Local kinetic sensors for adaptive mesh and algorithm refinement

This paper introduces novel, locally computable refinement sensors based on the one-particle distribution function to enable efficient and scalable adaptive mesh and algorithm refinement for kinetic models like lattice Boltzmann methods, effectively resolving complex features in compressible, turbulent, and non-equilibrium flows.

The Gist

Imagine you are trying to watch a movie, but instead of a single screen, you have a giant, flexible canvas that can change its resolution on the fly.

In the world of fluid dynamics (studying how air, water, and gases move), scientists use powerful computers to simulate everything from weather patterns to jet engines. However, simulating every single molecule in a room is impossible; it would take a supercomputer forever. So, they use a trick: they zoom in (high resolution) only where things are chaotic and messy, and zoom out (low resolution) where things are calm and smooth. This is called Adaptive Mesh Refinement (AMR).

The problem? The computer needs a "smart guide" to know where to zoom in. If the guide is wrong, the computer wastes time zooming in on empty space, or worse, misses a critical event like a shockwave or a tornado forming.

This paper introduces a new, super-smart guide specifically for Kinetic Models (a way of simulating fluids by tracking the movement of individual particles, like a swarm of bees, rather than just treating the fluid as a continuous liquid).

Here is the breakdown of their innovation using everyday analogies:

1. The Old Way: Looking at the Weather Report (Macroscopic Sensors)

Traditionally, computers decide where to zoom in by looking at "macroscopic" variables. Think of this like checking a weather report.

• The Sensor: "Is the wind speed high here?" or "Is the temperature changing fast?"
• The Flaw: To know the wind speed is changing, you have to compare the wind speed at point A with point B. This requires looking at neighbors, which is slow and computationally expensive. It's like trying to figure out if a crowd is moving by asking everyone to shout their location to the person next to them.

2. The New Way: Reading the Mind of a Bee (Local Kinetic Sensors)

The authors of this paper realized that Kinetic Models have a superpower: they don't just know the wind speed; they know the exact position and speed of every single particle (the "one-particle distribution function").

Instead of asking neighbors, they can look at a single particle and ask: "Are you acting weird?"

They created a new set of "sensors" (indicators) that check the particles directly. They fall into two categories:

Category A: The "Local Weather Report" (Class 1 Sensors)

These sensors do the same job as the old macroscopic ones (checking for stress, heat, or speed changes) but do it locally.

• The Analogy: Instead of asking the crowd next to you, you just look at your own shoes. If your shoes are vibrating violently, you know the ground is shaking right here.
• The Benefit: The computer doesn't need to talk to its neighbors to calculate gradients. It just sums up the data it already has sitting in its own memory. This makes the simulation much faster and allows it to run on massive supercomputers with thousands of processors without them getting stuck waiting for each other.

Category B: The "Stress Detector" (Class 2 Sensors - The Real Magic)

These are sensors that cannot be calculated using the old macroscopic methods. They only exist because we are looking at the individual particles.

• The Analogy: Imagine a calm lake. If you drop a stone, the water ripples. But if you look at the molecules of water, you might see that some are vibrating with "non-equilibrium" energy before the ripple even forms.
• What they detect:
  • Knudsen Sensor: Checks if the gas is so thin that particles are flying past each other without bumping (like a sparse crowd vs. a packed mosh pit).
  • Entropy Sensor: Checks how "disordered" the particles are. If the particles are chaotic, the simulation knows to zoom in.
  • Relaxation Sensor: Checks if the particles are struggling to settle down into a calm state. If they are struggling, something interesting is happening nearby.

3. The Result: A Smart, Self-Adjusting Camera

The authors tested these new sensors on two classic fluid problems:

1. The Shock Tube (Sod Shock): Imagine a tube with a wall in the middle. One side has high-pressure gas, the other low. When the wall breaks, a shockwave explodes through.
2. The 2D Riemann Problem: A complex square of gas where four different pressures meet, creating a chaotic dance of shockwaves and vortices.

The Outcome:
When they turned on these new "Local Kinetic Sensors," the computer's "camera" automatically focused its high-resolution lens exactly where the shockwaves and turbulence were happening.

• It ignored the calm, empty spaces (saving massive amounts of computing power).
• It caught the tiny, chaotic details of the gas particles that the old methods might have missed or calculated too slowly.

Why This Matters

Think of this as upgrading from a magnifying glass to a microscope with a brain.

• Old Method: You have to manually move the magnifying glass to every spot to see if it's interesting.
• New Method: The microscope has a built-in AI that says, "Hey, look here! The particles are freaking out! Let's zoom in!" and "Look there, everything is boring, let's zoom out."

This allows scientists to simulate complex, real-world problems—like how air flows over a supersonic jet, how blood flows through a heart, or how pollutants spread in the atmosphere—much faster and with greater accuracy. It paves the way for solving fluid dynamics problems that were previously too expensive or too slow to compute.

Technical Summary

1. Problem Statement

Adaptive Mesh and Algorithm Refinement (AMAR) is a critical technique for improving the computational efficiency and physical accuracy of fluid dynamics simulations. It dynamically adjusts grid resolution (AMR) and solution algorithms (AAR) based on local flow features.

While AMAR is well-established for macroscopic solvers (e.g., Navier-Stokes-Fourier equations using Finite Volume or Finite Difference methods), its application to kinetic models (such as the Discrete Velocity Boltzmann Method and Lattice Boltzmann Method) faces specific challenges:

• Reliance on Macroscopic Sensors: Existing kinetic AMR implementations largely rely on refinement sensors derived from macroscopic variables (density, velocity, temperature) and their gradients. Calculating these gradients often requires non-local operations (e.g., finite differences), which can hinder scalability in parallel High-Performance Computing (HPC) environments.
• Underutilization of Kinetic Data: Kinetic models possess a unique advantage: direct access to the one-particle distribution function ($f$) and its moments. This contains information about non-equilibrium states, entropy production, and internal energy modes that is not explicitly available in macroscopic descriptions.
• Lack of Generic Kinetic Sensors: There was a lack of a comprehensive, generic set of refinement sensors that leverage this kinetic information to drive AMAR efficiently and locally.

2. Methodology

The authors propose a novel framework for local kinetic refinement sensors that operate purely on the discrete distribution functions, avoiding the need for non-local gradient calculations.

A. Kinetic Model

The study utilizes a double-distribution quasi-equilibrium (QE) kinetic model for compressible flows with variable Prandtl numbers ($Pr$).

• Equations: Based on the Boltzmann transport equation with a Bhatnagar-Gross-Krook (BGK) approximation.
• Distributions: It employs two distribution functions:
  1. $f$: Represents mass, momentum, and translational energy.
  2. $g$: Represents internal roto-vibrational energy (allowing for variable specific heat and $Pr \neq 1$).
• Collision Operator: Uses a two-relaxation-time approach ($\tau_1, \tau_2$) to separate fast (viscous) and slow (thermal) modes, enabling the recovery of the Navier-Stokes-Fourier equations with arbitrary Prandtl numbers.

B. Refinement Sensors

The paper categorizes sensors into two classes:

Class 1: Kinetic Sensors with Macroscopic Counterparts
These sensors replicate macroscopic refinement criteria (like viscous stress or heat flux) but compute them locally by summing over discrete populations rather than calculating spatial gradients.

• Examples:
  • Viscous Stress Sensor: Computed via the non-equilibrium pressure tensor ($P^{neq} = \sum v_i \otimes v_i f^{neq}_i$).
  • Rate-of-Compression/Shear/Strain Sensors: Derived from the trace and deviatoric parts of the non-equilibrium tensors.
  • Energy Flux Sensors: Computed directly from non-equilibrium energy flux vectors.
• Advantage: These are purely local operations (summations), making them highly scalable for parallel codes compared to gradient-based macroscopic sensors.

Class 2: Purely Kinetic Sensors
These sensors exploit information unique to kinetic theory, which cannot be reconstructed solely from macroscopic variables.

• Knudsen Sensor: Estimates the local Knudsen number ($Kn$) by measuring the relative deviation of the distribution from equilibrium ($|f^{neq}/f^{eq}|$). This identifies regions where the continuum hypothesis breaks down.
• Non-Equilibrium Sensor: Measures the absolute magnitude of the departure from equilibrium ($|f^{neq}| + |g^{neq}|$).
• Quasi-Equilibrium Sensors: Measures the deviation from the intermediate quasi-equilibrium state ($|f^* - f^{eq}|$), relevant for models with multiple relaxation times.
• Entropy-Based Sensors:
  • Relative H-Sensors: Measures the distance between the current state and equilibrium using the discrete H-function (entropy) or Kullback-Leibler divergence.
  • H-Dissipation Sensors: Quantifies the local rate of entropy production ($-\sigma_H$).
  • Entropic Estimate ($\alpha$): A sensor derived from the Entropic Lattice Boltzmann Method (ELBM) to detect stability risks.

C. AMR Implementation

The sensors were integrated into a Space-Time Adaptive Conservative Mesh Refinement (SAMR) framework:

• Parallelism: Fully parallel implementation using MPI+OMP/GPU.
• Conservation: Strict flux conservation is maintained across grid levels using flux registers and refluxing.
• Sub-cycling: Time steps are adjusted per level based on the refinement ratio to maintain a constant CFL number.

3. Key Contributions

1. Novel Sensor Palette: Introduction of a comprehensive set of refinement sensors specifically designed for kinetic models, categorized into macroscopic-equivalent (local) and purely kinetic types.
2. Local Computation: Demonstration that macroscopic refinement criteria (like stress and heat flux) can be computed more efficiently and scalably using local summations of distribution functions, eliminating the need for non-local gradient stencils.
3. Access to Non-Equilibrium Physics: Provision of sensors that detect non-equilibrium phenomena (high Knudsen numbers, entropy production) which are invisible to standard macroscopic solvers, enabling AMR for rarefied and high-speed flows.
4. Validation and Demonstration: Rigorous validation of these sensors against macroscopic counterparts and successful application to complex flow scenarios (shock tubes, Riemann problems) using a variable Prandtl number solver.

4. Results

The authors validated the sensors using two benchmark cases:

• 1D Sod Shock Tube:
  • Class 1 sensors (kinetic) showed excellent agreement with macroscopic gradient-based sensors across multiple orders of magnitude.
  • Class 2 sensors successfully identified distinct features: The Knudsen sensor highlighted contact discontinuities and shock fronts; entropy sensors highlighted regions of high dissipation.
  • The sensors were smooth and converged with grid refinement.
• 2D Riemann Problem (Case No. 3):
  • The sensors successfully resolved complex interactions between shocks, contact discontinuities, rarefaction waves, and vortices.
  • Class 2 sensors revealed fine-grained structures (e.g., distinct peaks in the Knudsen and entropic estimate sensors) that provided precise localization of flow features.
• AMR Application:
  • When applied to AMR, the kinetic sensors effectively concentrated grid resolution on critical flow features (shocks, interfaces) while keeping smooth regions coarse.
  • The adaptive meshes successfully captured the physics with fewer cells than a uniform grid, demonstrating significant computational efficiency.
  • Different sensors triggered different refinement patterns, allowing users to "steer" the refinement based on specific physical interests (e.g., focusing on entropy production vs. viscous stress).

5. Significance

• Scalability for HPC: By replacing non-local gradient calculations with local summations, the proposed method significantly improves the scalability of kinetic solvers on modern distributed memory machines.
• Bridging Scales: The framework allows for seamless simulation of flows ranging from continuum (Navier-Stokes) to rarefied (non-equilibrium) regimes within a single adaptive framework.
• Future-Proofing: The approach is not limited to the specific BGK model used; the concepts can be extended to other kinetic models, including multiphase flows, non-ideal fluids, and reactive flows.
• Foundation for AI/ML: The paper suggests that these rich, local kinetic data points could serve as inputs for future data-driven optimization of refinement criteria using machine learning.

In conclusion, this work establishes a robust, efficient, and physically consistent foundation for adaptive simulations of complex fluid systems using kinetic models, moving beyond the limitations of macroscopic-only refinement strategies.