Solving BDNK diffusion using physics-informed neural networks

arXiv:2602.16117 · nucl-th, astro-ph.HE, gr-qc · Submitted 2026-02-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.

Vera: Today's paper: "Solving BDNK diffusion using physics-informed neural networks".

Jocelyn: The gist Solving BDNK diffusion using physics-informed neural networks reformulates the relativistic BDNK diffusion equation into a flux-conservative form and solves it in (1+1)D using both a second-order Kurganov-Tadmor finite…

Vera: First, who's behind it and why it matters.

Title and authors: Vera: So we're looking at this paper called "Solving BDNK diffusion using physics-informed neural networks," and what it does is take that complex relativistic diffusion equation and turn it into a flux-conservative form. It then solves that in one plus one dimensions using two different methods: a finite volume scheme and these physics-informed neural networks, or PINNs.

Jocelyn: That sounds like they're trying to find a way to make the math easier for computers to handle when dealing with fluid dynamics in space where relativity matters. It’s about taking something that might be hard to solve directly and reformulating it so the numerical solvers can work better, right?

Subrahmanyan: Exactly, it tackles the BDNK diffusion equation, which is a way we look at how matter and charge move around in fluids under relativistic conditions. The authors are looking at this as a more tractable approach to relativistic viscous hydrodynamics near equilibrium.

Vera: So what’s the core of what they found? They set up these equations in flux-conservative form, which means they write everything in terms of densities and fluxes so it looks like a conservation law, and then they use that structure to build their numerical solver.

Jocelyn: And for the solving part, they tested both a standard second-order Kurganov-Tadmor finite volume scheme and these PINNs, which is basically training a neural network to follow the physics of the equation.

Subrahmanyan: The key innovation here seems to be their introduction of something called the SA-PINN-ACTO framework, which combines self-adaptive techniques with an algebraic transform specifically for enforcing initial and boundary conditions exactly.

Vera: That’s what interests me—how they’re handling those tricky starting points and boundaries without messing up the physics. It sounds like they’ve tried to let the neural network focus only on solving the equation itself.

Jocelyn: So, when we look at their summary, it seems like they are showing that this combination of techniques works for both smooth and discontinuous initial data, which is important because real physical systems aren't always perfectly smooth.

Subrahmanyan: They tested this framework on two characteristic speeds and for two different kinds of BDNK backgrounds, which suggests the method has some stability across different conditions.

Vera: That brings us to the improvements they suggest they’ve made. They’re essentially showing how their SA-PINN-ACTO setup can handle sharper gradients better by using a self-adaptive PINN technique that weights different parts of the solution based on where it needs more attention.

Jocelyn: And this self-adaptation is linked to redefining the loss function, where each term in the residual is weighted by a learned factor lambda i, which lets the network learn what’s hard to approximate.

Subrahmanyan: That addresses one of the challenges with PINNs, which is that they often struggle near sharp features or discontinuities without extra tuning. By allowing them to adapt their weighting, they aim to improve accuracy in those difficult spots.

Vera: It also mentioned a comparison against traditional methods like the Kurganov-Tadmor scheme and showed that their SA-PINN-ACTO produced solutions with relative L2 errors of order ten to the power of negative three when compared to the KT solution on smooth problems <ref:2602.16117#pg2>.

Jocelyn: But they also found that when dealing with shock initial conditions, the PINN smoothed them out a bit, reaching error levels in the ten to negative two or ten to negative one range, which tells us something about how well it handles those sharp jumps.

Subrahmanyan: The comparison between these two approaches—the KT method and their SA-PINN-ACTO—gives some strong evidence about the reliability of the PDE solutions they are getting from the PINN.

Vera: So to wrap up, what does this paper actually mean for us on a day-to-day level? It suggests that physics-informed machine learning can be a flexible tool for solving these kinds of relativistic fluid dynamics problems.

Jocelyn: It points toward using these types of neural networks not just for simple curve fitting, but as actual solvers for complex physical laws when the equations are complicated.

Subrahmanyan: For someone looking at the bigger picture in astrophysics or heavy-ion collisions, it means we have a new framework to explore how matter and charge behave in extreme environments where relativity is key.

Vera: It’s a way to bridge the gap between traditional numerical methods and machine learning tools for these kinds of problems. The SA-PINN-ACTO framework is presented as a viable tool for this job, which motivates more research into physics-informed machine learning in relativistic viscous hydrodynamics.

Jocelyn: So while the KT method stays faster and handles those discontinuous solutions really well, the PINN approach offers that flexibility because you don't have to rewrite the whole equation into a flux-conservative form first.

Subrahmanyan: Future work they mention is focusing on accelerating how fast these PINNs can be trained and trying to improve their performance specifically near discontinuities using hybrid finite volume and machine learning schemes.

Vera: So that’s where we’ll leave it for now, but the idea is that this paper shows a promising path forward for applying these tools to things like astrophysics or maybe even heavy-ion collisions.

Jocelyn: It's cool to see how they combine different mathematical ideas—like algebraic transforms and self-adaptation—to tackle a problem in fluid dynamics.

Subrahmanyan: It gives us a new way to model the transport of conserved charges in fluids under relativistic conditions, which is fundamental stuff for understanding how things evolve in space and time.

The paper's summary: Vera: So, to recap, this paper is taking that complicated relativistic diffusion equation and rewriting it in a way that numerical solvers can actually handle, using these physics-informed neural networks, or PINNs.

Jocelyn: Right. Basically, they’re not trying to solve the whole thing perfectly with just traditional math; they're letting an AI learn how to solve it by looking at the physical rules we already know.

Subrahmanyan: That reformulation into a flux-conservative form is pretty smart because it sets up the conservation law structure, which makes it much easier for any numerical scheme to work with.

Vera: It’s about making the math cleaner so we can actually run simulations of relativistic fluids, especially near equilibrium where things are simpler.

Jocelyn: The big part they show is this SA-PINN-ACTO framework. It uses self-adaptive techniques combined with a specific way to enforce the starting and ending conditions exactly, without having to mess with the network's structure itself.

Subrahmanyan: That’s interesting because it lets the AI spend its processing power focusing only on figuring out where the solution is actually hard to approximate, instead of wasting time trying to satisfy boundary conditions every single step.

Vera: But they did have some trade-offs, right? When they tested those shock initial conditions, even with this advanced PINN setup, it smoothed things out a bit compared to the traditional finite volume methods like Kurganov-Tadmor.

Jocelyn: Yeah, the paper shows that the SA-PINN-ACTO got a relative L2 error of about one thousandths when compared to those finite volume solutions on smooth problems. That's pretty close for a neural network tackling something this complex.

Subrahmanyan: The caveat is that for discontinuities, which are common in real astrophysical shocks, the PINN version reached errors in the ten to the negative two or ten to the negative one range. So it handles those sharp changes less naturally than a dedicated finite volume scheme does.

Vera: It really highlights that while these machine learning tools are flexible—they don't need you to rewrite equations into flux-conservative form—there’s still a trade-off in accuracy depending on the physical setup you’re testing it on.

Jocelyn: So, for someone just listening, this means we have a new kind of tool that lets us model relativistic transport without having to invent entirely new mathematical frameworks every single time we want to see how matter moves.

Subrahmanyan: It opens up possibilities for applying these PINNs in areas like heavy-ion collisions or even studying the structure of accretion flows around black holes, where those extreme conditions are common.

Vera: Exactly. It’s about giving us a flexible way to test ideas quickly, but we still need to figure out how to push that accuracy past those discontinuities for the next generation of simulations.

The paper's improvements: Tom: So, we’re looking at how they plan to make this AI solver even better based on their findings.

Vera: The paper suggests they can refine how the network learns by using a self-adaptive technique that weights different parts of the solution differently based on where it needs more attention.

Jocelyn: So instead of treating every part of the simulation equally, the AI learns to pay closer attention to regions where it’s struggling to get an accurate answer.

Subrahmanyan: That is a practical way to handle those sharp gradients and discontinuities we talked about earlier by essentially telling the AI where its focus should be.

Vera: They're also using a loss function that gets redefined; they're weighting each term in the equation by a learned factor, which means the network learns what matters most for the final answer.

Jocelyn: That sounds like they’re essentially giving the neural network a built-in way to judge its own performance on different parts of the physics.

Subrahmanyan: This helps address one of those main hurdles with PINNs, which is making sure they don't just get stuck in an area where the math is easy and ignore a region where the physics are really complex.

Vera: It’s about improving robustness so that when you feed it a messy physical scenario, like a real astrophysical shock, it can adjust its learning strategy instead of just producing a blurry average.

Jocelyn: So, the implication is that this framework could be much more reliable for modeling dynamic processes where things are changing very rapidly.

Subrahmanyan: If we can get better handling of those sharp features, it helps us move closer to applying these methods in areas where shocks and rapid changes are the norm, like in high-energy astrophysics.

Vera: It’s a step toward making physics-informed machine learning tools more versatile for simulating complex fluid dynamics across a wider range of physical conditions.

Jocelyn: So the next thing we need to watch is how fast the AI can actually learn all this, because training these complex models can still take a significant amount of time.

Conclusion: Vera: So, to wrap up this discussion on "Solving BDNK diffusion using physics-informed neural networks," we’ve seen how this AI framework tries to bridge traditional numerical methods and machine learning for relativistic fluid dynamics.

Jocelyn: Right. It tackles that complex BDNK equation by turning it into a flux-conservative form and then solving it using a combination of finite volume schemes and these self-adaptive PINNs.

Subrahmanyan: I just think the connection between the theory—the conservation laws—and the practical solution provided by this AI is what makes this framework useful for understanding how matter moves in extreme cosmic environments.

Vera: It really does. The paper shows that even with these limitations, like smoothing out shocks a bit, we’re getting results that are comparable to established finite volume methods on smooth data sets.

Jocelyn: And the point is that this approach gives us flexibility; you don't have to rewrite everything just because you want to use an AI solver for a problem.

Subrahmanyan: It suggests that physics-informed learning can be a viable tool for exploring relativistic viscous hydrodynamics, which has big implications if we want to model things like the early universe or heavy-ion collisions accurately.

Vera: Exactly. We’ve got this new way of looking at how we can use neural networks to solve these kinds of tough physical laws.

Jocelyn: And while they pointed out that training speed is still a challenge, it’s a solid foundation for future research into applying these tools across different areas of astrophysics.

Subrahmanyan: Moving forward, the real test will be seeing if we can push that accuracy past those discontinuities, which is where the true complexity lies in real astrophysical systems.

Illinois Center for Advanced Studies of the Universe and Department of Physics, University of Illinois Urbana-Champaign · Department of Physics, North Carolina State University

nucl-th, astro-ph.HE, gr-qc

Submitted: 2026-02-18

Updated: 2026-02-18

Comments: 30 pages, 11 figures, 1 table

Journal ref: Phys. Rev. D 114, 076013 (2026)

DOI: 10.1103/shpm-ksq9

Code: https://github.com/vchomalicastro/1-1D-BDNK-diffusion-simulations

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 74/100

The gist: The gist Solving BDNK diffusion using physics-informed neural networks reformulates the relativistic BDNK diffusion equation into a flux-conservative form and solves it in (1+1)D using both a

Key concepts

BDNK Diffusion
This theory simplifies relativistic viscous hydrodynamics by focusing on diffusion, which governs how matter and charge transport in fluids relax toward equilibrium. It uses quantities like density, chemical potential, temperature, and fluid velocity to describe the system's behavior.
Flux-Conservative Formulation
Rewriting equations in this form means expressing them as changes in conserved quantities (fluxes) rather than just rates of change. This structure is mathematically useful for numerical methods because it ensures that conserved quantities are properly accounted for across different parts of the domain.
SA-PINN-ACTO
This is a physics-informed neural network enhanced with self-adaptive techniques and an algebraic transform (ACTO). The self-adaptive part helps the network focus on difficult areas, while the ACTO transform enforces initial and boundary conditions exactly, allowing the network to concentrate only on solving the main differential equation.

Terminology

Summary

The gist Solving BDNK diffusion using physics-informed neural networks reformulates the relativistic BDNK diffusion equation into a flux-conservative form and solves it in (1+1)D using both a second-order Kurganov-Tadmor finite volume scheme and a self-adaptive PINN framework, SA-PINN-ACTO, which combines self-adaptive techniques with an algebraic transform for exact enforcement of boundary conditions.

Theory of BDNK Diffusion

The paper reformulates the relativistic BDNK diffusion equation in flux-conservative form and solves the resulting equations in (1+1)D using both a secondorder Kurganov-Tadmor finite volume scheme and physics-informed neural networks (PINNs) The BDNK theory offers a simpler and more tractable approach to relativistic viscous hydrodynamics, at least near equilibrium Diffusion plays a central role in relativistic hydrodynamics as it governs the relaxation of conserved-charge imbalances and determines how matter and charge are transported in fluids In equilibrium, the only quantities are the density n, as well as the chemical potential µ, temperature T, and the background fluid velocity uµ Dissipative terms can be included through a gradient expansion in terms of derivatives of the chemical potential divided by temperature The conserved current in equilibrium takes the form Jµeq = nuµ In this work, we consider a “probe approximation” where the temperature and fluid velocity are taken to be background fields, and backreactions between the diffusion and energy-momentum sector are neglected for simplicity

Flux-Conservative Formulation

For numerical applications, it is useful to write the equations of motion in flux-conservative form, meaning that they take the form ∂tN + ∂iJi = Φ, for some set of densities N, fluxes Ji, and sources Φ It might seem that the equation of motion as written already takes this form since it comes from a conservation law ∂tJ0 + ∂iJi = 0 To proceed, we would like to write both the fluxes and sources as functions of these densities This can be done by writing the BDNK conserved current of Eq. (5) in terms of Nµ as Jµ = nuµ + (−λT uµuν + σT ∆µν) Nν Using these relations and the equations derived above, we can write the equations of motion in flux-conservative form ∂t (J0 α Ni) + ∂j (Jⱼ0 − δj i N0) = (0 − N0 0)

Numerical Methods

The paper employs two independent numerical methods: a second-order KT finite-volume solver and a physics-informed neural network (PINN) On the finite volume side, we implemented an SSP-RK2 Kurganov-Tadmor scheme and verified its convergence for both smooth and discontinuous initial data The semi-discrete form of the (1 + 1)D KT scheme is d t qi = − Hi−1/2 − Hi−2/2 ∆x + S[qi] For the SA-PINN-ACTO, we introduce a self-adaptive PINN technique with an algebraic transform for exact enforcement of initial and periodic boundary conditions through an algebraic transform of the network’s raw output This handling of initial and boundary conditions via post-processing allows the neural network to focus solely on minimizing the PDE residual

Results and Comparison

Across the set of smooth test problems, the SA-PINN-ACTO produced solutions that closely match the convergent KT solutions, with spacetime relative L2 errors of order 10−3 For the shock initial condition, however, while the KT method correctly preserved the discontinuities, the PINN smoothed them out and reached error levels in the 10−2–10−1 range The agreement between the two numerical methods, KT and SA-PINN-ACTO, gives credence to the reliability of the predicted PDE solutions

Conclusions

The paper concludes that the SA-PINN-ACTO framework provides a viable and flexible tool for solving PDEs and motivates further exploration of physics-informed machine learning in relativistic viscous hydrodynamics The close agreement between the two fundamentally different approaches indicates that the SA-PINN-ACTO framework introduced in this paper provides a viable and flexible tool for solving PDEs and motivates further exploration of physics-informed machine learning in relativistic viscous hydrodynamics with potential applications in heavy-ion collisions and astrophysics The KT method remains significantly faster and more accurate, especially for discontinuous solutions, while the PINN approach offers greater flexibility by not requiring rewriting the equations in flux-conservative form Future work should focus on accelerating the training process of PINNs and improving their performance near discontinuities through hybrid finite volume-machine learning schemes

How it works

The SA-PINN-ACTO method is a combination of techniques that enhances the standard PINN approach It incorporates the self-adaptive collocation weights technique of the SA-PINN [88] with an algebraic (also known as hard) enforcement of initial and periodic boundary conditions via only an output transform, which we call the ACTO transform This guarantees that those conditions are exactly satisfied without requiring any modification to the network architecture, while allowing the network to focus its attention on the regions of the domain where it is harder to properly approximate the solution to a given PDE, as well as having to minimize only the PDE residual

The SA-PINN-ACTO architecture takes spacetime inputs (t, x) and predicts normalized raw outputs Jˆ0/sJ0 and ˆα/sα These are first denormalized, then passed through the IC-enforcing transform in Eq. (52), and then through the BC-enforcing transform in Eq.

Improvements for AI systems

  1. Bold header: Enhanced PDE Solving for Relativistic Hydrodynamics

The improved system can solve relativistic viscous hydrodynamic equations in (1+1)D using a physics-informed neural network (PINN) with an algebraic enforcement of initial and periodic boundary conditions via the SA-PINN-ACTO framework. This allows the network to focus solely on minimizing the PDE residual, which is beneficial for handling complex, non-trivial BDNK backgrounds.

  1. Bold header: Robust Handling of Discontinuities

The system can accurately approximate solutions containing sharp gradients and discontinuities by using a self-adaptive PINN technique (SA-PINN) combined with network-internal normalization. This mechanism allows the loss function to be redefined as R2 PDE,i ≡ ∂tJ0θ + ∂xJ xθ2 / sJ0 + ∂tαθ + N0,θ2 / sα, where each term is weighted by a learned factor λi.

  1. Bold header: Comparison and Validation of Numerical Methods

The system can rigorously compare the SA-PINN-ACTO approach against traditional finite volume methods like the Kurganov-Tadmor (KT) scheme. This comparison provides quantitative metrics, such as relative L2 error between the SA-PINN-ACTO and KT solutions of order 10−3, validating its performance across smooth and discontinuous initial data.

Abstract

In this work, we reformulate the relativistic BDNK (Bemfica-Disconzi-Noronha-Kovtun) diffusion equation in flux-conservative form, and solve the resulting equations in (1+1) D using both a second-order Kurganov-Tadmor finite volume scheme and physics-informed neural networks (PINNs). In particular, we introduce the SA-PINN-ACTO framework, which combines the self-adaptive PINN technique with an exact enforcement of initial and periodic boundary conditions through an algebraic transform of the network's raw output, allowing the network to focus solely on minimizing the PDE residual. We test both approaches on smooth and discontinuous initial data, for both trivial and dynamically evolving velocity and temperature BDNK backgrounds, and for two characteristic speeds. The SA-PINN-ACTO method matches the converged Kurganov-Tadmor solutions for smooth profiles, while for discontinuous profiles the errors increase, reflecting an expected limitation of PINNs near sharp gradients.

Sources

Related papers