Generative marginal learning for finite-size molecular simulation with application to stress calculation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Generative marginal learning for finite-size molecular simulation with application to stress calculation".
Mira: Here's the gist of the paper. Generative marginal learning (GML) is a framework designed to learn the equilibrium marginal distribution of a retained atomic region from large-system simulations using denoising and force-alignment training, which substantially improves pair structure and local stress accuracy compared to small periodic molecular dynamics simulations Problem Definition The aim is to calculate macroscopic physical quantities, such as stress or strain, from the atomistic information in an interior region of a large crystalline system at equilibrium The core idea is to learn an effective potential energy for the interior system that allows sampling the same marginal distribution as full-system simulation This involves defining a retained region R such that C ⊂ R ⊂ omega, where C is the observation core and R satisfies C ⊂ R ⊂ omega The full system configuration q can be expressed in terms of a homogeneous deformation and atomic displacements How the training signal is built The framework employs a temperature-conditioned diffusion model to learn the conditional marginal distribution pR(· A, T) The network learns a residual correction sθ(zτ, τ, T) that approximates the score function sτ (zτ; T) Two primary training objectives are used to learn this score: 1.
Kai: Denoising loss (LN), which is motivated by the relationship between the conditional score and the Gaussian noise introduced in the forward process, defined as LN (θ) = EP∥εbθ − ε∥2D 2.
Mira: Force-alignment loss (LF), which utilizes a mean-force relation derived from Eq.
Kai: (11) to provide an alternative training objective, defined as LF (θ) = EP∥Rθ − RF∥2D Where the gains come from The benefits of GML are demonstrated through its application in estimating the spatially averaged Irving–Kirkwood stress near a vacancy defect Numerical tests show that GML substantially improves pair structure and local stress accuracy compared to small periodic molecular dynamics simulations, mitigating finite-size effects in calculating local molecular observables Specifically, force-alignment training gives the smallest stress discrepancy among the five comparison methods at all nine temperatures Sampling Performance and Further Acceleration The GML approach achieves a throughput of 2294.72 saved configurations per second using a GPU, which is higher than that of large-cell MD, although this does not establish a lower cost per statistically independent sample The method allows for the estimation of the mean core stress by sampling from the generated density p aR(· A, T), which preserves the equilibrium mean Irving–Kirkwood (IK) stress in the observation core Future work will focus on accurate score estimation in the zero-noise limit and extensions to a broader range of deformation states and defect types Application to Stress Calculation at Finite Temperature The exact marginal sampling preserves the mean core stress by applying Eq.
Title and authors: Mira: Now let's talk about the title itself, "Generative marginal learning for finite-size molecular simulation with application to stress calculation." It tells you immediately what this is about: using generative learning for sampling a part of a simulation to figure out stress.
Kai: And the authors are Cheng Yuana and their team from Wuhan University. They are taking this concept, which involves training conditional diffusion models, and applying it specifically to calculate things like the spatially averaged Irving–Kirkwood stress near defects.
Lev: It's interesting that they mentioned including methods like the quasicontinuum method and the heterogeneous multiscale method in their background reading. That suggests they are building on existing physics models while using this AI to handle the sampling problem.
The paper's summary: Kai: So what is the actual mechanism they introduce? They propose a framework called Generative Marginal Learning, or GML. This lets you learn the conditional marginal distribution of a retained region, which means you can generate configurations that look statistically correct for that area.
Mira: And to do this, they derive two key things: a marginal effective potential and an associated conditional mean-force identity. These are what enable both denoising loss and force-alignment training to work effectively.
Lev: That conditional mean-force identity is crucial because it directly gives you the forces conditioned on the retained coordinates, which is what you need for physics calculations like stress estimation.
Kai: And their application is estimating the spatially averaged Irving–Kirkwood stress near a vacancy defect. They claim this substantially improves pair structure and local stress accuracy when compared to small periodic molecular dynamics simulations.
The paper's improvements: Mira: The authors point out that this approach helps mitigate finite-size effects when calculating local molecular observables, which is a big deal because those artifacts usually creep in with small systems.
Kai: They specifically highlight that force-alignment training gives the smallest stress discrepancy among the five comparison methods they tested across nine different temperatures. That's a concrete result showing where this method shines compared to other sampling techniques.
Lev: If you were trying to run this on real hardware, I'd be looking at how much overhead the AI part adds. The paper mentions that for equilibrium constitutive calculations, the statistical objective is to recover the local configuration distribution needed for those averages of quantities like stress.
Kai: They also show that this method can achieve a throughput of two thousand two hundred ninety-four point seven two saved configurations per second when run on a GPU, which is faster than large-cell molecular dynamics, even though that doesn't necessarily mean it's cheaper per independent sample.
Conclusion: Mira: So to wrap up the GML paper, the main takeaway is that you can use generative models to learn how to sample the correct distribution of a retained region without needing periodic boundary conditions on that region itself. It preserves equilibrium properties while giving you access to local information.
Kai: And they showed this works by deriving that marginal effective potential and using both denoising and force-alignment training, with force alignment being particularly good for stress accuracy early on.
Lev: For real hardware deployment, the conditioning of forces on the retained coordinates is what makes it physically meaningful for things like calculating stress, provided you can get those accurate conditional forces reliably.
Mira: It seems like this framework offers a direct way to sample the exact distribution needed to evaluate averages of quantities such as stress in regions near defects. It's a useful tool for incorporating the statistical influence of the surrounding material into your local description.
Kai: This paper, "Generative marginal learning for finite-size molecular simulation with application to stress calculation," shows that this approach is an effective way to get better pair structure and local stress accuracy than what you get from small periodic molecular dynamics simulations.
School of Artificial Intelligence, Wuhan University
physics.comp-ph, cond-mat.stat-mech
Submitted: 2026-10-18
Updated: 2026-11-05
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 82/100
The gist: Generative marginal learning (GML) is a framework designed to learn the equilibrium marginal distribution of a retained atomic region from large-system simulations using denoising and force-alignment…
Key concepts
- Generative Marginal Learning (GML)
- A framework that uses deep learning to learn the equilibrium marginal distribution of a retained atomic region from large-scale simulations. It creates configurations for this region without needing periodic boundary conditions, capturing the statistical influence of the surroundings.
- Denoising Loss (LN)
- A training objective where an AI network learns to predict and remove noise from a process. This loss function guides the model to accurately approximate the conditional marginal distribution $p_R(ullet | A, T)$, which is crucial for generating realistic local atomic configurations.
- Force-Alignment Loss (LF)
- A training objective that uses a mean-force relation derived from physical principles. It trains the model to generate configurations where the learned forces align correctly with the true mean force, leading to better accuracy in stress calculations.
Terminology
Summary
Generative marginal learning (GML) is a framework designed to learn the equilibrium marginal distribution of a retained atomic region from large-system simulations using denoising and force-alignment training, which substantially improves pair structure and local stress accuracy compared to small periodic molecular dynamics simulations
Problem Definition
The aim is to calculate macroscopic physical quantities, such as stress or strain, from the atomistic information in an interior region of a large crystalline system at equilibrium The core idea is to learn an effective potential energy for the interior system that allows sampling the same marginal distribution as full-system simulation This involves defining a retained region R such that C ⊂ R ⊂ omega, where C is the observation core and R satisfies C ⊂ R ⊂ omega The full system configuration q can be expressed in terms of a homogeneous deformation and atomic displacements
How the training signal is built
The framework employs a temperature-conditioned diffusion model to learn the conditional marginal distribution pR(· A, T) The network learns a residual correction sθ(zτ, τ, T) that approximates the score function sτ (zτ; T) Two primary training objectives are used to learn this score:
-
Denoising loss (LN), which is motivated by the relationship between the conditional score and the Gaussian noise introduced in the forward process, defined as LN (θ) = EP∥εbθ − ε∥2D
-
Force-alignment loss (LF), which utilizes a mean-force relation derived from Eq. p. 9
(11) to provide an alternative training objective, defined as LF (θ) = EP∥Rθ − RF∥2DWhere the paper says it
(23) Here, all network outputs are evaluated at (zτ , τ, T), and ∥·∥ denotes the Euclidean norm. Force-alignment loss When full-system forces are available, the mean-force relation in Eq.
Where the gains come from
The benefits of GML are demonstrated through its application in estimating the spatially averaged Irving–Kirkwood stress near a vacancy defect Numerical tests show that GML substantially improves pair structure and local stress accuracy compared to small periodic molecular dynamics simulations, mitigating finite-size effects in calculating local molecular observables Specifically, force-alignment training gives the smallest stress discrepancy among the five comparison methods at all nine temperatures
Sampling Performance and Further Acceleration
The GML approach achieves a throughput of 2294.72 saved configurations per second using a GPU, which is higher than that of large-cell MD, although this does not establish a lower cost per statistically independent sample The method allows for the estimation of the mean core stress by sampling from the generated density p aR(· A, T), which preserves the equilibrium mean Irving–Kirkwood (IK) stress in the observation core Future work will focus on accurate score estimation in the zero-noise limit and extensions to a broader range of deformation states and defect types
Application to Stress Calculation at Finite Temperature
The exact marginal sampling preserves the mean core stress by applying Eq. (7) to the aligned coordinates, yielding ⟨σC⟩omega = Z σC(x a, T)p aR(x aA,T) dx a = Z e−βV a(x a;A,T) dx a The estimator used in experiments is σC(q, T) = −kBT NC(q)CI − 1/CX i<j fij ⊗ rijλij,C (33) This velocity average preserves the mean kinetic contribution and removes the fluctuations associated with instantaneous velocities The required boundary forces are therefore the original pair forces fij on these bonds, evaluated from the specified pair potential
Conclusion
GML develops a generative model for a retained atomic region that captures the statistical influence of its surroundings without imposing periodic boundary conditions on the region itself The two training strategies exhibit complementary strengths, with force alignment achieving accurate stress estimates earlier in training and denoising yielding better generalization at convergence These results support retained-region marginal learning as an effective approach to incorporating the statistical influence of the surrounding material while preserving an explicit atomistic description within the region of interest The framework extends to other equilibrium ensembles by using their macroscopic constraints as conditioning variables The learned score can also supply the conditional atomic forces needed for stress evaluation
Acknowledgments
This work was supported by the National Natural Science Foundation of China (No. 12301558) The paper is supported by the National Natural Science Foundation of China (No. p. 0 Generative marginal learning for finite-size molecular
simulation with application to stress calculation
Cheng Yuana,∗
aSchool of Artificial Intelligence, Wuhan University, Wuhan, 430072, Hubei, China
Abstract
We propose generative marginal learning (GML), a framework for equilibrium
sampling and local observable estimation in molecular simulations. By training a conditional diffusion model on large-system simulations, GML learns
the marginal distribution of a retained region, enabling configuration generation without explicitly simulating the surrounding environment or imposing
periodic boundary conditions on the retained region.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved AI systems can learn marginal distributions of retained regions without imposing periodic boundary conditions on those regions, enabling configuration generation
without explicitly simulating the surrounding environment or imposing periodic boundary conditions on the retained region.
-
The system can derive a marginal effective potential and associated conditional mean-force identity to support both denoising and force-alignment training, facilitating more accurate sampling of local configurations.
-
The AI system can estimate the spatially averaged Irving–Kirkwood stress near a vacancy defect by using
generated configurations and an interaction buffer,
which is shown tosubstantially improve pair structure and local stress accuracy compared with small periodic molecular dynamics simulations.
-
The system can utilize force-alignment training, which provides a score label defined as s a R(x a; T) = -betaEomega[f corr R x a], to obtain physical forces that are conditioned on the retained coordinates.
-
The AI system can calculate the mean core stress using an estimator derived from Eq. (35), which
is unbiased when p a R,theta = p a R,
thus preserving the equilibrium mean Irving–Kirkwood stress in the observation core. -
The system can achieve higher configuration-generation throughput by running on a GPU, achieving
a throughput of 2294.72 saved configurations per second, compared with 460.88 for large-cell MD.
Abstract
We propose generative marginal learning (GML), a framework for equilibrium sampling and local observable estimation in molecular simulations. By training a conditional diffusion model on large-system simulations, GML learns the marginal distribution of a retained region, enabling configuration generation without explicitly simulating the surrounding environment or imposing periodic boundary conditions on the retained region. We derive a marginal effective potential and the associated conditional mean-force identity, which provide the basis for both denoising and force-alignment training. As an application, we estimate the spatially averaged Irving--Kirkwood stress near a vacancy defect. Numerical tests on a two-dimensional Lennard--Jones crystal and three-dimensional BCC Fe under combined tension and shear demonstrate improved local stress accuracy relative to small periodic molecular dynamics simulations.
Sources
- Score-Based Generative Modeling through Stochastic Differential Equations
- Score-based Generative Modeling in Latent Space
- Denoising Diffusion Probabilistic Models
- Target Score Matching
- Latent Target Score Matching, with an application to Simulation-Based Inference
- Decoupled Weight Decay Regularization
Related papers
- Gradient-Based Inverse Design of Free-Energy Landscapes with Diffusion Models
- Mixture of Polyconvex Neural Potentials for Parametric Hyperelasticity: Towards Foundation Material Models
- Development and large-scale benchmarks of a protein--ligand absolute binding free energy toolkit
- Spin disorder competing with positional symmetry breaking governs the metal-insulator behavior in oxide paramagnets
- Investigation of the Asymptotic Properties of Active Impedance in Large Finite Array Antennas
- Automatic generation of exchange-correlation response kernels