Adaptive Equilibration of Molecular Dynamics Simulations

Luciano G. Silvestri, Zachary A. Johnson, Michael S. Murillo · 2025

Read Adaptive Equilibration of Molecular Dynamics Simulations by Luciano G. Silvestri, Zachary A. Johnson, Michael S. Murillo (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-04362.txt.

Introduction

Luciano Silvestri, Zachary Johnson, and Michael Murillo study how to get computer models of matter ready for use faster, in a 2025 paper. Such models start with atoms placed at random, then need a warm-up phase before results can be trusted. That warm-up can waste much computer time if the start is poor.

Using a simple charged plasma as their test case, the team compares seven ways to place atoms at the start. They also test heat controls and a way to forecast the final heat level, so users can stop the warm-up once error bars are small enough. At strong coupling, smart starts cut the warm-up a lot, while weak heat control also helped. The method turns a rule-of-thumb step into one with clear stop rules.

Catalog record

CREATORLuciano G. Silvestri, Zachary A. Johnson, Michael S. Murillo
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE71 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-04362

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Adaptive Equilibration of Molecular Dynamics Simulations Luciano G. Silvestri, Zachary A. Johnson, Michael S. Murillo Abstract We present a systematic framework for shortening and automating molecular dynamics equilibration through improved position initialization methods and uncertainty quantification analysis, using the Yukawa one-component plasma as an exemplar system. Our comprehensive evaluation of seven initialization approaches (uniform random, uniform random with rejection, Halton and Sobol sequences, perfect and perturbed lattices, and a Monte Carlo pair distribution method) demonstrates that initialization significantly impacts equilibration efficiency, with microfield distribution analysis providing diagnostic insights into thermal behaviors. Our results establish that initialization method selection is relatively inconsequential at low coupling strengths, while physics-informed methods demonstrate superior performance at high coupling strengths, reducing equilibration time. We establish direct relationships between temperature stability and uncertainties in transport properties (diffusion coefficient and viscosity), comparing thermostating protocols including ON-OFF versus OFF-ON duty cycles, Berendsen versus Langevin thermostats, and thermostat coupling strengths. Our findings demonstrate that weaker thermostat coupling generally requires fewer equilibration cycles, and OFF-ON thermostating sequences outperform ON-OFF approaches for most initialization methods. The methodology implements temperature forecasting as a quantitative metric for system thermalization, enabling users to determine equilibration adequacy based on specified uncertainty tolerances in desired output properties, thus transforming equilibration from a heuristic process to a rigorously quantifiable procedure with clear termination criteria. I Introduction Molecular dynamics (MD) simulations constitute an important tool for investigating both classical and quantum many-body systems across physics, chemistry, materials science, and biology [ 1 , 2 , 3 ] . Obtaining physically meaningful results from these simulations requires an equilibration stage—a period during which the system reaches a stable, thermodynamically consistent state before data collection commences. This step is essential to ensure that the subsequent production run yields results that are neither biased by the initial configuration nor deviate from the target thermodynamic state [ 4 ] . The efficiency of the equilibration phase is largely determined by the initial configuration of the system in phase space. In classical MD simulations without magnetic fields, including Born-Oppenheimer style density functional theory MD (DFT-MD) simulations, the phase space decouples and the velocity distribution is readily obtainable by sampling from the Maxwell-Boltzmann distribution. However, generating a set of initial positions consistent with the specified thermodynamic state presents a significantly greater challenge, necessitating an equilibration or thermalization phase where the system is driven to the required state, typically through the application of thermostats and/or barostats. Traditionally, initial spatial configurations have been generated using methods appropriate for high-temperature situations such as uniform random placement (with or without a rejection radius), or methods that more closely approximate low-temperature configurations by starting from a known lattice. Such approaches are frequently accompanied by an initial optimization stage in the form of energy minimization. However, these methods can provide poor representations of the true desired state, resulting in characteristic temperature changes [ 5 , 6 ] , and extended equilibration times or persistent biases in the resulting production runs. This has lead to recent work in improving the configuration initialization through running cheaper auxiliary MD simulations [ 7 ] . Despite the fundamental importance of the equilibration phase, the selection of equilibration parameters remains largely heuristic. Some knowledge of best practices exists [ 4 ] , but researchers must often rely on experience, trial and error, or consultation with experts to determine appropriate thermostat strengths, equilibration durations, and thermostat algorithms. This lack of systematic methodology introduces potential inconsistencies across studies and raises questions about the reliability of subsequent production runs. In this work, we tackle two main aspects of improving the equilibration of MD simulations. First, we reduce the arbitrary components of MD equilibration by casting the problem as a uncertainty quantification (UQ). In this framework, instead of guessing an adequate equilibration procedure a priori we use estimates of the target of interest to inform how much equilibration is necessary. To illustrate our approach, we consider the example of the Yukawa one-component plasma for which the outputs of interest are transport coefficients such as self-diffusion and viscosity. We use approximate models for these numbers [ 8 , 9 ] and use their numeric temperature dependence to turn MD temperature uncertainty into an uncertainty in the targeted output. This criterion-driven approach transforms equilibration from an open-ended preparatory step into a quantifiable part of the simulation with clear success/failure feedback. The second main goal is a systematic look at initialization methodology, encompassing both established methods and novel techniques. We compare the amount of undesired heating and equilibration time for methods including uniform random, Monte Carlo-based techniques (uniform random with reject, and pair distribution random), low-discrepancy sequence methods, and lattice-based methods examining both perturbed and unperturbed lattices. Additionally, we compare the effectiveness of different thermostat algorithms, analyze the impact of thermostat coupling strengths, and explore cycling strategies ( e.g. , OFF-ON versus ON-OFF sequences). The remainder of this paper is organized as follows: Section II provides comprehensive descriptions of the seven position initialization algorithms examined in this study along with their theoretical foundations and computational scaling properties. In section III we introduce the Yukawa system as an example to investigate equilibration techniques, detail our simulation methods. We go into thermostating protocols, temperature deviation metrics and introduce an uncertainty quantification framework and statistical measures for establishing equilibration sufficiency based on target property uncertainties. Section IV presents a systematic comparative analysis of initialization-thermalization method combinations across multiple coupling regimes, including microfield distribution characterization, temperature forecasting validation, and quantitative assessment of thermostating duty cycles (ON-OFF versus OFF-ON). Finally, Section V synthesizes our findings into practical guidelines for optimizing MD equilibration. II Position Initialization Methods Molecular dynamics simulations are initialized with the specification of the positions and velocities of all particles. For classical simulations at equilibrium in the absence of a magnetic field, the phase space decouples and the specification of the velocities is simply that of drawing randomly from the Maxwell-Boltzmann distribution. In this section we describe several particle placement algorithms, summarized in Table 1 , with the goal of determining a set of best choices for certain physical situations. Method Description Section Uni Uniform random placement II.1 Uni Rej Uniform random with rejection radius II.2 Halton Low discrepancy, quasi-random sequence generator, Halton method II.3.1 Sobol Low discrepancy, quasi-random sequence generator, Sobol method II.3.2 MCPDF Mesh based Monte Carlo method that matches input pair distribution function II.4 BCC Uni Body-centered cubic (BCC) lattice initialization II.5 BCC Beta BCC lattice with physical perturbations using compact beta function II.6 Table 1: Summary of initialization methods II.1 Uniform Random One of the simplest methods for initializing the particles is to sample each coordinate uniformly from the available position space, \mathbf{r}\sim\mathcal{U}(0,L_{x})\mathcal{U}(0,L_{y})\mathcal{U}(0,L_{z}) . This approach is easy to implement, and the initialization itself is very fast and \mathcal{O}(N) . The drawbacks are a non-zero probability of coincident particle placement, which becomes increasingly likely with larger particle number. This clumping behavior can be quantified mathematically. For any two particles in a cubic simulation box with side length L , the probability that they fall within a distance a of each other is approximately P(d\leq a)\approx\frac{4\pi a^{3}}{3L^{3}}. (1) While this probability may seem small for typical values of a\ll L , the scaling with particle number is problematic. For N particles, we

Read the complete text