly on full Schrödinger–Poisson simulations.

What carries the argument

Cosmo-SPINN: a residual U-Net that outputs the real and imaginary parts of the wavefunction plus the gravitational potential, trained on the sum of a data MSE and a weighted physics loss built from Schrödinger–Poisson residuals (or a linearized kick–drift–kick step).

Load-bearing premise

That residual physics losses evaluated only on narrow windows of scale factor, together with ordinary field matching, are enough to fix the wavefunction phase and make the method work beyond the small boxes and short intervals used in the tests.

What would settle it

Train the same architecture with and without the physics loss on a held-out set of full Schrödinger–Poisson runs at a later redshift or larger box; if the physics-informed model no longer recovers the target power spectrum to within ~20 percent up to the Jeans scale, or if phase errors grow without bound, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Dense intermediate snapshots of fuzzy-dark-matter simulations need not be stored once an evolution emulator is trained.
  • Low-resolution Schrödinger–Poisson runs can be upscaled to high resolution while remaining consistent with the governing equations.
  • Physics-informed generative models become a practical route to larger-volume or longer-time fuzzy-dark-matter statistics that pure spectral solvers cannot yet reach.
  • Comparing residual versus kick–drift–kick physics losses gives a concrete design choice for future wave-dark-matter emulators.

Where Pith is reading between the lines

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

  • The same residual-plus-physics pattern should transfer to mixed cold-plus-fuzzy dark-matter models where only a fraction of the matter is ultralight.
  • Autoregressive or neural-operator versions of the same loss could remove the need for any high-resolution supervised targets, provided phase degeneracy is controlled.
  • If the method scales, observational forecasts that currently rely on a handful of expensive fuzzy-dark-matter boxes can be rerun with large ensembles of emulated realizations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript introduces Cosmo-SPINN, a U-Net generative framework with a Schrödinger–Poisson (SP) physics-informed loss for fuzzy dark matter in a 1 h⁻¹ Mpc box. It addresses (i) mapping initial conditions at z=127 to fields near a target scale factor and (ii) 32³→128³ super-resolution at fixed epoch. Physics losses are either SP PDE residuals (Eqs. 37–40) or a first-order kick–drift–kick residual (Eqs. 41–42), combined with MSE on normalized (R,I,V). For evolution, adding the SP loss after a data-only warm-up improves residuals and data fidelity when only ~20% of snapshots in a∈[0.149,0.157] are labeled. For SR, models trained on full-SP Jaxion simulations recover power spectra to ~100 h Mpc⁻¹ with mean fractional errors typically ≲20% on held-out realizations, with SP vs KDK ablations and data-only baselines reported.

Significance. If the claims hold in the demonstrated regime, this is a useful step for FDM emulation: full-SP-trained generative SR is new relative to prior FDM GANs that did not use full quantum-pressure simulations, and the evolution results show clear data-efficiency gains from an explicit SP residual versus data-only training (Figs. 2–3, 5–7). Public PyTorch code and systematic SP/KDK ablations with power-spectrum scatter are strengths. The work is relevant to Epoch-of-Reionization-scale FDM studies where SP simulations are expensive. Significance is currently limited by the narrow scale-factor windows, small box, and incomplete phase/small-scale transfer across realizations, so the path to large-volume emulators remains prospective rather than demonstrated.

major comments (4)
  1. [Abstract; §3.2–3.3; Eqs. (37)–(45); Appendix A] Abstract and §1/§6 claim evolution “to an arbitrary cosmological scale factor” and effective generalization across unseen ICs. Training and physics losses are confined to a∈[0.149,0.157] (evolution) and a∈[0.0798,0.0803] (SR) (§3.2; Eqs. 37–41). The evolution map is a residual from z=127 ICs to fields inside that thin window (Eq. 45), with SP/KDK residuals formed only at a and a+Δa drawn from the same window—not along the IC→target path. Appendix A already shows degraded power-spectrum error at interval edges. Please revise abstract/conclusion language to the demonstrated Δa≲0.01 regime, or add experiments that sample physics residuals and targets over a substantially broader a range.
  2. [Abstract; Appendix B; Fig. 9] Appendix B states that multi-realization evolution recovers large-scale morphology but fails on high-k modes, and that “the physics-informed loss alone admits multiple solutions that satisfy the governing equations but differ in phase.” This undercuts the Abstract claim that the framework “generalizes effectively across previously unseen realizations.” The single-realization 20%-snapshot result (Fig. 4) should not be conflated with cross-IC generalization. Either qualify the generalization claim to large-scale morphology only, or strengthen phase constraints (e.g., explicit phase/velocity losses, denser temporal supervision) and report quantitative high-k metrics for multi-IC evolution comparable to Fig. 7.
  3. [§4.2.2; Figs. 5–7; §6] For SR, ensemble statistics (Fig. 7) show the SP-residual model has larger high-k bias and realization scatter than KDK, while selected examples (Figs. 5–6) favor SP at high-k. §4.2.2 and §6 still lean toward SP as improving small-scale fidelity. Please reconcile example-level vs ensemble conclusions explicitly (e.g., preferred loss by metric and k-range), and state which checkpoint/selection rule (min physics test loss) is recommended for downstream use so the central SR claim is not overstated.
  4. [§5.1; §3.1] §5.1 acknowledges that particle-based 2LPT/CIC ICs are suboptimal for field-based FDM (mesh error; CDM-suited velocities) yet all results use this setup. Given that phase learning is already identified as the weak point (Appendix B), the IC choice is load-bearing for claimed physical consistency. At minimum, quantify sensitivity (e.g., one comparison with field-consistent ICs as in Luu et al. 2025) or clearly limit claims to “standard particle-initialized SP suites” rather than general FDM dynamics.
minor comments (5)
  1. [§2.1–2.2] Equation numbering is inconsistent or duplicated in places (e.g., Poisson/mean density and timestep constraints reuse numbers such as (16), (34) in §2). Please renumber uniquely through the manuscript.
  2. [§4.1; Fig. 2] Fig. 2 caption and §4.1 refer to a∈[0.0798,0.0803] when discussing the evolution model’s snapshot count; evolution training is a∈[0.149,0.157] (§3.2). Correct the interval in the evolution discussion.
  3. [Title; §2] Title and running text alternate “F uzzy” / “Schr¨ odinger” spacing artifacts from LaTeX; clean typography throughout.
  4. [§3.4; §4.2] State explicitly whether test realizations in Figs. 5–6 are drawn from the 20% held-out IC set and whether any hyperparameter (α_max, λ_w, residual scale 0.1) was tuned on test data.
  5. [§1] Cite and briefly contrast related physics-informed or operator-learning cosmology emulators beyond the FDM GAN of Sipp et al. (2023) to clarify novelty boundaries.

Circularity Check

0 steps flagged

No significant circularity: physics loss is residual regularization against independent Jaxion targets, not a self-defining prediction.

full rationale

The paper’s load-bearing chain is: (i) standard SP/KDK dynamics as known PDEs; (ii) independent pseudo-spectral ground truth from Jaxion; (iii) a U-Net trained with MSE to those targets plus residual (or linearized KDK) penalties on network outputs; (iv) evaluation on held-out scale factors or realizations, with an explicit data-only baseline. The physics term L_physics = ⟨R_R²+R_I²+R_V²⟩ or ∥ψ_θ(a+Δa)−F_Δa[ψ_θ(a)]∥² enforces consistency of the generator with the governing equations; it does not define the predicted fields as equal to the training labels by construction, nor is any cosmological parameter fitted and then rediscovered as a ‘prediction.’ Self-citation to SPINN (Mishra & Tolley 2025) is methodological background that PINNs can handle SP collapse, not a uniqueness theorem or ansatz that forces the reported spectra or the 20%-data claim. Narrow a-windows, phase degeneracy (Appendix B), and limited transfer are scope/correctness issues, not circular reductions. The derivation is therefore self-contained against external simulation benchmarks.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The work sits on standard SP cosmology plus ordinary deep-learning inductive biases. Load-bearing choices are the small-box full-SP training distribution, particle-based ICs, hand-set loss weights/schedules, and the claim that PDE/KDK residuals plus MSE on (R,I,V) adequately regularize generative maps. No new physical entity is postulated; free parameters are training knobs and fixed simulation settings, not cosmological inferences.

free parameters (5)
  • physics loss ramp α_max and N_ramp (evolution) = α_max=0.02, N_ramp=1000
    α_max=0.02 over 1000 epochs is chosen empirically so SP loss does not overwhelm data fidelity; larger values hurt training reconstruction (§4.1).
  • adaptive gradient-balance weight λ_w = adaptive (equal gradient magnitudes)
    Set so data and physics losses contribute equally in gradient space; not derived from first principles (§3.3).
  • evolution residual scale factor 0.1 = 0.1 (then 1.0 in multi-realization)
    Output formed as ψ(a0)+0.1δ in single-realization evolution; later removed for multi-realization runs after empirical excess power (§3.4, App. B).
  • FDM particle mass and box cosmology = m=2.5e-22 eV, L=1 h^{-1} Mpc
    Fixed simulation settings defining the training distribution (m=2.5×10^{-22} eV, L=1 h^{-1} Mpc, Ωm=0.27, etc.), not inferred by the network (§3.1–3.2).
  • U-Net width/depth, Fourier feature K=8, c=64, Adam lr=1e-4, epoch budgets = K=8, c=64, lr=1e-4; evo 250+2750 ep; SR 20+60 ep
    Standard architecture/optimization hyperparameters that control capacity and convergence; results depend on these choices (§3.4).
axioms (5)
  • domain assumption Non-relativistic fuzzy dark matter in an expanding FLRW background is governed by the Schrödinger–Poisson system (Eqs. 4–6 / 31–33).
    Entire physics loss and simulation suite assume this classical-field SP description.
  • domain assumption Second-order unitary split-step pseudo-spectral KDK integration of SP provides trustworthy ground-truth targets when timesteps satisfy the stated CFL-like bounds.
    Supervised labels and KDK physics loss both inherit accuracy limits of Jaxion’s scheme (§2.2).
  • domain assumption Particle 2LPT + CIC density and velocity-derived phase are acceptable FDM initial conditions for evaluating the ML method.
    Authors note this is conventional but not ideal for field-based FDM; still used for all results (§3.1, §5.1).
  • ad hoc to paper MSE on normalized (R,I,V) plus SP residual or first-order KDK residual is a sufficient training objective for physically consistent generative maps (no explicit mass, phase, or BC losses).
    Stated that mass conservation and periodic BCs are expected to emerge when optimized (§3.3); Appendix B shows phase degeneracy remains.
  • standard math Standard facts of multivariable calculus, FFTs for spectral derivatives, and backprop through the U-Net.
    Used for residual assembly and optimization; not controversial.
invented entities (1)
  • Cosmo-SPINN (physics-informed generative U-Net framework) no independent evidence
    purpose: Name the combined evolution and super-resolution architecture with SP/KDK losses for FDM cubes.
    Methodological construct, not a new physical object; independent evidence is empirical performance on held-out sims, not an external detection channel.

pith-pipeline@v1.2.0-daily-grok45 · 26203 in / 3953 out tokens · 74595 ms · 2026-07-31T02:24:48.612634+00:00 · methodology

0 comments
read the original abstract

Generative machine learning models have recently emerged as powerful tools for producing cosmological simulations. However, many existing emulators do not explicitly enforce the underlying physical dynamics governing cosmological evolution, often leading to artifacts and poor adherence to the evolution equations. In this work, we present a physics-informed generative U-Net framework for fuzzy dark matter (FDM) that addresses two complementary tasks: (i) the evolution of cosmological fields from initial conditions to an arbitrary cosmological scale factor and (ii) the super-resolution of FDM simulations at a specified cosmological scale factor. Our model incorporates a physics-informed loss function that explicitly enforces consistency with the underlying Schr\"odinger-Poisson (SP) dynamics during training. For the evolution task, we find that the inclusion of this physics-based loss significantly improves the quality of the predicted simulations, even when only a small amount of training data is available. Using only 20% of the training data, the model accurately reproduces the target simulations in a 1 $h^{-1}$ Mpc box. Furthermore, the framework generalizes effectively across previously unseen realizations of the initial conditions. For the super-resolution task, we present, for the first time, a generative super-resolution model trained on FDM simulations obtained by solving the full SP equations, considering both single and multiple realizations of the initial conditions and analyzing the role of the physics-informed loss in each case. Our approach enables modern generative modeling of cosmological simulations while maintaining physical consistency and substantially reducing generative artifacts.

Figures

Figures reproduced from arXiv: 2607.28604 by Ashutosh Kumar Mishra, Emma Tolley, Nicolas Cerardi.

Figure 1
Figure 1. Figure 1: Schematic of the Cosmo-SPINN framework applied to two tasks: (a) (left) Evolution Model, which predicts R,I,V (Re(Ψ) = R,Im(Ψ) = I) at a target scale factor a from the initial conditions, and (b) (right) Super-Resolution Model, which reconstructs high-resolution fields from low-resolution inputs at the same scale factor. In both cases, the outputs serve as approximate solutions to the SP system and are con… view at source ↗
Figure 2
Figure 2. Figure 2: Training and test data losses (left) and normalized SP loss (right) for the evolution model. Training is performed using only the data loss for the first 250 epochs, after which the SP loss is added to the objective (vertical dotted line) and optimization continues for the remaining 2750 epochs. Solid and dashed curves denote the training and test losses (shown losses are smoothed with an exponential movin… view at source ↗
Figure 3
Figure 3. Figure 3: Evolution of the normalized physics loss for the super-resolution model. Left: physics loss based on the SP residual. Right: physics loss based on a first-order KDK approximation. Physics-informed training significantly reduces the losses on both the training and test sets. Solid and dashed lines denote the training and test sets, respectively (shown losses are smoothed with an exponential moving average w… view at source ↗
Figure 4
Figure 4. Figure 4: Evolution Model: Comparison of the ground-truth and predicted projected fields ⟨|R|⟩z, ⟨|I|⟩z, and density log10⟨ρ⟩z at scale factor a=0.156, obtained by averaging the corresponding 643 fields along the z-axis. Columns show the initial condition, reference solution, neural-network prediction, and residual (prediction-truth). The predicted wavefunction components and density closely match the reference solu… view at source ↗
Figure 5
Figure 5. Figure 5: Examples of super-resolution density fields generated by the model trained with the SP physics loss for the unseen test realizations. The corresponding scale factors are (from top to bottom) a ≈ 0.08039, 0.08042, 0.08043, and 0.08039. The input low-resolution fields (323 , LR), generated super-resolved fields (1283 , SR), and target high-resolution fields (1283 , HR) are shown as density projections along … view at source ↗
Figure 6
Figure 6. Figure 6: Examples of super-resolution density fields generated by the model trained with the first-order KDK approx. for the unseen test realizations. The corresponding scale factors are (from top to bottom) a ≈ 0.08043, 0.08039, 0.08042, and 0.08039. The input low-resolution fields (323 , LR), generated super-resolved fields (1283 , SR), and target high-resolution fields (1283 , HR) are shown as density projection… view at source ↗
Figure 7
Figure 7. Figure 7: Mean fractional power-spectrum error, (PSR/PHR −1), across the test realizations at the mid-scale factor a ≈ 0.08041. The shaded regions denote the corresponding 1σ scatter across realizations. Left: SR model trained with the SP-residual based physics loss. Right: SR model trained with the first-order KDK-based physics loss. exhibits a larger test loss than the KDK model, it more accurately reproduces the … view at source ↗
Figure 8
Figure 8. Figure 8: Mean absolute relative error in the predicted power spectrum, ⟨|Ppred/Ptrue −1|⟩ as a function of the scale factor a. The shaded vertical bands indicate the boundaries of the full data interval. The prediction error is minimized within the central region and increases toward the edges of the sampled range. A. EVOLUTION MODEL STABILITY ACROSS REDSHIFTS Here we test the stability of evolution model across re… view at source ↗
Figure 9
Figure 9. Figure 9: Examples of density fields generated by evolution model the unseen test realizations. The generated fields (643 , Gen), and target density fields (643 , Real) are shown as density projections along the z-axis on a logarithmic scale. Insets show zoomed-in regions. The right panels compare the dimensionless power spectra k 3P(k) of the Gen, and Real fields, while the lower panels show the fractional error (P… view at source ↗

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

130 extracted references · 17 canonical work pages · 7 internal anchors

  1. [1]

    2025 , eprint=

    QCPINN: Quantum-Classical Physics-Informed Neural Networks for Solving PDEs , author=. 2025 , eprint=

    2025

  2. [2]

    2025 , eprint=

    Quantum physics informed neural networks for multi-variable partial differential equations , author=. 2025 , eprint=

    2025

  3. [3]

    Workshop Proceedings of the 53rd International Conference on Parallel Processing , pages =

    Hegde, Pratibha Raghupati and Markidis, Stefano , title =. Workshop Proceedings of the 53rd International Conference on Parallel Processing , pages =. 2024 , isbn =. doi:10.1145/3677333.3678272 , abstract =

    arXiv 2024

  4. [4]

    Entropy , VOLUME =

    Trahan, Corey and Loveland, Mark and Dent, Samuel , TITLE =. Entropy , VOLUME =. 2024 , NUMBER =

    2024

  5. [5]

    Variational quantum simulation using non-Gaussian continuous-variable systems , author =. Phys. Rev. Res. , volume =. 2024 , month =. doi:10.1103/PhysRevResearch.6.043212 , url =

    work page doi:10.1103/physrevresearch.6.043212 2024

  6. [6]

    Quantum error mitigation for rotation-symmetric bosonic codes with symmetry expansion , author =. Phys. Rev. A , volume =. 2025 , month =. doi:10.1103/PhysRevA.111.062402 , url =

    work page doi:10.1103/physreva.111.062402 2025

  7. [7]

    Bosonic quantum error correction codes in superconducting quantum circuits , journal =

    Weizhou Cai and Yuwei Ma and Weiting Wang and Chang-Ling Zou and Luyan Sun , keywords =. Bosonic quantum error correction codes in superconducting quantum circuits , journal =. 2021 , issn =. doi:https://doi.org/10.1016/j.fmre.2020.12.006 , url =

    doi:10.1016/j.fmre.2020.12.006 2021

  8. [8]

    PRX Quantum , volume =

    Quantum Error Correction with the Gottesman-Kitaev-Preskill Code , author =. PRX Quantum , volume =. 2021 , month =. doi:10.1103/PRXQuantum.2.020101 , url =

    work page doi:10.1103/prxquantum.2.020101 2021

  9. [9]

    Continuous-variable quantum neural networks , author =. Phys. Rev. Res. , volume =. 2019 , month =. doi:10.1103/PhysRevResearch.1.033063 , url =

    doi:10.1103/physrevresearch.1.033063 2019

  10. [10]

    Gaussian quantum information , author =. Rev. Mod. Phys. , volume =. 2012 , month =. doi:10.1103/RevModPhys.84.621 , url =

    doi:10.1103/revmodphys.84.621 2012

  11. [11]

    Inequivalence between the Schr\"odinger equation and the Madelung hydrodynamic equations , author =. Phys. Rev. A , volume =. 1994 , month =. doi:10.1103/PhysRevA.49.1613 , url =

    work page doi:10.1103/physreva.49.1613 1994

  12. [12]

    Physical Review Research , keywords =

    From Vlasov-Poisson to Schr \"o dinger-Poisson: Dark matter simulation with a quantum variational time evolution algorithm. Physical Review Research , keywords =. doi:10.1103/PhysRevResearch.6.013282 , archivePrefix =. 2307.06032 , primaryClass =

    Pith/arXiv arXiv

  13. [13]

    , keywords =

    Toward Cosmological Simulations of Dark Matter on Quantum Computers. , keywords =. doi:10.3847/1538-4357/abe6ac , archivePrefix =. 2101.05821 , primaryClass =

    Pith/arXiv arXiv

  14. [14]

    Peccei, R. D. and Quinn, Helen R. , journal =. 1977 , month =. doi:10.1103/PhysRevLett.38.1440 , url =

    doi:10.1103/physrevlett.38.1440 1977

  15. [15]

    arXiv e-prints , keywords =

    Sub-Sequential Physics-Informed Learning with State Space Model. arXiv e-prints , keywords =. doi:10.48550/arXiv.2502.00318 , archivePrefix =. 2502.00318 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.00318

  16. [16]

    arXiv e-prints , keywords =

    Physics-Informed Kolmogorov-Arnold Networks for Power System Dynamics. arXiv e-prints , keywords =. doi:10.48550/arXiv.2408.06650 , archivePrefix =. 2408.06650 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2408.06650

  17. [17]

    SIAM Journal on Scientific Computing , keywords =

    Understanding and Mitigating Gradient Flow Pathologies in Physics-Informed Neural Networks. SIAM Journal on Scientific Computing , keywords =. doi:10.1137/20M1318043 , archivePrefix =. 2001.04536 , primaryClass =

    Pith/arXiv arXiv 2001

  18. [18]

    arXiv e-prints , keywords =

    When and why PINNs fail to train: A neural tangent kernel perspective. arXiv e-prints , keywords =. doi:10.48550/arXiv.2007.14527 , archivePrefix =. 2007.14527 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2007.14527 2007

  19. [19]

    Pearlmutter and Alexey Andreyevich Radul and Jeffrey Mark Siskind , title =

    Atilim Gunes Baydin and Barak A. Pearlmutter and Alexey Andreyevich Radul and Jeffrey Mark Siskind , title =. Journal of Machine Learning Research , year =

  20. [20]

    Deep learning

    Yann Lecun and Yoshua Bengio and Geoffrey Hinton. Deep learning. Nature. 2015. doi:10.1038/nature14539

    doi:10.1038/nature14539 2015

  21. [21]

    Journal of Heat Transfer , volume =

    Cai, Shengze and Wang, Zhicheng and Wang, Sifan and Perdikaris, Paris and Karniadakis, George Em , title =. Journal of Heat Transfer , volume =. 2021 , month =. doi:10.1115/1.4050542 , url =

    doi:10.1115/1.4050542 2021

  22. [22]

    NSFnets (Navier-Stokes flow nets): Physics-informed neural networks for the incompressible Navier-Stokes equations , journal =

    Xiaowei Jin and Shengze Cai and Hui Li and George Em Karniadakis , keywords =. NSFnets (Navier-Stokes flow nets): Physics-informed neural networks for the incompressible Navier-Stokes equations , journal =. 2021 , issn =. doi:https://doi.org/10.1016/j.jcp.2020.109951 , url =

    arXiv 2021

  23. [23]

    arXiv e-prints , keywords =

    Hidden Fluid Mechanics: A Navier-Stokes Informed Deep Learning Framework for Assimilating Flow Visualization Data. arXiv e-prints , keywords =. doi:10.48550/arXiv.1808.04327 , archivePrefix =. 1808.04327 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.1808.04327

  24. [24]

    , keywords =

    Towards super-resolution simulations of the fuzzy dark matter cosmological model. , keywords =. doi:10.1093/mnras/stad2341 , archivePrefix =. 2210.12907 , primaryClass =

    Pith/arXiv arXiv

  25. [25]

    , keywords =

    Super-resolution emulator of cosmological simulations using deep physical models. , keywords =. doi:10.1093/mnras/staa1428 , archivePrefix =. 2001.05519 , primaryClass =

    Pith/arXiv arXiv 2001

  26. [26]

    Proceedings of the National Academy of Science , keywords =

    AI-assisted superresolution cosmological simulations. Proceedings of the National Academy of Science , keywords =. doi:10.1073/pnas.2022038118 , archivePrefix =. 2010.06608 , primaryClass =

    Pith/arXiv arXiv 2010

  27. [27]

    , keywords =

    SPINN: Advancing Cosmological Simulations of Fuzzy Dark Matter with Physics Informed Neural Networks. , keywords =. doi:10.3847/1538-4357/ade43e , archivePrefix =. 2506.02957 , primaryClass =

    Pith/arXiv arXiv

  28. [28]

    arXiv e-prints , keywords =

    Updated bounds on ultra-light dark matter from the tiniest galaxies. arXiv e-prints , keywords =. doi:10.48550/arXiv.2509.02781 , archivePrefix =. 2509.02781 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2509.02781

  29. [29]

    , keywords =

    Warm and fuzzy dark matter: Free streaming of wave dark matter. , keywords =. doi:10.1103/PhysRevD.111.023535 , archivePrefix =. 2406.12970 , primaryClass =

    Pith/arXiv arXiv

  30. [30]

    Universe , keywords =

    Warm Dark Matter in Simulations. Universe , keywords =. doi:10.3390/universe8020076 , adsurl =

    work page doi:10.3390/universe8020076

  31. [31]

    Fuzzy Cold Dark Matter: The Wave Properties of Ultralight Particles , author =. Phys. Rev. Lett. , volume =. 2000 , month =. doi:10.1103/PhysRevLett.85.1158 , url =

    doi:10.1103/physrevlett.85.1158 2000

  32. [32]

    Dynamical evidence of a dark solitonic core of 109M⊙ in the milky way , journal =

    Ivan. Dynamical evidence of a dark solitonic core of 109M⊙ in the milky way , journal =. 2020 , issn =. doi:https://doi.org/10.1016/j.dark.2020.100503 , url =

    arXiv 2020

  33. [33]

    Limits on primordial black holes from M87 , author =. Phys. Rev. D , volume =. 2022 , month =. doi:10.1103/PhysRevD.105.063506 , url =

    work page doi:10.1103/physrevd.105.063506 2022

  34. [34]

    Nature Astronomy , keywords =

    Baryonic solutions and challenges for cosmological models of dwarf galaxies. Nature Astronomy , keywords =. doi:10.1038/s41550-022-01689-w , archivePrefix =. 2206.05295 , primaryClass =

    Pith/arXiv arXiv

  35. [35]

    Physics informed neural networks for simulating radiative transfer , journal =

    Siddhartha Mishra and Roberto Molinaro , keywords =. Physics informed neural networks for simulating radiative transfer , journal =. 2021 , issn =. doi:https://doi.org/10.1016/j.jqsrt.2021.107705 , url =

    arXiv 2021

  36. [36]

    Physics of Fluids , volume =

    Ren, Xiang and Hu, Peng and Su, Hua and Zhang, Feizhou and Yu, Huahua , title =. Physics of Fluids , volume =. 2024 , month =. doi:10.1063/5.0200384 , url =

    work page doi:10.1063/5.0200384 2024

  37. [37]

    , keywords =

    Large-scale structure of the Universe and cosmological perturbation theory. , keywords =. doi:10.1016/S0370-1573(02)00135-7 , archivePrefix =. astro-ph/0112551 , primaryClass =

    Pith/arXiv arXiv

  38. [38]

    Computer Methods in Applied Mechanics and Engineering , keywords =

    Physics-informed neural networks for high-speed flows. Computer Methods in Applied Mechanics and Engineering , keywords =. doi:10.1016/j.cma.2019.112789 , adsurl =

    arXiv 2019

  39. [39]

    Zhao, D. H. and Mo, H. J. and Jing, Y. P. and Börner, G. , title =. Monthly Notices of the Royal Astronomical Society , volume =. 2003 , month =. doi:10.1046/j.1365-8711.2003.06135.x , url =

    arXiv 2003

  40. [40]

    , keywords =

    Core condensation in heavy halos: a two-stage theory for galaxy formation and clustering. , keywords =. doi:10.1093/mnras/183.3.341 , adsurl =

    doi:10.1093/mnras/183.3.341

  41. [41]

    , keywords =

    A Direct Empirical Proof of the Existence of Dark Matter. , keywords =. doi:10.1086/508162 , archivePrefix =. astro-ph/0608407 , primaryClass =

    Pith/arXiv arXiv

  42. [42]

    , keywords =

    Rotational properties of 21 SC galaxies with a large range of luminosities and radii, from NGC 4605 (R=4kpc) to UGC 2885 (R=122kpc). , keywords =. doi:10.1086/158003 , adsurl =

    doi:10.1086/158003

  43. [43]

    arXiv e-prints , keywords =

    ProPINN: Demystifying Propagation Failures in Physics-Informed Neural Networks. arXiv e-prints , keywords =. doi:10.48550/arXiv.2502.00803 , archivePrefix =. 2502.00803 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2502.00803

  44. [44]

    PinnDE: Physics-Informed Neural Networks for Solving Differential Equations

    PinnDE: Physics-Informed Neural Networks for Solving Differential Equations. arXiv e-prints , keywords =. doi:10.48550/arXiv.2408.10011 , archivePrefix =. 2408.10011 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2408.10011

  45. [45]

    arXiv e-prints , keywords =

    A hands-on introduction to Physics-Informed Neural Networks for solving partial differential equations with benchmark tests taken from astrophysics and plasma physics. arXiv e-prints , keywords =. doi:10.48550/arXiv.2403.00599 , archivePrefix =. 2403.00599 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2403.00599

  46. [46]

    Journal of Computational Physics , keywords =

    Physics-informed neural networks for inverse problems in supersonic flows. Journal of Computational Physics , keywords =. doi:10.1016/j.jcp.2022.111402 , archivePrefix =. 2202.11821 , primaryClass =

    arXiv 2022

  47. [47]

    DAE-PINN: A Physics-Informed Neural Network Model for Simulating Differential-Algebraic Equations with Application to Power Networks

    DAE-PINN: A Physics-Informed Neural Network Model for Simulating Differential-Algebraic Equations with Application to Power Networks. arXiv e-prints , keywords =. doi:10.48550/arXiv.2109.04304 , archivePrefix =. 2109.04304 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2109.04304

  48. [48]

    arXiv e-prints , keywords =

    PyTorch: An Imperative Style, High-Performance Deep Learning Library. arXiv e-prints , keywords =. doi:10.48550/arXiv.1912.01703 , archivePrefix =. 1912.01703 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.1912.01703 1912

  49. [49]

    , keywords =

    Observational Evidence from Supernovae for an Accelerating Universe and a Cosmological Constant. , keywords =. doi:10.1086/300499 , archivePrefix =. astro-ph/9805201 , primaryClass =

    Pith/arXiv arXiv

  50. [50]

    2015 , publisher=

    Neural Networks and Deep Learning , author=. 2015 , publisher=

    2015

  51. [51]

    arXiv e-prints , keywords =

    Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. arXiv e-prints , keywords =. doi:10.48550/arXiv.2006.10739 , archivePrefix =. 2006.10739 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2006.10739 2006

  52. [52]

    Computer Methods in Applied Mechanics and Engineering , keywords =

    Multilevel domain decomposition-based architectures for physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering , keywords =. doi:10.1016/j.cma.2024.117116 , archivePrefix =. 2306.05486 , primaryClass =

    arXiv 2024

  53. [53]

    arXiv e-prints , keywords =

    Finite Basis Physics-Informed Neural Networks (FBPINNs): a scalable domain decomposition approach for solving differential equations. arXiv e-prints , keywords =. doi:10.48550/arXiv.2107.07871 , archivePrefix =. 2107.07871 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2107.07871

  54. [54]

    arXiv e-prints , keywords =

    On the Spectral Bias of Neural Networks. arXiv e-prints , keywords =. doi:10.48550/arXiv.1806.08734 , archivePrefix =. 1806.08734 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.1806.08734

  55. [55]

    A Tutorial on the Use of Physics-Informed Neural Networks to Compute the Spectrum of Quantum Systems

    A Tutorial on the Use of Physics-Informed Neural Networks to Compute the Spectrum of Quantum Systems. arXiv e-prints , keywords =. doi:10.48550/arXiv.2407.20669 , archivePrefix =. 2407.20669 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2407.20669

  56. [56]

    Journal of Computational Physics , keywords =

    B-PINNs: Bayesian physics-informed neural networks for forward and inverse PDE problems with noisy data. Journal of Computational Physics , keywords =. doi:10.1016/j.jcp.2020.109913 , archivePrefix =. 2003.06097 , primaryClass =

    arXiv 2020

  57. [57]

    Computer Methods in Applied Mechanics and Engineering , keywords =

    Surrogate modeling for fluid flows based on physics-constrained deep learning without simulation data. Computer Methods in Applied Mechanics and Engineering , keywords =. doi:10.1016/j.cma.2019.112732 , archivePrefix =. 1906.02382 , primaryClass =

    arXiv 2019

  58. [58]

    Journal of Computational Physics , keywords =

    Physics-constrained deep learning for high-dimensional surrogate modeling and uncertainty quantification without labeled data. Journal of Computational Physics , keywords =. doi:10.1016/j.jcp.2019.05.024 , archivePrefix =. 1901.06314 , primaryClass =

    Pith/arXiv arXiv 2019

  59. [59]

    and Geiger, Bernhard C

    Hoffer, Johannes G. and Geiger, Bernhard C. and Ofner, Patrick and Kern, Roman , TITLE =. Applied Sciences , VOLUME =. 2021 , NUMBER =

    2021

  60. [60]

    DPM: A Novel Training Method for Physics-Informed Neural Networks in Extrapolation

    DPM: A Novel Training Method for Physics-Informed Neural Networks in Extrapolation. arXiv e-prints , keywords =. doi:10.48550/arXiv.2012.02681 , archivePrefix =. 2012.02681 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2012.02681 2012

  61. [61]

    Numerical Optimization

    Line Search Methods. Numerical Optimization. 2006. doi:10.1007/978-0-387-40065-5_3

    work page doi:10.1007/978-0-387-40065-5_3 2006

  62. [62]

    SIAM Review , volume =

    Wolfe, Philip , title =. SIAM Review , volume =. 1971 , doi =. https://doi.org/10.1137/1013035 , abstract =

    doi:10.1137/1013035 1971

  63. [63]

    SIAM Review , volume =

    Wolfe, Philip , title =. SIAM Review , volume =. 1969 , doi =. https://doi.org/10.1137/1011036 , abstract =

    doi:10.1137/1011036 1969

  64. [64]

    arXiv e-prints , keywords =

    PINNsFormer: A Transformer-Based Framework For Physics-Informed Neural Networks. arXiv e-prints , keywords =. doi:10.48550/arXiv.2307.11833 , archivePrefix =. 2307.11833 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.11833

  65. [65]

    Learning in Sinusoidal Spaces with Physics-Informed Neural Networks

    Learning in Sinusoidal Spaces with Physics-Informed Neural Networks. arXiv e-prints , keywords =. doi:10.48550/arXiv.2109.09338 , archivePrefix =. 2109.09338 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2109.09338

  66. [66]

    Updating Quasi-Newton Matrices with Limited Storage , urldate =

    Jorge Nocedal , journal =. Updating Quasi-Newton Matrices with Limited Storage , urldate =

  67. [67]

    arXiv e-prints , keywords =

    Challenges in Training PINNs: A Loss Landscape Perspective. arXiv e-prints , keywords =. doi:10.48550/arXiv.2402.01868 , archivePrefix =. 2402.01868 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.01868

  68. [68]

    arXiv e-prints , keywords =

    Adam: A Method for Stochastic Optimization. arXiv e-prints , keywords =. doi:10.48550/arXiv.1412.6980 , archivePrefix =. 1412.6980 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980

  69. [69]

    arXiv e-prints , keywords =

    Implicit Neural Representations with Periodic Activation Functions. arXiv e-prints , keywords =. doi:10.48550/arXiv.2006.09661 , archivePrefix =. 2006.09661 , primaryClass =

    Pith/arXiv arXiv doi:10.48550/arxiv.2006.09661 2006

  70. [70]

    Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics , pages =

    Understanding the difficulty of training deep feedforward neural networks , author =. Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics , pages =. 2010 , editor =

    2010

  71. [71]

    , keywords =

    The unexpected diversity of dwarf galaxy rotation curves. , keywords =. doi:10.1093/mnras/stv1504 , archivePrefix =. 1504.01437 , primaryClass =

    Pith/arXiv arXiv

  72. [72]

    A Hybrid Scheme for Fuzzy Dark Matter Simulations Combining the Schr\"odinger and Hamilton-Jacobi-Madelung Equations

    A Hybrid Scheme for Fuzzy Dark Matter Simulations Combining the Schr \"o dinger and Hamilton-Jacobi-Madelung Equations. arXiv e-prints , keywords =. doi:10.48550/arXiv.2411.17288 , archivePrefix =. 2411.17288 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2411.17288

  73. [73]

    Frontiers in Astronomy and Space Sciences , keywords =

    Cosmological Simulation for Fuzzy Dark Matter Model. Frontiers in Astronomy and Space Sciences , keywords =. doi:10.3389/fspas.2018.00048 , archivePrefix =. 1809.09848 , primaryClass =

    arXiv 2018

  74. [74]

    , keywords =

    Numerical and perturbative computations of the fuzzy dark matter model. , keywords =. doi:10.1103/PhysRevD.99.063509 , archivePrefix =. 1810.01915 , primaryClass =

    Pith/arXiv arXiv

  75. [75]

    , keywords =

    AX-GADGET: a new code for cosmological simulations of Fuzzy Dark Matter and Axion models. , keywords =. doi:10.1093/mnras/sty1224 , archivePrefix =. 1801.08144 , primaryClass =

    Pith/arXiv arXiv

  76. [76]

    , keywords =

    Contrasting Galaxy Formation from Quantum Wave Dark Matter, DM, with CDM, using Planck and Hubble Data. , keywords =. doi:10.3847/0004-637X/818/1/89 , archivePrefix =. 1508.04621 , primaryClass =

    Pith/arXiv arXiv

  77. [77]

    Journal of Computational Physics , keywords =

    Solution of the Schr \"o dinger Equation by a Spectral Method. Journal of Computational Physics , keywords =. doi:10.1016/0021-9991(82)90091-2 , adsurl =

    doi:10.1016/0021-9991(82)90091-2

  78. [78]

    , keywords =

    A search for ultralight axions using precision cosmological data. , keywords =. doi:10.1103/PhysRevD.91.103512 , archivePrefix =. 1410.2896 , primaryClass =

    Pith/arXiv arXiv

  79. [79]

    Mocz, Philip , doi =

  80. [80]

    and Nelson, Dylan and Pillepich, Annalisa , title =

    Thiele, Leander and Villaescusa-Navarro, Francisco and Spergel, David N. and Nelson, Dylan and Pillepich, Annalisa , title =. The Astrophysical Journal , abstract =. 2020 , month =. doi:10.3847/1538-4357/abb80f , url =

    work page doi:10.3847/1538-4357/abb80f 2020

Showing first 80 references.