Physics Informed Neural Networks for Simulating Radiative Transfer

Roberto Molinaro; Siddhartha Mishra

REVIEW 3 major objections 5 minor 1 cited by

Physics Informed Neural Networks for Simulating Radiative Transfer

T0 review · 3 major / 5 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read A neural network trained on the residual of the radiative transfer equation approximates the solution accurately, with the true error controlled by the training error and only logarithmic dimension dependence.

desk verdict A genuine but conditionally supported extension of PINNs to radiative transfer: the numerics are suggestive, the steady-state theory does not cover the paper's own steady experiments, and the K=2 'beats diffusion' claim lacks a ground truth. read the letter →

arxiv 2009.13291 v3 pith:NM2G2GBB submitted 2020-09-25 cs.LG stat.ML

Siddhartha Mishra , Roberto Molinaro This is my paper · ORCID

classification cs.LGstat.ML MSC 68T0765C0565D30 PACS 44.40.+a

keywords radiativetransferphysics-informedneuralnetworksgeneralizationerrorquasi-MonteCarloSobolsequencesinverseproblemspartialintegro-differentialequationsdeeplearningforPDEs

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that physics-informed neural networks (PINNs) — neural networks trained to minimize the residual of the radiative transfer equation at collocation points — form a practical simulator for radiative transfer, a high-dimensional linear partial integro-differential equation that underlies astrophysics, climate modeling, and nuclear engineering. The central theoretical result is a generalization error bound: the error between the trained network and the true radiative intensity is controlled by the training errors plus quadrature terms that decay with the number of training points, with only a logarithmic dependence on dimension. In experiments ranging from one-dimensional stationary slabs to time-dependent, polychromatic three-dimensional problems, the trained networks match reference solutions with low error. A small modification of the algorithm also recovers the absorption coefficient from measurements of the incident radiation.

What carries the argument

The central object is the PINN residual loss: the neural-network output $u_\theta$ is substituted into the radiative transfer equation, and the squared mismatches of the interior residual, the inflow boundary, and the initial datum are summed at quadrature points (Sobol sequences in the interior, Gauss-Legendre for the scattering integral). The estimate that carries the argument is a stability inequality for the error $\hat u = u^* - u$, derived by energy estimates and Grönwall's lemma, which bounds the true $L^2$ error by the $L^2$ norms of the residuals; those norms are then converted to the discrete training errors plus quadrature remainder terms by the Koksma-Hlawka inequality and Gauss-quadrature error bounds.

What would settle it

Construct a steady radiative-transfer test problem with zero absorption, unit scattering, and isotropic kernel — exactly the regime of the paper's section 3.3, where assumption (B.6) fails — and compare the trained PINN against a high-resolution reference solution. If a network with very small training loss can be produced whose true $L^2$ error is large, the 'train well, generalize well' transfer claim would be shown to fail whenever the Hardy-Krause and coefficient assumptions are violated.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a feedforward neural network trained by gradient descent on the loss function (2.14) — the squared residual of the radiative transfer equation plus initial and boundary residuals, collocated at low-discrepancy Sobol points — approximates the radiative intensity accurately, and that the approximation error is bounded by the computable training errors. Concretely, for the time-dependent problem, the generalization error satisfies $E_G^2 \le C\big((E_{tb}^T)^2 + c(E_{sb}^T)^2 + c(E_{int}^T)^2\big) + CC^*\big(\frac{(\log N_{tb})^{2d}}{N_{tb}} + \frac{(\log N_{sb})^{2d}}{N_{sb}} + \frac{(\log N_{int})^{2d+1}}{N_{int}} + N_S^{-2s}\big)$, so that 'as long as the PINN is trained well, it generalizes well.' In the steady case, a similar bound (B.8) is proven under the coefficient condition (B.6). The same residual-minimization idea, extended with a data-fidelity term, solves the inverse problem of determining the absorption coefficient from incident-radiation measurements.

Load-bearing premise

The transfer from a small training loss to a small true error assumes the trained network and its PDE residual have bounded Hardy-Krause variation (and, in the steady case, that absorption is strong enough relative to scattering), conditions that the training process does not enforce and that are not verified for any of the paper's steady-state experiments.

Editorial extensions

If this is right

If the error bound holds as stated, training loss is a trustworthy a posteriori error indicator: a user can stop training once the residual loss is small and know the radiative intensity is accurate.
The method extends to full seven-dimensional time-dependent polychromatic problems without a curse of dimensionality beyond logarithmic factors, so it can serve as the radiation module in hydrodynamics codes.
The inverse algorithm recovers absorption coefficients from incident-radiation measurements, opening the same treatment for scattering coefficients, emission terms, and scattering kernels.
Because the optimizer and loss are simple, the algorithm is far easier to implement than sparse-grid or discrete-ordinate methods while matching their published accuracy.
Low-discrepancy training points plus Gauss quadrature for scattering give a dimension-robust recipe applicable to other linear transport equations.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the Hardy-Krause assumption could be verified or enforced during training, the bound would become fully rigorous; one testable route is adding a penalty on the total variation of the residual during training and checking whether generalization improves on the paper's own steady benchmarks.
The steady-state experiments run in a regime where the proven steady bound does not apply, so the paper's steady results would need a separate stability argument, for example based on a different norm or on the time-dependent estimate with large but finite speed of light.
The same collocation-plus-quadrature recipe may transfer to other integro-differential equations, such as the Boltzmann equation with a collision kernel, where the scattering integral plays the role of the kernel integration.
For inverse problems, the observed accuracy in recovering the absorption coefficient suggests the method could be extended to joint estimation of several coefficients from partial moment measurements, though uniqueness would then require stronger data or regularization.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: this is a real extension of PINNs to radiative transfer, with an RT-specific residual and a generalization bound that explicitly tracks scattering quadrature. But the advertised steady-state bound does not apply to any of the steady experiments, because assumption (B.6) and strict positivity of k and sigma fail there. The paper is worth engaging, not worth taking as validated.

What's new: the scattering integral is written as a quadrature in the PINN residual, and the time-dependent bound (A.3) includes the N_S^{-2s} term from the Gauss rule. That is a genuine contribution over the authors' prior framework. The 1D steady test against the exact boundary solution (0.07-0.29% L2) and the 6D polychromatic flux test (2.1% error) are real, reproducible checks, and the code is linked. The inverse problem is a standard synthetic demonstration; that's fine as a proof-of-concept.

Soft spots, in order of importance. First, Lemma B.1's hypotheses are not met in Sections 3.2-3.4: k_min = 0 in every steady case, and in 3.3 the left side of (B.6) is negative (-1/4pi). So low training error does not, by the paper's own theory, imply the claimed steady generalization. The text even invokes the time-dependent bound (A.3) in 3.3, which is the wrong tool for c -> infinity. Second, the claim that PINN beats the diffusion approximation at K=2 in Section 3.5 has no full-RT ground truth; the comparison only shows that the diffusion solution differs, which is expected. Third, the no-curse claim rests on unverified Hardy-Krause variation of trained networks and residuals. Fourth, the Discussion sentence that the error estimates were validated by the experiments is not supported, since the bounds don't apply where they're tested.

The conditional estimates themselves look conventional and likely correct. The issue is the gap between theory and experiments, not the algebra. Citation pattern is fine; the self-citations are to the framework being extended.

Who it's for: anyone working on PINN error estimates or on ML solvers for radiative transfer. It deserves peer review---it's a legitimate first step with reproducible experiments---but a referee should require the authors to either extend the steady theory to the weakly absorbing/thin cases or clearly state that the experiments are not covered by the bound, and to temper the K=2 claim.

Referee Report

3 major / 5 minor

Summary. The paper adapts physics-informed neural networks (PINNs) to the radiative transfer equation, both forward and inverse. The forward algorithm minimizes a loss composed of PDE, boundary, and initial residuals at quadrature points, and the authors derive conditional generalization-error estimates (A.3) and (B.8) that bound the L2 error by training errors plus quasi-Monte Carlo and quadrature terms. They present five forward experiments, from 1D steady slab transport to time-dependent polychromatic 3D transport, and one inverse experiment recovering an absorption coefficient from incident-radiation data. The paper claims that the experiments validate the error estimates and that the method is simple, fast, and accurate.

Significance. If the theoretical estimates are valid in the regimes where they are applied, the paper provides a useful conditional explanation of PINN performance for a high-dimensional linear kinetic equation, with only logarithmic dimension dependence in the quadrature terms. The paper also ships reproducible code and has concrete, checkable empirical successes: the 1D steady experiment matches the exact boundary solution to 0.07–0.29%, the 6D polychromatic flux matches the analytic flux to 2.1%, and the inverse problem achieves errors below 3% for the recovered coefficient. These anchor results give the paper genuine value even where the theory is not fully matched to the experiments.

major comments (3)

[§3.2–3.4, Lemma B.1 (Appendix B)] The steady-state error estimate (B.8) cannot be invoked for the steady numerical experiments, because its hypotheses fail in each case. Section 3.2 has k=0 and σ(x)=x, so k_min=σ_min=0; Section 3.3 has k=I_b (hence k_min=0), σ=1, Φ=1, s_d=4π and ||Ψ||_∞=4π, so the left-hand side of (B.6) is 0+1-(1+4π)/(4π)=-1/(4π)<0; Section 3.4 explicitly uses zero absorption. Therefore the low training errors in Tables 2–4 do not, by the paper's own theorem, imply the claimed generalization accuracy. The reliance on estimate (A.3) in Section 3.3 is not a substitute, since (A.3) is the time-dependent bound and Remark 2.4 states that it is unsuitable for steady problems.
[§3.5] The conclusion that for K=2 the PINN solution of the full transport equation (3.9) is more accurate than the diffusion approximation is not supported, because the only comparison is with the diffusion solution (3.14) and there is no reference solution of the full transport equation. A mismatch between the PINN intensity and the diffusion approximation could also be caused by PINN approximation error; the low reported training error and estimate (A.3) do not close this gap. In addition, the theoretical QMC bound in Section 2.5 is stated for rectangular domains rescaled to [0,1]^{2d+1}, whereas this experiment uses a spherical shell with Sobol training points. Please either add a transport reference solution or soften the claim.
[§5] The statement that 'the predictions of the error estimates were validated by the experiments' is not supported as written: the steady experiments fall outside the hypotheses of Lemma B.1, and for the 3D steady experiment in Section 3.3 no quantitative generalization error is computed. The sentence should be restricted to regimes where the assumptions hold, or reformulated as consistency with, rather than validation of, the estimates.

minor comments (5)

[Eq. (3.1)] The scattering integral appears to contain a typo: the integrand should be u(x, μ′) rather than u(z, μ).
[§3.5] The text sets Nsb = Ntb = 12228, whereas Table 5 reports 12288; please correct the inconsistent number.
[Eq. (A.4)] The symbol C is used both for the Grönwall constant in the first line of (A.4) and for the quadrature constant inside the definition of C*; this is confusing and somewhat circular. Please rename one of the two constants.
[Eq. (B.9)] Several terms in the definition of C appear to be missing the Hardy-Krause variation notation: the terms involving (R_int*)^2 and C N_S^{-2s} should likely be V_HK((R_int*)^2) and a variation-weighted quadrature term, consistent with (A.4).
[Throughout] There are several typos to correct: 'Guass-quadrature rule' in Lemmas A.1 and B.1, 'trainig' in Section 3.4, and 'T able 1' near Table 1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the generalization error bound is a genuine conditional stability estimate, not a restatement of the training loss.

full rationale

The paper's central estimate (2.18)/(A.3) is a conditional stability-plus-quadrature bound. The generalization error (2.16) is the true L2 error against the PDE solution, while the training errors (2.17) are residual norms at collocation points. Lemma A.1 derives the bound by propagating the residuals through the error equation (A.6) with a Gronwall argument and controlling quadrature errors via Koksma-Hlawka and Gauss-quadrature estimates (A.13)-(A.15). The generalization error is not defined as, nor equated to, the training loss; it is bounded by it under explicit regularity assumptions. The steady-state Lemma B.1 follows the same structure from the error equation (B.11) and assumption (B.6). No step reduces to its own input by construction. The cited prior work [30,31] is from the same authors, but the load-bearing estimates are re-derived in this paper's appendices rather than merely imported, so the self-citations are not circular. The skeptical concern that assumption (B.6) is violated by the steady experiments in Sections 3.2-3.4 is a correctness and validity gap, not a circularity, and does not affect this score.

Assumptions & free parameters 5 free parameters · 8 assumptions · 0 invented entities

No new physical entities are introduced; the absorption-coefficient network k_theta_k in Section 4 is a parameterization of an existing quantity, not a new entity. The paper's theoretical claims rest on four structural assumptions beyond standard background math: bounded Hardy-Krause variation of the trained network and its residual (A.2)/(B.7), the steady-state coercivity condition (B.6), the success of non-convex training in reaching a small-residual minimum, and the correctness of the diffusion model as a benchmark in the K = 20 regime. The first and third are unverified by the experiments; the second is provably violated by the paper's own steady experiments. Free parameters are the hand-selected and ensemble-selected hyperparameters lambda, lambda_reg, lambda_k, network sizes, and retraining budgets; no constants are fitted to data within the error bounds, so the numerical comparisons are not data-fitted.

free parameters (5)

lambda (loss-balance weight) = 0.1 (best in Tables 2-5); 1.0 (Table 6)
Chosen from {0.1, 1, 10} by ensemble selection on training loss; balances PDE residual versus initial and boundary residuals in (2.14) and (4.3).
lambda_reg (weight regularization) = 0 (most runs); 1e-6, 1e-5 (Example 3.4)
Regularization weight in (2.15); hand-set and ensemble-tuned, influences the reported training errors.
lambda_k (Tikhonov weight for inverse problem) = 0.001
Hand-set in (4.5) to select a unique absorption coefficient; directly shapes the recovered k in the inverse experiment.
Network depth K-1 and width d_tilde = 8x24 (Tables 2-3), 8x20 (Table 4), 4x40 (Table 5), 8x20 (Table 6)
Architecture choice from ensemble search; determines expressiveness and the reported training errors; no theory in the paper links architecture size to the constants in the bounds.
Number of retrainings n_theta = 5 to 20 (Table 1)
Best-of-n_theta selection over random initializations; results depend on this uncontrolled search budget, and no variance over seeds is reported.

assumptions (8)

standard math Koksma-Hlawka inequality and low-discrepancy of Sobol points
Invoked at (A.13)-(A.14) to convert the training sums into L2 norms plus discrepancy terms; standard QMC background, cited to [2, 44].
standard math Gauss-Legendre quadrature error of order s for the scattering integral
Estimate (A.15) bounds the scattering-quadrature error by C N_S^{-2s}; requires smoothness of Phi and u*, as assumed in Lemma A.1.
domain assumption Unique weak solution u of the radiative transfer equation exists under stated coefficient bounds
Lemma A.1 and Lemma B.1 assume u in L2 exists and is unique; standard for linear transport with L-infinity coefficients, cited to [33], not proven in the paper.
domain assumption Symmetric scattering kernel with bounded marginal Psi in L-infinity (A.1)
Needed for the stability/coercivity estimates; holds for the isotropic and Legendre kernels used in the experiments, but is an input assumption of the theory.
ad hoc to paper The trained network u* and its residuals have bounded Hardy-Krause variation (A.2), (B.7)
Required for the Koksma-Hlawka step; plausible for tanh networks but never verified for the output of a stochastic optimizer. This assumption is what converts low training error into a small generalization error, and it is load-bearing for the no-curse-of-dimensionality claim.
ad hoc to paper Steady-state coercivity condition (B.6): k_min + sigma_min - (sigma_max + ||Psi||_inf)/s_d > kappa > 0, and 0 < k_min, 0 < sigma_min
Needed for Lemma B.1. Violated in the paper's own steady experiments: section 3.2 gives -3/2 and section 3.3 gives -1/(4pi). The paper does not state that its steady theory excludes these regimes.
ad hoc to paper The optimizer reaches a minimum with small training error
The entire conditional bound requires small Eint_T, Esb_T, Etb_T. Remark 2.2 concedes these are only computable after training; the ability of ADAM/LBFGS to find such minima in high dimensions is empirical, not proven.
domain assumption The diffusion approximation (3.11) is accurate for large Knudsen numbers
Used to treat the analytic solution (3.14) as a benchmark for the K = 20 case and to motivate the claim that deviation at K = 2 is the diffusion model's fault; standard in radiation hydrodynamics, cited to [3, 11].

how reviews work

0 comments

Cite this review

Pith. "Pith review of Physics Informed Neural Networks for Simulating Radiative Transfer." pith.science (2026). https://pith.science/paper/NM2G2GBB

@misc{pith2026200913291,
  author       = {Pith},
  title        = {Pith review of: Physics Informed Neural Networks for Simulating Radiative Transfer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NM2G2GBB}},
  note         = {Machine review of arXiv:2009.13291}
}

read the original abstract

We propose a novel machine learning algorithm for simulating radiative transfer. Our algorithm is based on physics informed neural networks (PINNs), which are trained by minimizing the residual of the underlying radiative tranfer equations. We present extensive experiments and theoretical error estimates to demonstrate that PINNs provide a very easy to implement, fast, robust and accurate method for simulating radiative transfer. We also present a PINN based algorithm for simulating inverse problems for radiative transfer efficiently.

Figures

Figures reproduced from arXiv: 2009.13291 by the authors.

**Figure 1.** An illustration of a (fully connected) deep neural network. The red neurons represent the inputs to the network and the blue neurons denote the output layer. They are connected by hidden layers with yellow neurons. Each hidden unit (neuron) is connected by affine linear maps between units in different layers and then with nonlinear (scalar) activation functions within units. 2.4 Training PINNs: Loss functions and op… view at source ↗

**Figure 2.** Contour plot of the radiative intensity u(x, µ) for the 1D monochromatic experiment Another attractive feature of this simplified problem lies in the fact that the authors in [4] obtained an exact analytical solution for it. Although it is very complicated to evaluate this solution for the whole (x, µ)-plane, its values on the boundaries can be readily evaluated i.e. we can readily compute u−(µ) = u(0, µ) and u+(µ) … view at source ↗

**Figure 3.** Comparison of the analytical and PINNN radiative intensity at the physical domain boundaries for the stationary monochromatic radiative transfer in one-space dimension. 3.3 Monochromatic stationary radiative transfer in three space dimensions Next, we consider a monochromatic and stationary version of the general radiative transfer equations (1.1), but in three space dimensions. Already, this problem is in 5 dimensi… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Contour plot of the incident radiation G(x) for the 3D monochromatic experiment 3.4 Polychromatic stationary radiative transfer in three space dimensions Next, we consider the most general case of the steady state radiative transfer equation (B.1) by following the setu…

**Figure 5.** Figure 5: Comparison of incident radiationn with respect to the exact solution of the diffusion approximation (3.12) and PINN approximation of the full radiative transfer equation (3.9) at rescaled time τ = 1 for two different values of the absorption coefficient kν = 1, 10 prov…

**Figure 6.** Figure 6: Comparison of exact solutions of the diffusion approximation (3.12) with the PINN approximation of the full radiative transfer equation (3.9) for two different values of the absorption coefficient kν = 1, 10 at different radial locations and at rescaled time τ = 1 k(x,…

**Figure 7.** Figure 7: Results for the PINNs algorithm 4.1 for the inverse problem for radiative transfer. The PINNs approximation to the incident radiation and absorption coefficient are plotted along the diagonal of the unit cube and compared with the measured data G¯ and ground truth abso…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

Physics Enhanced Deep Surrogates for the Phonon Boltzmann Transport Equation
physics.comp-ph 2025-11 conditional novelty 5.0 of 10

Embedding a differentiable Fourier solver as a low-fidelity physics core lets a neural surrogate predict phonon-BTE conductivity of porous nanostructures to ~5% error with 300 BTE simulations and design targets at ~4%...

Reference graph

Works this paper leans on

48 extracted references · 43 canonical work pages · cited by 1 Pith paper

[1]

A. R. Barron. Universal approximation bounds for superpositions of a sigmoidal function. IEEE Trans. Inform. Theory., 39(3):930–945, 1993

work page 1993
[2]

R. E. Caﬂisch. Monte carlo and quasi-monte carlo methods. Acta Numerica, 7:1–49, 1998

work page 1998
[3]

J. I. Castor. Radiation hydrodynamics. Cambridge University Press, 2004

work page 2004
[4]

Cengel, M

Y. Cengel, M. ¨Ozi, et al. Radiation transfer in an anisotropically scattering plane-parallel medium with space-dependent albedo ω(x). Journal of Quantitative Spectroscopy and Radiative Transfer , 34(3):263–270, 1985

work page 1985
[5]

Y. Chen, L. Lu, G. E. Karniadakis, and L. D. Negro. Physics-informed neural networks for inverse problems in nano-optics and metamaterials. Preprint, available from arXiv:1912.01085, 2019

work page Pith review arXiv 1912
[6]

W. E, J. Han, and A. Jentzen. Deep learning-based numerical methods for high-dimensional parabolic partial diﬀerential equations and backward stochastic diﬀerential equations. Communications in Mathematics and Statistics , 5(4):349–380, 2017

work page 2017
[7]

Evans, J

R. Evans, J. Jumper, J. Kirkpatrick, L. Sifre, T. Green, C. Qin, A. Zidek, A. Nelson, A. Bridgland, H. Penedones, et al. De novo structure prediction with deep-learning based scoring. Annual Review of Biochemistry, 77(363-382):6, 2018

work page 2018
[8]

Fletcher

R. Fletcher. Practical methods of optimization . John Wiley and sons, 1987. 23

work page 1987

Show all 48 references

[9]

M. Frank. Approximate models for radiative transfer. Bull. Inst. Math. Acad. Sinica (New Series) , 2:409–432, 2007

2007
[10]

Goodfellow, Y

I. Goodfellow, Y. Bengio, and A. Courville. Deep learning. MIT press, 2016

2016
[11]

Graziani

F. Graziani. The prompt spectrum of a radiating sphere: Benchmark solutions for diﬀusion and transport. In Computational methods in transport: veriﬁcation and validation, LNCSE-62 , pages 151–167. Springer, 2008

2008
[12]

K. Grella. Sparse tensor approximation for radiative transport . PhD thesis, ETH Zurich, 2013

2013
[13]

Grella and C

K. Grella and C. Schwab. Sparse tensor spherical harmonics approximation in radiative transfer. J. Comput. Phys., 230:8452–8473, 2011

2011
[14]

J. Han, A. Jentzen, and W. E. Solving high-dimensional partial diﬀerential equations using deep learning. Proceedings of the National Academy of Sciences , 115(34):8505–8510, 2018

2018
[15]

A. D. Jagtap and G. E. Karniadakis. Extended physics-informed neural networks (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial diﬀerential equations. Communications in Computational Physics , 28(5):2002–2041, 2020

2002
[16]

A. D. Jagtap, K. Kawaguchi, and G. E. Karniadakis. Adaptive activation functions accelerate conver- gence in deep and physics-informed neural networks. Journal of Computational Physics , 404:109126, 2020

2020
[17]

A. D. Jagtap, E. Kharazmi, and G. E. Karniadakis. Conservative physics-informed neural networks on discrete domains for conservation laws: Applications to forward and inverse problems. Computer Methods in Applied Mechanics and Engineering , 365:113028, 2020

2020
[18]

Kanschat and et

G. Kanschat and et. al. Numerical methods in multi-dimensional radiative transfer . Springer, 2008

2008
[19]

Kaushik, M

D. Kaushik, M. Smith, A. Wollaber, B. Smith, A. Siegel, and W. S. Yang. Enabling high ﬁdelity neutron transport simulations on petascle architectures. In Proceedings of the Conference on High Performance Computing Networking, Storage, and Analysis, volume 67, Portland, Oregon , 2009

2009
[20]

D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In 3rd International Conference on Learning Representations, ICLR 2015 , 2015

2015
[21]

I. E. Lagaris, A. Likas, and P. G. D. Neural-network methods for bound- ary value problems with irregular boundaries. IEEE Transactions on Neural Networks , 11:1041–1049, 2000

2000
[22]

I. E. Lagaris, A. Likas, and D. I. Fotiadis. Artiﬁcial neural networks for solving ordinary and partial diﬀerential equations. IEEE Transactions on Neural Networks , 9(5):987–1000, 1998

1998
[23]

K. D. Lathrop. Ray eﬀects in discrete ordinates equations. Nucl. Sci. Eng , 32:357, 1968

1968
[24]

LeCun, Y

Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 521(7553):436–444, 2015

2015
[25]

Y. Liu, X. Meng, and G. E. Karniadakis. B-pinns: Bayesian physics-informed neural networks for forward and inverse pde problems with noisy data. Preprint, available from arXiv:2003.06097, 2020

2003 arXiv
[26]

L. Lu, X. Meng, Z. Mao, and G. E. Karniadakis. Deepxde: A deep learning library for solving diﬀerential equations. Preprint, available from arXiv:1907.04502, 2019

1907 arXiv
[27]

K. . O. Lye, S. Mishra, P. Chandrasekhar, and D. Ray. Iterative surrogate model optimization (ismo): An active learning algorithm for pde constrained optimization with deep neural networks. Preprint, available as arXiv:2008.05730, 2020

2008 arXiv
[28]

K. O. Lye, S. Mishra, and D. Ray. Deep learning observables in computational ﬂuid dynamics. Journal of Computational Physics , page 109339, 2020. 24

2020
[29]

Z. Mao, A. D. Jagtap, and G. E. Karniadakis. Physics-informed neural networks for high-speed ﬂows. Computer Methods in Applied Mechanics and Engineering , 360:112789, 2020

2020
[30]

Mishra and R

S. Mishra and R. Molinaro. Estimates on the generalization error of physics informed neural networks (pinns) for approximating pdes. Preprint, available from arXiv:2006:16144v1, 2020

2006
[31]

Mishra and R

S. Mishra and R. Molinaro. Estimates on the generalization error of physics informed neural net- works (pinns) for approximating pdes ii: A class of inverse problems. Preprint, available from arXiv:2007:01138v1, 2020

2007
[32]

Mishra and T

S. Mishra and T. K. Rusch. Enhancing accuracy of deep learning algorithms by training with low-discrepancy sequences. Preprint, available as arXiv:2005.12564, 2020

2005 arXiv
[33]

M. F. Modest. Radiative heat transfer. Elsevier, 2003

2003
[34]

M. F. Modest and J. Yang. Elliptic pde formulation and boundary conditions of the spherical har- monics method of arbitrary order for general three-dimensional geometries. Journal of Quantitative Spectroscopy and Radiative Transfer, 109:1641–1666, 2008

2008
[35]

Mohri, A

M. Mohri, A. Rostamizadeh, and A. Talwalkar. Foundations of machine learning . MIT press, 2018

2018
[36]

G. Pang, L. Lu, and G. E. Karniadakis. fpinns: Fractional physics-informed neural networks. SIAM journal of Scientiﬁc computing , 41:A2603–A2626, 2019

2019
[37]

Paszke, S

A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer. Automatic diﬀerentiation in pytorch. In Workshop Proceedings of Neural Information Processing Systems, 2017

2017
[38]

Pontaza and J

J. Pontaza and J. Reddy. Least-squares ﬁnite element formulations for one-dimensional radiative transfer. Journal of Quantitative Spectroscopy and Radiative Transfer , 95(3):387–406, 2005

2005
[39]

Raissi and G

M. Raissi and G. E. Karniadakis. Hidden physics models: Machine learning of nonlinear partial diﬀerential equations. Journal of Computational Physics , 357:125–141, 2018

2018
[40]

Raissi, P

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

2019
[41]

Raissi, A

M. Raissi, A. Yazdani, and G. E. Karniadakis. Hidden ﬂuid mechanics: A navier-stokes informed deep learning framework for assimilating ﬂow visualization data. arXiv preprint arXiv:1808.04327 , 2018

2018 arXiv
[42]

Richling, E

S. Richling, E. Meink¨ ohn, N. Kryzhevoi, and G. Kanschat. Radiative transfer with ﬁnite elements-i. basic method and tests. Astronomy & Astrophysics, 380(2):776–788, 2001

2001
[43]

Shukla, P

K. Shukla, P. C. Di Leoni, J. Blackshire, D. Sparkman, and G. E. Karniadakis. Physics-informed neural network for ultrasound nondestructive quantiﬁcation of surface breaking cracks. Journal of Nondestructive Evaluation, 39(3):1–20, 2020

2020
[44]

I. M. Sobol’. On the distribution of points in a cube and the approximate evaluation of integrals. Zhurnal Vychislitel’noi Matematiki i Matematicheskoi Fiziki , 7(4):784–802, 1967

1967
[45]

Stamatellos and A

D. Stamatellos and A. P. Whitworth. Probing the initial conditions for star formation with monte carlo radiative transfer simulations. In Numerical Methods in Multidimensional Radiative Transfer, G. Kanschat, E. Meink¨ ohn, R. Rannacher, and R. Wehrse, eds, pages 289–298. Spri...

2008
[46]

Stoer and R

J. Stoer and R. Bulirsch. Introduction to numerical analysis . Springer Verlag, 2002

2002
[47]

G. Widmer. Sparse Finite Elements for Radiative Transfer. PhD thesis, ETH Zurich, 2009

2009
[48]

Zhang, L

W. Zhang, L. Howell, A. Almgren, A. Burrows, J. . Dolence, and J. Bell. Castro: A new compressible astrophysical solver. iii. multigroup radiation hydrodynamics. Astrophysical Journal (supplement series), 204(7):27 pp, 2013. 25

2013

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.

[1] [1]

A. R. Barron. Universal approximation bounds for superpositions of a sigmoidal function. IEEE Trans. Inform. Theory., 39(3):930–945, 1993

work page 1993

[2] [2]

R. E. Caﬂisch. Monte carlo and quasi-monte carlo methods. Acta Numerica, 7:1–49, 1998

work page 1998

[3] [3]

J. I. Castor. Radiation hydrodynamics. Cambridge University Press, 2004

work page 2004

[4] [4]

Cengel, M

Y. Cengel, M. ¨Ozi, et al. Radiation transfer in an anisotropically scattering plane-parallel medium with space-dependent albedo ω(x). Journal of Quantitative Spectroscopy and Radiative Transfer , 34(3):263–270, 1985

work page 1985

[5] [5]

Y. Chen, L. Lu, G. E. Karniadakis, and L. D. Negro. Physics-informed neural networks for inverse problems in nano-optics and metamaterials. Preprint, available from arXiv:1912.01085, 2019

work page Pith review arXiv 1912

[6] [6]

W. E, J. Han, and A. Jentzen. Deep learning-based numerical methods for high-dimensional parabolic partial diﬀerential equations and backward stochastic diﬀerential equations. Communications in Mathematics and Statistics , 5(4):349–380, 2017

work page 2017

[7] [7]

Evans, J

R. Evans, J. Jumper, J. Kirkpatrick, L. Sifre, T. Green, C. Qin, A. Zidek, A. Nelson, A. Bridgland, H. Penedones, et al. De novo structure prediction with deep-learning based scoring. Annual Review of Biochemistry, 77(363-382):6, 2018

work page 2018

[8] [8]

Fletcher

R. Fletcher. Practical methods of optimization . John Wiley and sons, 1987. 23

work page 1987