Noise-balanced multilevel on-the-fly sparse grid surrogates for coupling Monte Carlo models into continuum models with application to heterogeneous catalysis

This paper proposes a novel noise-balanced, multilevel, on-the-fly sparse grid interpolation approach to efficiently create surrogate models for high-fidelity Monte Carlo simulations, overcoming challenges like sampling noise and the curse of dimensionality in multiscale applications such as heterogeneous catalysis.

The Gist

Imagine you are trying to predict how a massive, complex forest will grow over the next fifty years.

To be perfectly accurate, you would need to track every single leaf, every insect, and every drop of rain on every single tree. This is a "High-Fidelity Model" (the microscopic view). It is incredibly accurate, but it is also impossible to do because the sheer amount of data would crash even the world's most powerful supercomputers.

Instead, scientists usually use a "Continuum Model" (the macroscopic view). This is like looking at the forest from a satellite. You don't see individual leaves; you see "greenness" and "moisture levels." It’s fast and easy, but it’s just a guess. The problem is that the "satellite view" needs to know how the "leaf view" works to be accurate. If the leaves change how they breathe, the satellite view needs to update its math.

This paper introduces a clever way to bridge that gap.

The Problem: The "Noisy" Messenger

The researchers are specifically looking at Monte Carlo models. Think of a Monte Carlo model like a person trying to predict the weather by flipping a coin millions of times. It’s a great way to simulate randomness (like chemical reactions), but it has a major flaw: Noise.

If you flip a coin 10 times, you might get 7 heads just by luck. That’s "noise." If you use that "7 heads" result to predict the weather, your prediction will be wrong. To get a "clean" answer, you have to flip the coin millions of times, which takes forever.

So, scientists have two bad choices:

1. The Slow Way: Flip the coin a billion times to get a perfect answer (too slow).
2. The Fast but Wrong Way: Flip the coin 10 times and hope for the best (too inaccurate).

The Solution: The "Smart Sketch Artist"

The authors created a new tool called ML-OTF-SG (Multilevel On-the-Fly Sparse Grid). Think of this as a Smart Sketch Artist who is helping the satellite view.

Here is how the Sketch Artist works:

1. The "On-the-Fly" Sketch (Efficiency)
Instead of drawing a map of the entire forest before the simulation starts, the artist only draws the parts of the forest the satellite is actually looking at right now. If the satellite is looking at a specific valley, the artist only sketches that valley. This saves a massive amount of time.

2. The "Noise-Balancer" (The Secret Sauce)
This is the most brilliant part. The artist knows that the "coin-flippers" (the Monte Carlo models) are noisy.

• If the artist is sketching a very simple, flat area of the forest, they don't ask the coin-flippers to work hard. They just take a quick, "noisy" glance.
• But, if the artist is sketching a complex, jagged mountain peak where accuracy is vital, they tell the coin-flippers: "Stop! Don't just flip the coin 10 times. Flip it a million times! I need a crystal-clear answer here because this part matters."

By automatically deciding where to be precise and where to be quick, the artist balances "drawing time" with "accuracy."

Why does this matter? (The Real-World Application)

The researchers tested this on Heterogeneous Catalysis. This is a fancy way of saying "how chemicals react on the surface of a solid material" (like the catalytic converter in your car that cleans exhaust fumes).

In these reactions, atoms are dancing around on a surface in a chaotic, random way. Predicting how a whole factory reactor will behave based on those tiny atomic dances is incredibly hard.

The researchers proved that their "Smart Sketch Artist" could predict how these chemical reactors work with incredible accuracy, using a fraction of the computing power previously required. It allows scientists to design better catalysts and cleaner industrial processes without needing a computer the size of a planet.

Summary in a Nutshell

• The Goal: Connect tiny, random atomic movements to big, industrial chemical processes.
• The Obstacle: Tiny models are too slow; big models are too "dumb"; and random models are too "noisy."
• The Innovation: A smart mathematical "sketch artist" that only draws what is necessary and only asks for high-accuracy data when the math gets complicated.

Technical Summary

Technical Summary: Noise-Balanced Multilevel On-the-Fly Sparse Grid Surrogates

Authors: Tobias Hülser and Sebastian Matera (Fritz Haber Institute of the Max Planck Society)
Date: February 12, 2026

---

1. The Problem: The Multiscale Bottleneck

The paper addresses a fundamental challenge in Heterogeneous Multiscale Modeling (HMM): coupling high-fidelity microscopic models (HFMs)—specifically stochastic Kinetic Monte Carlo (kMC) simulations—into macroscopic continuum models (e.g., reactor simulations).

Two primary obstacles prevent direct coupling:

• Computational Intractability: HFMs are extremely expensive to evaluate. In a continuum simulation, the HFM must be queried repeatedly, making direct coupling computationally impossible.
• The Noise-Discretization Dilemma: Unlike deterministic HFMs, kMC models are stochastic and produce "noisy" data. If a surrogate model is trained on noisy data, the sampling error can corrupt the surrogate, even if the discretization (the number of data points) is high. Conversely, spending excessive CPU time to reduce noise at every point is inefficient if the surrogate itself is too coarse to capture the underlying function.
• Curse of Dimensionality: As the number of chemical species or macro-variables increases, the number of required training points for traditional grid-based surrogates grows exponentially.

2. Methodology: ML-OTF-SG with Noise Balancing

The authors propose a novel approach called Multilevel On-the-Fly Sparse Grid (ML-OTF-SG) interpolation, enhanced by a noise-balancing strategy.

A. Multilevel On-the-Fly Sparse Grids (ML-OTF-SG):

• Sparse Grids (SG): Instead of a full tensor grid, the method uses sparse grids to mitigate the curse of dimensionality, allowing for accurate interpolation in higher dimensions with fewer points.
• On-the-Fly (OTF) Construction: The surrogate is not built entirely upfront. Instead, it is constructed locally as the continuum solver queries specific points. This exploits the fact that the continuum solution typically only visits a small, low-dimensional subdomain of the total input space.
• Multilevel Strategy: The approach starts with a coarse interpolation level and iteratively refines it, using the previous solution as a starting point to save computational effort.

B. Noise-Balancing Strategy:
The core innovation is the mathematical decomposition of the total expected error into two parts:

1. Discretization Error: The error from using a finite number of grid points.
2. Sampling (Noise) Error: The error from the stochastic nature of the kMC data.

The authors derive a rule to determine the optimal number of Monte Carlo samples ($N_{l,i}$) for each specific grid point. By setting the sampling effort such that the noise error is roughly proportional to the discretization error (controlled by a single hyperparameter, the maximum level $L$), they ensure that computational resources are not wasted on "over-sampling" points in a coarse surrogate or "under-sampling" points in a fine one.

3. Key Contributions

• Error-Balanced Coupling: A framework that treats sampling effort as a controllable variable to match the accuracy of the interpolation grid.
• Single Hyperparameter Control: The entire complexity of the surrogate—from the accuracy of the interpolation to the required precision of the kMC data—is controlled by one parameter: the maximum sparse grid level $L$.
• Self-Consistency: The on-the-fly construction is designed to be self-consistent, ensuring that the surrogate behaves like a fixed, deterministic function during the iterative steps of a nonlinear solver, which is critical for numerical stability.
• Nonlinear Transformation: The use of an inverse hyperbolic sine ($\text{arcsinh}$) transformation to handle the high nonlinearity and wide dynamic ranges of reaction rates, preventing high-rate values from "poisoning" the interpolation of lower-rate regimes.

4. Results and Applications

The methodology was tested on three realistic heterogeneous catalysis scenarios:

1. Semi-analytical Test (Reverse Water Gas Shift on Cu/ZnO): Demonstrated that the error converges predictably with both grid refinement and increased sampling, following the theoretical $O(N^{-1/2})$ scaling of Monte Carlo methods.
2. CO-oxidation on $\text{RuO}_2$: A 2D first-principles kMC model. The surrogate achieved high accuracy with extremely low computational cost (less than two core hours for the highest level), proving the efficiency of the OTF approach.
3. Water-Gas Shift on $\text{Pt}(111)$: A high-dimensional, highly nonlinear, and noisy 1p-kMC model. The method successfully handled the increased dimensionality and noise, allowing for a temperature scan where data from previous steps was recycled to minimize new sampling costs.

5. Significance

This work provides a robust, "black-box" compatible bridge between the atomic scale and the macroscopic scale. It allows researchers to perform complex reactor simulations using high-fidelity kinetic models that were previously considered too expensive or too noisy to couple directly. The ability to control accuracy via a single parameter makes it a highly practical tool for engineers and chemists working in multiscale modeling and computational fluid dynamics (CFD).