Higher-order methods of radiative transfer in simulations of the epoch of reionisation: Pn versus M1

The paper investigates Higher-order methods of radiative transfer in simulations of the epoch of reionisation: Pn versus M1, comparing the performance and reliability of higher-order methods (P n) against a reference model (M1).

Need for Higher Orders and Convergence:

The authors note that using higher orders is crucial because lower orders exhibit significant limitations. Regarding convergence, they state that the oscillations of the models are very strong under order 9. For orders above 9 however, the output becomes almost indistinguishable from M1. Consequently, they argue that we can assume that orders above P9 are able to handle the worst case of Pn the best, leading them to focus mainly on P 9.

Limitations of Low-Order Models (The Clump Test):

A key area of concern is the performance of low-order models in unfavorable, non-isotropic scenarios. Specifically, in the clump test... where all sources are directional, thus non isotropic, the results using P 3 are problematic. The authors state that P 3 cannot reproduce correctly what we physically expect from this test. Furthermore, they point out that P 3 creates visible local maxima in the photon density, and tends to over-ionise the sphere, and critically, its modes output negative photon densities in a large area close to the sources... which creates artefacts in the neutral fraction and temperature that should not exist. To mitigate this negativity and perform comparably to M1, they conclude that a higher order is required, and as such, a higher computational cost.

Analysis of Negative Photon Densities (The Cosmological Map Test):

To assess the stability of the models at high orders, the authors examine the fraction of negative cells and the minimum photon density in a cosmological map test. They report that while the fraction of negative cells seems to be almost independent of the order, peaking around 7.5% at 0.2 Myr, they observe that for minimum density, all models above P7 have an almost negligible negative outputs, which significantly reduces the impact on the simulation compared to P 3 or P 5. This is because when approximating a negative cell value as zero (Eq. 14), the impact of this modification will be smaller if the value of our negative cell is already almost a 0.

Methodological Improvements and Boundary Conditions:

For specific tests, such as the shadowing a dense clump test, which uses anisotropic sources, the interaction between non-isotropic P n sources and boundary conditions can create oscillations. To minimize this impact and ensure a fair comparison between M1 and P n, the authors implemented nested boxes, creating buffers between the main simulation box (the 643 times 6.6 kpc box) and a larger outer buffer (1283 times 13.2 kpc). This setup is designed to minimize the impact of boundary conditions on the simulation box.

Artifacts and Model Comparison:

The investigation was also motivated by known artifacts in previous simulations. The authors mention the Dark Sombrero artefact, which is a property of the M1 model observed in large published simulations (e.g., CoDa II). The discovery of this artifact was one of the reasons that led our team to investigate the potential issues of the M1 closure and to compare it with another state of the art model.

Based on a rigorous analysis of this research paper, the transition from the standard M1 approximation to a higher-order model like Pn presents several critical opportunities for optimizing and validating complex AI systems used in astrophysical simulations. The following improvements detail how an advanced AI framework can leverage these findings.

Improvement: Replace or augment the standard M1 (Moment-based, Order 1) RT kernel within the simulation pipeline with a high-order Pn implementation (e.g., P 9 or P 25).

• Specific Action: The AI system must be updated to handle the the (n+1) squared coefficients required by Pn, moving beyond the simple four coefficients of M1. This requires a dynamic memory allocation and processing architecture capable of handling high-dimensional state vectors for each grid cell.

Improvement: The AI system must be configured to enforce physical conservation laws regarding photon directionality, particularly at source interactions.

• Specific Action: When the simulation encounters two intersecting radiation beams (as seen in Figure 3), the AI kernel must execute a vector-sum avoidance mechanism. Instead of merging and adding direction vectors (M1's approach), the Pn implementation requires calculating the path where photons simply cross, maintaining their individual angular momentum, thereby preventing unphysical coalescence artifacts.

Improvement: Integrate a predictive error model derived from the PN vs M1 comparison to quantify uncertainty in observable outputs (x HI).

• Specific Action: The AI system should utilize the statistical properties of the relative error (e.g., a Gaussian distribution with sigma = 0.27 dex in 10(x HI)) as a confidence interval when generating predictions for the Epoch of Reionization (EoR). This allows researchers to not only provide an answer but also quantify the expected discrepancy between M1 and Pn, ensuring that model-driven conclusions are statistically robust.

Improvement: Adopt the Reduced Kinetic Model Solver (RKMS) architecture for massive parallelization across multiple GPUs.

• Specific Action: The AI system must be optimized to run the P n calculation using an OpenCL/GPU framework, rather than sequential CPU loops. This requires restructuring the simulation time-stepping logic into a massively parallel kernel where each grid cell updates its P n state simultaneously, achieving computational efficiency far superior to previous non-parallel CPU implementations.

The system can autonomously identify and flag unphysical simulation artifacts, specifically the Dark Sombrero.

• Mechanism: The AI monitors photon density profiles around sources. If a spherical, localized photon-deficit shell appears (as detailed in Figure 21), the system flags this as an M1 failure mode.

• Utility: This provides a high-confidence validation metric for running simulations, allowing the researchers to discard any results derived from M1 and ensuring that the final reported data only reflects physically plausible phenomena.

The system can provide more accurate predictions regarding the timing of reionization.

• Mechanism: By comparing Pn's convergence rate in the Strömgren sphere tests (Figure 4) with M1, the the AI can predict when a given source intensity (S 0) will achieve a specific ionization radius faster or slower than M1.

• Utility: It can generate simulations that converge toward an accurate, physically expected result (e.g., matching the analytical Strömgren sphere progression) rather than being misled by M1's fluid-like behavior, directly aiding the timing of reionisation open question.

The system can map and quantify the spatial distribution of model discrepancies across large cosmological domains.

• Mechanism: By calculating the ratio x HI P n over x HI M1 (Figure 23) across various lines of sight, the AI can generate a precise map of where M1 is under-ionizing or over-ionizing relative to Pn.

• Utility: This allows scientists to pinpoint exactly where the directional loss in M1 is causing physical inconsistencies, guiding future targeted tests and ensuring that complex, multi-source simulations are not compromised by local artifacts.

The system can automatically determine the optimal trade-off between computational cost and physical accuracy for a given problem.

• Mechanism: By running the test cases across various Pn orders (P 3, P 5, P 7, P 9), the AI identifies where higher-order models become negligible in their oscillations (e.g., above P 10).

• Utility: This prevents unnecessary computational waste. Instead of always running P 25, the AI can recommend that a P 7 implementation is sufficient for standard EoR simulations, balancing accuracy against the massive memory and time cost of higher orders.

In current cosmological simulations, the radiative transfer modules generally rely on the M1 approximation, which has some glaring flaws related to its fluid-like behaviour, such as spurious pseudo-sources and loss of directionality when radiation fronts from different directions collide. Pn, another moment-based model used in other fields of physics, may correct these issues. We aim at testing out Pn in an astrophysical setting and compare it to M1, in order to see if it can indeed correct M1's flaws. Also, we want to use Pn's solutions to better pinpoint M1 errors. We implement a Pn radiation transport method and couple it to a photo-thermo-chemistry module to account for the interaction of ionising radiation with the Hydrogen gas, and benchmark it using tests for radiative transfer models comparison in astrophysics as defined in arXiv:astro-ph/0603199. We find that high order P n (e.g. P9) indeed correct M1's flaws, while faring as well or even better in some aspects in the tests, in particular when directionality is important or colliding radiation fronts occur. By comparing P9 and M1 radiation fields in an idealised and cosmological test case, we highlight a new, thus far unreported artefact of M 1, the 'dark sombrero'. A dark sombrero appears as a spherical photon-deficit shell around the source. The photon density in dark sombreros can be underestimated by a factor up to 2-3. They occur in regions where a source's radiation field connects with that of another source or group of sources. These basic properties (position and amplitude) of the dark sombreros may depend on the sources' relative intensities, positions, spatial resolution, although we have not been able to test this in detail in this study.