<?xml-model href='http://www.tei-c.org/release/xml/tei/custom/schema/relaxng/tei_all.rng' schematypens='http://relaxng.org/ns/structure/1.0'?><TEI xmlns="http://www.tei-c.org/ns/1.0">
	<teiHeader>
		<fileDesc>
			<titleStmt><title level='a'>A Generalized Machine-Learning Framework for Developing Alchemical Many-Body Interaction Models for Polymer-Grafted Nanoparticles</title></titleStmt>
			<publicationStmt>
				<publisher>American Chemical Society</publisher>
				<date>10/14/2025</date>
			</publicationStmt>
			<sourceDesc>
				<bibl> 
					<idno type="par_id">10678929</idno>
					<idno type="doi">10.1021/acs.jctc.5c00901</idno>
					<title level='j'>Journal of Chemical Theory and Computation</title>
<idno>1549-9618</idno>
<biblScope unit="volume">21</biblScope>
<biblScope unit="issue">19</biblScope>					

					<author>Melody Yiyuan Zhang</author><author>Shih-Kuang Alex Lee</author><author>Sharon C Glotzer</author><author>Rebecca K Lindsey</author>
				</bibl>
			</sourceDesc>
		</fileDesc>
		<profileDesc>
			<abstract><ab><![CDATA[Polymer-grafted nanoparticles (PGNs) serve as highly customizable building blocks for technologically relevant self-assembled nanomaterials. Physics informedInverse design strategies are crucial for expediting exploration of the massive associated design space by identifying optimal PGN attributes such aspolymer length and grafting density to meet a targetself-assembled structure. However, their success hingeson accurate description of how PGN interactions varyas a function of both particle positions and physical attributes.We introduce a framework to meet this need.Specifically, we develop an “alchemical” ML-IAM thatdescribes how PGN interactions vary as a function ofboth inter-PGN distances and tunable PGN attributes,simultaneously. This model is an extension of thephysics-informed and explicitly many-bodied ChIMESmodel. The resulting extended ChIMES (X-ChIMES)model is trained on PMF data for PGNs with variedpolymer ligand length. We enable efficient training datageneration by combining the forward-reverse steeredmolecular dynamics enhanced sampling approach witha grid-sampling scheme in the HOOMD-blue softwarepackage. We demonstrate the efficacy of ChIMES forgenerating CG models for PGNs with fixed design attributes,the development of the X-ChIMES model,and the application of the X-ChIMES model that enhancesdigital alchemy inverse-design simulations usingHOOMD-blue and the ChIMES Calculator. This is thefirst application of ChIMES in modeling CG systemsand coupling with an inverse design method to targetnanomaterial self-assembly.]]></ab></abstract>
		</profileDesc>
	</teiHeader>
	<text><body xmlns="http://www.tei-c.org/ns/1.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:xlink="http://www.w3.org/1999/xlink">
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="1">Introduction</head><p>Polymer-grafted nanoparticles (PGNs) provide an endlessly customizable base unit that can be used to build a manifold of complex self-assembled structures with technological applications spanning catalysis, 1 drug delivery, 2 optics, 3 battery technologies, <ref type="bibr">4</ref> and more. These particles are broadly comprised of an inner core nanoparticle (e.g., SiO  2 , Ag, Au, or CdSe) with chains such as alkyl thiol ligands grafted onto the surface, allowing precise modulation of interparticle interactions through mechanisms such as hybridized quantum states, <ref type="bibr">5</ref> covalent interactions, <ref type="bibr">6</ref> and solvent-particle interactions. <ref type="bibr">7</ref> PGNs can self-assemble into a variety of structures, each exhibiting unique and tunable properties pertaining to thermoelectric behavior, <ref type="bibr">8</ref> structural order and symmetry, <ref type="bibr">9</ref> electric permittivity, <ref type="bibr">10</ref> and selective permeability. <ref type="bibr">11</ref> However, this tunability gives rise to the practical challenge of identifying specific PGN configurations (e.g., polymer chain length, grafting density, chemical functionality, and solvent composition) that are suitable for producing a target self-assembled structure (SAS).</p><p>Significant e&#8629;ort has been directed toward addressing this challenge, with early work taking a "forward design" approach, utilizing theoretical frameworks and simulations to predict SAS given an input PGN description. For studies of this natures, accuracy of the underlying models describing inter-PGN interactions is criti-  cal. For instance, initial studies using the hard-sphere model revealed a strong correlation between binary particle size and resulting packing fractions. <ref type="bibr">12</ref> However, phase diagrams produced by these models are not entirely consistent with experiments <ref type="bibr">13,</ref><ref type="bibr">14</ref> due to model's inability to account for properties of soft and compressible polymer ligand layers. More recent studies that employ adjustable short-range pairwise inversepower law potentials have produced more realistic predictions, <ref type="bibr">[15]</ref><ref type="bibr">[16]</ref><ref type="bibr">[17]</ref> and theoretical approaches that implicitly or explicitly account for polymer distribution around the PGN core and many-body e&#8629;ects have enabled predictions with yet better experimental agreement. <ref type="bibr">[18]</ref><ref type="bibr">[19]</ref><ref type="bibr">[20]</ref><ref type="bibr">[21]</ref> While e&#8629;ective, the forward design approach can be ine cient when an SAS is known a priori. In such cases, users must analyze an assortment of PGNs until they identify one that can form the desired SAS. In these cases, an "inverse design" approach becomes attractive, wherein the design space is autonomously and e ciently searched to identify an optimal PGN building block for producing a user-specified SAS. Inverse design can be achieved through a variety of approaches. For instance, a machine-learning (ML)-based landscape engineering method that employs di&#8629;usion maps and hybrid Monte Carlo (MC) simulation has been used to optimize factors such as patch placement and strength on colloidal nanoparticles to ensure the formation of target structures. <ref type="bibr">22</ref> Similarly, genetic algorithm-based reinforcement learning schemes have shown promise in reverseengineering colloidal crystals and quasicrystals. <ref type="bibr">23</ref> Alternatively, the relative entropy method defines a target probability distribution of the desired microstructure and samples interaction parameters that maximize the likelihood of embodying the target system. <ref type="bibr">24,</ref><ref type="bibr">25</ref> Of the multitude of available inverse design schemes, the statistical mechanics-based digital alchemy (DA) method is a particularly attractive approach. <ref type="bibr">26,</ref><ref type="bibr">27</ref> DA can be used to optimize PGN design for a SAS by treating particle attributes such as shape or e&#8629;ective interactions as thermodynamic variables that can be adjusted to minimize the free energy of the target SAS, maximizing liklihood of its formation. <ref type="bibr">27,</ref><ref type="bibr">28</ref> Compared to other inverse design methods, DA is an intuitive, statistical mechanically rigorous, straightforward to implement, and e cient method. DA has been employed in a variety of self-assembly studies, including the inverse design of symmetric hard particles for assembling target colloidal crystals, <ref type="bibr">27</ref> shape-directed solid-solid phase transitions in colloidal systems, <ref type="bibr">29</ref> robust optimization of pair potentials through iterative information maximization, <ref type="bibr">30</ref> optimization of a generalized Fourier potential, <ref type="bibr">31</ref> alchemical molecular dynamics to optimize double well potential, <ref type="bibr">32</ref> and energy landscape analysis for the design of interaction potentials. <ref type="bibr">33</ref> Nevertheless, DA has not yet been applied to systems that exhibit many-body interactions. For instance, PGNs with longer polymer chains can introduce significant steric e&#8629;ects at short range, resulting in threebody or higher-order interactions. Moreover, in its current formulation, DA provides high-level information on optimal PGNs, such as e&#8629;ective interparticle interaction curves or PGN shape, but not detailed insights into specific physical properties of optimal PGNs, such as polymer length or grafting density, which govern e&#8629;ective interactions. Practically, this means experimentally realizing these recommendations requires trial-and-error, i.e., to synthesize PGNs that exhibit the DA-recommended properties. To bridge this gap, DA requires a high-dimensional "alchemical" interaction surface that depends on both system configuration and design parameters directly correlating with the physical properties. Conventional molecular mechanics (MM) models lack the ability to model such an interaction surface that incorporates design parameter descriptions. In contrast, ML-based methods are well suited for learning complex, high-dimensional interaction surfaces, <ref type="bibr">[34]</ref><ref type="bibr">[35]</ref><ref type="bibr">[36]</ref> o&#8629;ering a promising route to enhancing DA.</p><p>To this end, we extended the ChIMES ML interatomic model to establish X-ChIMES, a generalized ML framework for producing coarse-grained (CG) PGN de-sign surfaces describing how e&#8629;ective PGN interactions change as a function of physical properties, which can be integrated into inverse design tools like DA to provide chemically detailed information on optimal PGNs. The remainder of this paper is organized as follows: We begin with a high level overview of our X-ChIMES, followed by description of our training data generation strategy and overview of ChIMES in Sec. 2. We then present results for generating ChIMES inter-particle interaction models in Sec. 3.1 and 3.2. We follow with description of our X-ChIMES extension, its application to MD simulations of PGN systems in Sec. 3.3. We then demonstrate use of the resulting X-ChIMES model in DA calculations in Sec. 3.3. We conclude by discussing implications of the X-ChIMES method and future directions. Key acronyms used throughout the paper are given in Table <ref type="table">1</ref>. </p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2">Methods</head><p>A central goal of this work is to develop PGN design surfaces that describe how inter-PGN interactions vary with both inter-PGN distances and PGN attributes. We aim to describe inter-PGN interactions at the resolution of one-PGN to one interaction site (i.e., OS-CG resolution) due to the significant increase in computational e ciency a&#8629;orded compared to finer-grained descriptions. For fixed PGN attributes, interparticle interaction surfaces can exhibit a variety of shapes. Hence, rather than describing these interactions through conventional molecular mechanics-inspired models, we focus on the development of ML-based solutions. To account for the loss of all conformational degrees of freedom at the OS-CG resolution, our ML interparticle models describe PMF rather than potential energy between particles. <ref type="bibr">37,</ref><ref type="bibr">38</ref> Hence, in the following sections, we provide (1) description of the PGNs studied and our approach to inter-PGN PMF generation, (2) descrip-tion of the ML models developed in this work, and (3) general simulation protocols.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.1">PMF Generation</head><p>As shown in Figure <ref type="figure">1</ref>, we treat polymer-grafted nanoparticles with varied polymer lengths as a testbed system. For e ciency, the ground-truth multi-site coarse-grained (i.e., MS-CG) description is based on the Martini 3 model 39 . Each PGN comprises a 4.7 nm diameter core coated with 300 polymers, resulting in a grafting density of 4.32 chains/nm 2 . Polymer beads interact with the core with a shifted Lennard-Jones (LJ) potential. These polymer beads are modeled as Martini 3 C1 beads, and the bond and angle bending interactions are considered to describe CG beads' connectivity to the core and other CG beads within the polymer chain. Detailed descriptions of coarse-graining and our MS-CG model parameters are provided in Supplementary Information Sections S1, S2, and Table <ref type="table">S1</ref>. PGN PMF calculations are non-trivial due to the many conformation degrees of freedom that must be accounted for. Enhanced sampling techniques an accelerate these calculations. For example, umbrella sampling (US) <ref type="bibr">40</ref> coupled with weighted histogram analysis method (WHAM) <ref type="bibr">41</ref> is a well-established and robust equilibrium-enhanced sampling technique, where PMFs are calculated by analyzing the biased probability distribution sampled around the pre-selected windows distributed along the reaction coordinate. Nourian et al. used US to calculate PMFs for PGNs with various polymer architectures (e.g., linear, dendron, and comb) and found that precisely-tuned comb polymers introduce enhanced interaction range and repulsive strength between PGNs. <ref type="bibr">42</ref> Zhou et al. used the blue moon sampling method to sample and integrate the mean PGN force at various cluster configurations to calculate the accurate pairwise and three-body PMFs for studying the self-assembly of aggregate and dispersed phases. <ref type="bibr">43</ref> While e&#8629;ective, the primary drawback of these equilibrium enhanced sampling methods can be computationally expensive and data-intensive.</p><p>Non-equilibrium methods can also be used for PMF calculation, e.g., Steered Molecular Dynamics (SMD) based on the Jarzynski Equality (JE) formulation. SMD is an enhanced sampling method that applies a moving harmonic potential to steer systems along a chosen reaction coordinate in non-equilibrium molecular simulations. In SMD, the system is pulled across the selected reaction coordinate to calculate the PMFs using the average of the exponentiated work. Park et al. <ref type="bibr">44,</ref><ref type="bibr">45</ref> used the SMD method to calculate the exponentiated work and generate PMFs of deca-alanine oligopeptide molecules.</p><p>However, accurate calculation of PMFs often requires a large number of non-equilibrium SMD simulations to obtain a reliable estimation of the mean exponentiated work done along the reaction coordinate. Consequently, such a technique is rarely applied in the PMF calcula-tion for PGN systems containing large particle degreesof-freedom due to the significant computational burden. As a result, such an approach is limited to simpler systems where sampling e ciency is less challenging. <ref type="bibr">[46]</ref><ref type="bibr">[47]</ref><ref type="bibr">[48]</ref> In the present work, we use the JE formulation to generate PMFs between PGNs via SMD. To overcome the challenge of high computational cost, we employ the forward-reverse (FR) method proposed by Kosztin et al. <ref type="bibr">49</ref> to approximate PMFs. This method employs SMD to generate two time-reversed symmetric paths (i.e., forward and reverse paths) such that the PMFs are correlated to the di&#8629;erence between mean forward work and mean negative reversed work, thereby significantly reducing the number of required SMD runs without directly invoking JE.</p><p>It has been shown <ref type="bibr">[18]</ref><ref type="bibr">[19]</ref><ref type="bibr">[20]</ref><ref type="bibr">[21]</ref> that PGNs with large capping layer thicknesses can exhibit non-negligible many-body e&#8629;ects. Hence, we generate PMF data for both two-and three-body PGN interactions by implementing the FR method into the many-body PMF grid sampling scheme outlined in the work by Zhou et al., <ref type="bibr">43</ref> i.e., leveraging the symmetry of three-PGN clusters to design minimal grids for training data sampling. Additional details and theoretical derivations of these methods are provided in Supplementary Information Section S3.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.2">The ChIMES Machine-Learned Interaction Model</head><p>In the present work, the physics-informed ChIMES 50 machine-learned interaction model (ML-IAM) is used to describe OS-CG resolution inter-PGN interactions and is extended to produce corresponding X-ChIMES PGN design surfaces. Of the multitude of available ML-IAMs, we select ChIMES due to its simplicity, extensibility, inherent capability for describing many-body interactions, and capability to generate high accuracy 34,51-53 models using considerably less training data than more complex ML-IAM. <ref type="bibr">34,</ref><ref type="bibr">54</ref> The ChIMES model, originally developed for treating systems at atomistic resolution, describes system energy through an explicit many-body cluster expansion:</p><p>where E total represents the total energy, consisting of single-particle energy 1 E i1 , pairwise energy 2 E i1i2 , and up to maximum-bodied contributions nB E i1i2..in B . nB represents n-body interactions, i 1 , i 2 , ..., i n represents particle indices, and n a is the total number of particles in the system.</p><p>The single-particle energy is included when ChIMES models are trained on first-principles-derived data, which explicitly account for single-particle contributions to overall system energy and therefore exhibit a compositional dependence even in the case of a fully noninteracting system. In this study, this term is neglected.</p><p>The ChIMES two-body interactions between a pair of particles i and j are given by the following:</p><p>where T A is a Chebyshev polynomial of subscripted order, O 2B is the highest polynomial order, and c A is a set of model fitting coe cients. f s is a "Terso&#8629;" style <ref type="bibr">55</ref> smooth cuto&#8629; function, which ensures the two-body interactions smoothly decay to zero at the outer cuto&#8629;, r c,out . This function incorporates a hyperparameter f o that governs the function's activation point. <ref type="bibr">34</ref> f p is a penalty function designed to prevent the system from sampling distances below the model inner cuto&#8629;, r c,in . s ij is the transformed pair distance, which maps interparticle distances r to the [ 1, 1] domain over Chebyshev polynomials are defined. The transformation from r to s as achieved through the following equation:</p><p>where</p><p>A Morse-style transformation <ref type="bibr">[56]</ref><ref type="bibr">[57]</ref><ref type="bibr">[58]</ref> is commonly used, where x is taken as exp( r/ Morse ). <ref type="bibr">50</ref> Morse-style transformation has the e&#8629;ect of exaggerating changes at short interparticle distances and reducing e&#8629;ective changes in large separation distances. Morse is typically set to the pair distance corresponding to the first peak in the radial pair distribution function (RDF).</p><p>The ChIMES three-body energy is defined as the following:</p><p>For many-body interactions to be valid, all pair distances must fall within the model cuto&#8629;s, forming a fully connected graph between constituent particles. <ref type="bibr">59</ref> As indicated by the asterisk (*) below the triple summation symbol, a su cient number of non-zero polynomial orders must be present within the triple sum to qualify as a true many-body interaction. A three-body interaction is considered valid if all three distances are within the model cuto&#8629;s, and the polynomial orders of at least two of the constituent pairs must begin at one, since the first Chebyshev polynomial is a constant and does not vary with the interparticle distance.</p><p>In this work, ChIMES models are generated by minimizing the loss function between model prediction and ground truth:</p><p>where N train is the total number of training data points, W OS-CG is the OS-CG ChIMES model, and W MS-CG is the ground truth MS-CG PMF training data. Since ChIMES is entirely linear in its fitted parameters, the model optimization problem can be recast as the following overdetermined matrix equation:</p><p>where b MS-CG is the vector that consists W MS-CG for each training configuration, c is the vector consisting model parameters to be determined, and the elements of design matrix A are given by the following:</p><p>where a represents index over PMF components, i.e., sampled dimers or trimers, and b is the index over permutationally invariant model parameters. In this study, model parameters are determined via truncated singular value decomposition (TSVD), where the singular values are set to zero if the corresponding ratios between those and the largest singular value are below &#63743; = 10 5 . Additional information on TSVD is provided in Supplementary Information Section S4.</p><p>In this work, we extend ChIMES to incorporate an additional dimension, &#8629;, which represents a tunable PGN attribute. Henceforth, we refer to this "alchemical" model as X-ChIMES. In this study, &#8629; corresponds to polymer length in terms of the number of MS-CG beads. The X-ChIMES model embeds &#8629; dimension by introducing an additional product of polynomials to the energy expression as shown in Equation <ref type="formula">10</ref>. We note that this strategy can be extended to an arbitrary number of alchemical dimensions through the introduction of additional corresponding polynomials to the product. Consequently, the potential energy surface describing the system is lifted to a higher dimensionality, incorporating information on both inter-PGN distance and polymer length. The X-ChIMES two-and three-body energies are described as follows:</p><p>and</p><p>where s r,ij , O r,2B , and O r,3B are equivalent to s ij O 2B , and O 3B in Equations 2 and 6, and are used to di&#8629;erentiate from s &#8629; , O &#8629;,2B , and O &#8629;,3B , which are introduced to account for the new alchemical dimension. In contrast to the distance dimension which applies a Morse-style transformation to map inter-PGN distances to the [ 1, 1] domain, we employ a direct transformation for the &#8629; dimensions (i.e., x = &#8629; in Equations 3 -5), since we do not know a priori how distance polynomials should be modulated in response to changing &#8629;.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.3">Simulation protocols</head><p>Unless otherwise stated, all simulations were performed in HOOMD-blue <ref type="bibr">60</ref> . The MD simulations were conducted at T = 300 K with a timestep of 2.5 and 5.0 fs for MS-CG and OS-CG systems, respectively. For thermostatting, the Nos&#233;-Hoover thermostat <ref type="bibr">61</ref> was only applied for SMD simulations, since a deterministic integration is required to calculate PMFs via the FR method. For all other MD simulations, the Bussi-Donadio-Parrinello thermostatt <ref type="bibr">62</ref> was used, which is a stochastic thermostat and guarantees the correct sampling of the canonical ensemble. A Langevin piston was employed to control pressure in isobaric-isothermal (NPT) simulations. <ref type="bibr">63</ref> Custom force modules in HOOMD-blue were implemented to perform PMF calculations, including SMD and US simulations. Furthermore, we performed alchemical Monte Carlo (Alch-MC) simulations to inversely design the target SAS. The details about Alch-MC simulations are outlined in Section 3.3.2.</p><p>ChIMES models were generated using the ChIMES LSQ package. <ref type="bibr">64</ref> To enable DA simulations with the X-ChIMES model, a custom HOOMD-blue module was developed using the alchemical Monte Carlo (Alch-MC) algorithm. <ref type="bibr">26,</ref><ref type="bibr">27</ref> Subsequently, the ChIMES Calculator package <ref type="bibr">65</ref> was used to evaluate the ChIMES interaction potential to conduct simulations.</p><p>Trajectory analyses was carried out using freud, 66 and snapshots were rendered using OVITO. 67 signac 68,69 and row <ref type="bibr">70</ref> were implemented for simulation workflow and data management.</p><p>We provide detailed instructions for the packages we developed in Supplementary Information Section S8.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3">Results</head></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.1">PMF Generation</head></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.1.1">FR Method and Benchmarks</head><p>We sampled PMF data for generating ChIMES and X-ChIMES models via SMD simulations with the FR method. Two-body PMFs, W MS-CG 2B , were generated through the following FR SMD protocols: First, two PGNs of a given &#8629; were prepared at a separation distance large enough such that the PGNs were noninteracting. Then, a strong harmonic guiding potential was used to pull the PGNs toward an inner cuto&#8629; distance with a constant pulling velocity. Once the innermost inter-PGN distance was reached, the PGNs were equilibrated at that distance for a su cient time. Subsequently, the PGNs were pulled by the same harmonic guiding potential but in the reverse path (i.e., reversed pulling velocity) to their initial separation distance.</p><p>The FR method relies on the sti&#8629; spring approximation, <ref type="bibr">71</ref> which requires su ciently slow pulling velocities to ensure (1) the positions of the core particles remain in the vicinity of the reaction coordinate, which is the equilibrium distance of the guiding potential, and (2) the inter-PGN forces among the cores and polymer beads remain relatively small compared to the force exerted by the guiding potential, allowing them to be treated as perturbations to the system's potential energy. A challenge faced by directly applying JE is that simulations become considerably more expensive as the pulling velocity is decreased. Other free energy methods such as US circumvent this challenge but require considerably more computational e&#8629;ort due to the need for a large number of individual sampling simulations. Hence, a series of FR SMD simulations were conducted for pulling velocities varying from 1 &#8677; 10 5 to 2.5 &#8677; 10 7 nm/fs (Figure <ref type="figure">S1</ref>) to identify optimal values for our system. For each velocity, we ran five simulations to calculate the average work as required by the FR method. Our selections were validated by comparing the resulting PMFs against those obtained via US with WHAM. A total of 223 windows at spacings of 0.025 nm were used to generate a single PMF. Such a large number of windows is required to ensure adequate histogram overlap and successful PMF generation. For each window, we ran 4 million steps of MD simulation to obtain the biased equilibrium distribution. In Figure <ref type="figure">2</ref>, we show the PMFs of PGNs with 4 and 8 Martini 3 C1 beads per polymer chain (i.e., &#8629; = 4 and 8). We observe strong agreement between two-body PMFs obtained from US and FR SMD simulations with our selected pulling velocities of 1 &#8677; 10 6 nm/fs and 2.5 &#8677; 10 7 nm/fs for &#8629; = 4 and 8, respectively. We can then use the validated pulling velocities to calculate two-body PMFs for PGNs with &#8629; = 3 to 8, each obtained from five SMD simulations. Notably, for any given polymer length, the five combined SMD simulations achieve the same level of convergence while requiring an order of magnitude less simulation time than a single set of 223 US simulations. </p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.1.2">PMF Training Data Sampling</head><p>PGNs with di&#8629;erent &#8629; values have di&#8629;erent e&#8629;ective interaction ranges (e.g., particles with larger values of &#8629; interact over longer ranges than those with smaller &#8629;). Hence, di&#8629;erent initial separation distances were used depending on the &#8629; values of PGNs. We use 10 nm for &#8629; = 3 to 6, 11 nm for &#8629; = 7, and 12 nm for &#8629; = 8. The resulting two-body PMFs are shown in Figure <ref type="figure">3</ref>. As expected, we find an increase in attractive interactions and equilibrium inter-PGN distances in the PMFs as the polymer length increases.</p><p>In the present work, we consider many-body interactions up to three-body interactions, in addition to twobody interactions. The procedure for generating manybody PMFs is illustrated in the top panel of Figure <ref type="figure">4</ref>. We generate the initial configuration by placing three PGNs at separation distances of r 0,12 , r 0,13 , and r 0,23 , where r 0,13 and r 0,23 are large enough to be treated as the zero energy point for three-body PMFs. Following the strategy outlined by Zhou et al., <ref type="bibr">43</ref> non-degenerate sets of initial configurations for the three PGNs are considered such that PGN 1 and 2 are placed symmetrically along the x-axis with respect to the origin. For each fixed r 0,12 value, the position of PGN 3 is shifted by a fixed distance d x along the positive x-axis. This approach assumes that the resulting three-body PMFs inherit cylindrical symmetry, indicating that the energy remains constant when PGN 3 is positioned symmetrically about the y-axis.</p><p>For PGNs with &#8629; = 3 to 6, we selected a total of 319 di&#8629;erent initial configurations with unique combinations of 11 initial r 12 values and 29 d x values ranging between 0 and 9 nm. For PGNs with &#8629; = 7 and 8 that exhibit a longer interaction range, we selected a total of 672 initial configurations with unique combinations of 21 r 12 values and 32 d x values ranging between 0 and 14 nm. These values provide su cient sampling points to produce smooth PMFs. The three-body SMD simulations are similar to the two-body SMD simulations: PGNs 1 and 2 are fixed at their initial positions using strong harmonic potentials, and PGN 3 is pulled along the reaction coordinate, which is the path perpendicular to the x-axis from its initial position, to integrate work done on PGN 3 for PMF calculations.</p><p>As noted by Zhou et al., <ref type="bibr">43</ref> the resulting PMF from these SMD simulations, W MS-CG 3B , includes both the three-body PMF (i.e., W MS-CG</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>3B</head><p>) and two of the twobody PMFs (i.e., W MS-CG 2B (r 13 ) and W MS-CG 2B (r 23 )), for interactions between PGN 1 and 3 and PGN 2 and 3, accordingly:</p><p>We note that W MS-CG 2B (r 12 ) has a value of zero and is excluded, since the pulling direction of SMD is always perpendicular to the axis connecting PGN 1 and 2. Based on Equation <ref type="formula">12</ref>, we can calculate the pure three-body interaction term W MS-CG 3B by subtracting the</p><p>To approximate the many-body PMF for the PGNs, which is dominated by the two-body and three-body PMFs, we can then write:</p><p>To evaluate the significance of three-body interactions, Figure <ref type="figure">4</ref> bottom panel compares 2-dimensional slices of W MS-CG 2B, total and W MS-CG mB for PGNs with &#8629; = 4 and 8 by placing these PGNs on the vertices of isosceles triangles, where r 13 and r 23 are set to be equal, denoted as r, and r 12 are fixed at various distances. After sampling these configurations, we plot two-and many-body PMFs as a function of r. The total two-body PMFs, W MS-CG 2B, total , are obtained by summing the contributions from all three components of two-body interactions, i. , where W MS-CG 2B, total alone underestimates the inter-PGN interactions. We observe more comparable energy profiles between W MS-CG 2B and W MS-CG mB for &#8629; = 4, indicating weaker three-body interactions at shorter polymer length. In contrast, PGNs with &#8629; = 8 exhibit stronger and longer-ranged threebody repulsive interactions. The intensified three-body contribution is consistent with the results reported in literature, <ref type="bibr">[18]</ref><ref type="bibr">[19]</ref><ref type="bibr">[20]</ref> which suggests that PGNs with longer capping layers generally exhibit a stronger many-body e&#8629;ect due to the complex entanglement between polymer ligands.</p><p>A more detailed view of the three-body PMFs is provided in Figure <ref type="figure">5</ref>. Overall, we find that many-body interactions are more attractive and exhibit additional minima in systems with larger &#8629;. However, three-body interactions introduce repulsion at small inter-PGN separation distances that grows with increasing &#8629;. We showcase many-body PMF data across a broader range of r 12 at all &#8629; values in Figure <ref type="figure">S2</ref>-S7.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.2">OS ChIMES Models</head><p>This study represents the first application of ChIMES to CG systems and also presents the unique challenge of generating models describing many-body interactions between nanoparticles.Hence, before exploring the development and application of X-ChIMES models, we first assess the e cacy and accuracy of ChIMES for describing PGNs with fixed &#8629;.</p><p>In the following section, we will use the terms OS-CG and ChIMES interchangeably. We generated a total of 12 individual OS-CG models corresponding to PGNs of &#8629; = 3 to 8, described through either two-body interactions:</p><p>or many-body interactions that are the sum of two-body  </p><p>x q b p z 1 5 0 5 + 6 9 + w / W H s a P H j 9 5 u r 6 x u X V i V K M p j q j i S p 8 V Y J A z i S P L L M e z W i O I g u N p c f E u 5 E / n q A 1 T 8 t g u a s w F T C W r G A X r q Y + 7 8 / O N f j p I 2 0 h u g + w K 9 A 9 / 1 4 t f 3 3 Z / H J 1 v 9 r 5 </p><p>x q b p z 1 5 0 5 + 6 9 + w / W H s a P H j 9 5 u r 6 x u X V i V K M p j q j i S p 8 V Y J A z i S P L L M e z W i O I g u N p c f E u 5 E / n q A 1 T 8 t g u a s w F T C W r G A X r q Y + 7 8 / O N f j p I 2 0 h u g + w K 9 A 9 / 1 4 t f 3 3 Z / H J 1 v 9 r 5 </p><p>x q b p z 1 5 0 5 + 6 9 + w / W H s a P H j 9 5 u r 6 x u X V i V K M p j q j i S p 8 V Y J A z i S P L L M e z W i O I g u N p c f E u 5 E / n q A 1 T 8 t g u a s w F T C W r G A X r q Y + 7 8 / O N f j p I 2 0 h u g + w K 9 A 9 / 1 4 t f 3 3 Z / H J 1 v 9 r 5     (dashed lines) and many-body PMFs W MS-CG mB (solid lines) for PGNs with &#8629;=4 and 8. Each PMF is a function of an isosceles trimer configuration as illustrated in the right figure, where r is the two equal sides (magenta arrows) and r 12 is the fixed base (black arrow). The deviation between W MS-CG 2B and W MS-CG mB indicates the presence of non-negligible three-body interactions. and three-body interactions: W OS-CG mB (r ij , r ik , r jk ) = W OS-CG 2B (r ij ) + W OS-CG 2B (r ik ) +W OS-CG 2B (r jk ) + X i&lt;j&lt;k 3 E ijk . (15) Models were trained on two-body PMFs (i.e., W MS-CG 2B ) and three-body PMFs (i.e., W MS-CG 3B</p><p>) described in Section 3.1. Model hyperparameters are provided in Table 2. Polynomial orders were selected to achieve the best fit with the lowest possible polynomial orders to maximize model e ciency; we note that larger polynomial orders were used for PGNs with greater &#8629; to accommodate their more featured PMF surfaces (see, e.g., Fig. <ref type="figure">5</ref>).</p><p>Figure <ref type="figure">6</ref> provides the performance of two-and manybody models for systems with &#8629; = 4 and 8; we observe excellent model performance across all models with distinct &#8629;. Overall, we find Pearson correlation values of R 2 &#8672; 0.9998 and &#8672; 0.9999 for two-body &#8629; = 4 and 8 models, and &#8672; 0.9994 and &#8672; 0.9939 for many-body &#8629; = 4 and 8 models. A more comprehensive overview of W OS-CG mB including PES heatmaps and error analysis for broader range of r 12 at all &#8629; values is provided in Supplementary Information Figure <ref type="figure">S2-S7</ref>.</p><p>We further validate the performance of the manybody OS-CG models by comparing ChIMES-predicted and MS-CG equilibrium structures for PGN trimers ini-  <ref type="figure">7</ref>, trimer configurations were initialized with PGN-PGN-PGN angle &#10003; 0 of 120 and 90 ; the trimers were then equilibrated in NVT MD simulations. For each &#10003; 0 , we conducted five independent replica simulations. We note that due to the reduction in the degrees of freedom in the OS-CG resolution, simulation time cannot be directly compared between the MS-CG and OS-CG resolutions.  For &#8629; = 8, the top right of Figure <ref type="figure">7</ref> shows that trimers initialized at &#10003; 0 of 90 using both MS-CG and OS-CG models equilibrate to a larger equilateral triangular trimer configuration with (r 12 , r 13 , r 23 ) ' (6.5, 6.5, 6.5) nm within 10 ns, again corresponding to the global energy minima of both W MS-CG mB , and W OS-CG mB for &#8629; = 8 in bottom right of Figure <ref type="figure">5</ref> and bottom middle of Figure <ref type="figure">7</ref>. In contrast, &#8629; = 8 trimers initialized at &#10003; 0 = 120 equilibrate to a configuration with the larger enclosed angle of (r 12 , r 13 , r 23 ) ' (6.0, 6.0, 11.0) nm in simulations using both the MS-CG and OS-CG models, corresponding to the second lowest PMF minimum as illustrated in the bottom middle of Figure <ref type="figure">7</ref>. This local minimum corresponds to a nearly linear trimer configuration illustrated by the green star symbol in the bottom right of Figure <ref type="figure">7</ref>, which explains why the r 23 distance ; bottom: W MS-CG mB ). The red line is the y = x line.</p><p>remained at approximately 11 nm for the trimers initialized at &#10003; 0 = 120 . Such local minima are not observed in the &#8629; = 4 heat map in bottom left of Figure <ref type="figure">7</ref>, which aligns with the observation that for &#8629; = 4 trimers initiated at &#10003; 0 = 120 , both the r 13 and r 23 converged to the energy minimum location (i.e., around 6.0 nm). Ultimately, this study demonstrates (1) fidelity of the PMF calculated with the FR method, (2) e&#8629;ectiveness of ChIMES for learning highly coarse-grained systems with many-body interactions, and (3) acceleration of simulation dynamics a&#8629;orded by the OS-CG model due to the reduction of overall conformational degrees of freedom.</p><p>The performance of the OS-CG models was further evaluated by comparing equation of state (EOS) predictions for PGNs in a BCC and FCC lattice with those generated via the MS-CG models. To generate this data, a series of NPT MD simulations were run at pressures of 10, 100, and 1000 atm for 37.5 and 5 ns for the MS-CG and OS-CG models, respectively. As is shown in Figure <ref type="figure">8</ref>, we find that for &#8629; = 4, many-body OS-CG model predictions are in good agreement with MS-CG predictions at all pressures. In contrast, the two-body OS-CG model underpredicts equilibrium volume, and for the BCC system at 1000 atm, becomes unstable due to an absence of repulsive three-body interactions, allowing particles to sample unphysical distances below the model inner cuto&#8629;.</p><p>For PGNs with &#8629; = 8, two-body OS-CG simulations are only stable at the lowest pressure and severely</p><p>1 3 2 1 3 OS-CG MS-CG OS-CG MS-CG  underpredict the volume of the MS-CG system. The many-body OS-CG model exhibits the expected behavior, i.e., becoming considerably denser with increasing pressure. However, the MS-CG model only exhibits small changes in system volume over the duration of the simulations due to the explicit treatment of longer polymer ligands, which causes considerably slower relaxation dynamics. Consequently, the system remains in the initial metastable state throughout the simulation.</p><p>Figure <ref type="figure">9</ref> provides the corresponding RDF for each NPT simulation. We find excellent agreement between the many-body MS-CG and OS-CG models for PGNs with &#8629; = 4 in both FCC and BCC structures. However, for &#8629; = 8, a mismatch in both the FCC and BCC structures is observed due to the MS-CG systems being kinetically trapped in their initial metastable state.</p><p>The top right of Figure <ref type="figure">9</ref> also highlights the intriguing behavior of PGNs with &#8629; = 8 initialized in the FCC lattice sites, where OS-CG RDF peaks vanish at 10 and 100 atm. Upon visualizing the simulation trajectories, at 10 atm, two out of the five replicas contain a mixture of FCC, BCC, and liquid-like structure at late simulation times, while the rest remain stable in the FCC structure. Additionally, for 100 atm, all five replica simulations show the tendency to transition from FCC to BCC, resulting in a FCC/BCC/liquid mixed state. In contrast, at 1000 atm, all five replicas remain stable in the FCC structure. This behavior occurs be-</p><p>10 atm 100 atm 1000 atm 10 atm 100 atm 1000 atm 10 atm 100 atm 1000 atm 10 atm 100 atm 1000 atm cause at lower pressures, systems can undergo an FCC to BCC transition, while at higher pressures, they are more likely to become kinetically trapped in their initial crystal structures. To verify our observation, we calculated the Steinhardt order parameter 72 q 6 for the NPT MD simulation trajectories of &#8629; = 8 FCC structures, and we present the evolution of q 6 during simulations in Figure <ref type="figure">S10</ref>.</p><p>To qualitatively investigate our observations and to probe the most stable structure self-assembled by these PGNs, we equilibrate an &#8629; = 8 FCC crystal in a box with linearly expanding box length. By continuously increasing the box length without rescaling particle positions, we allow the kinetically trapped system to relax and reconfigure into its energetically favorable structure at lower pressures. We initialized the FCC crystal structure with lattice constant a 0 equal to 9.2 nm and gradually increased the box volume over the first 10 ns of the total 30 ns simulation time. We illustrate the snapshots during the expanding box simulation for &#8629; = 4 and 8 in the Supplementary Information Figure <ref type="figure">S8</ref>. During the box expansion, the FCC structure first melts into a liquid droplet. As the simulation progresses, BCC structure begins to nucleate within the liquid droplet and continues to grow. Eventually, the system converts into a liquid droplet with a BCC structure at the core, which demonstrates the potential of a BCC structure self-assembled from &#8629; = 8 PGNs. For PGNs with &#8629; = 4, we performed the same simulation in a linearly expanding box and observed that the &#8629; = 4 system is stable in a liquid droplet with co-existing FCC and BCC structure at the core. This result, together with the RDF data, shows that FCC and BCC structures have comparable stability for &#8629; = 4. To complement our finding, we provide the approximated phase diagram in the Supplementary Information Figure <ref type="figure">S9</ref> by comparing the enthalpy at T = 0 K between the perfect FCC and BCC crystals at the same statepoint via single-point energy calculation. We note that our observed tendency for phase transformation in PGNs with &#8629; = 8 qualitatively aligns with the OTM model developed by <ref type="bibr">Travesset,</ref><ref type="bibr">20</ref> which explore the relationship between the free energy of PGNs at zero pressure and ligand capping layer thickness , defined as the ratio of the maximal polymer length L to the core radius R, i.e., L/R. Note that this di&#8629;ers from the ChIMES fitting parameter Morse mentioned in Section 2.2. Travesset demonstrates that FCC and BCC structures are the most stable crystal configurations for PGNs, with a transition point from FCC to BCC occurring at &#8673; 0.66 and a flat free energy surface for the remaining values. This result is intuitive in the sense that an increase in corresponds to a thicker polymer grafting layer, and thus, the less dense BCC structures are favored over the more compact FCC structures. The transition from FCC to BCC also signifies a shift from the regime of two-body interaction dominance to where three-body interactions become significant.</p><p>Table <ref type="table">3</ref> provides values for our PGNs, determined based on the partial number integral between the MS-CG core particles and the MS-CG polymer beads of each PGN, as shown in Supplementary Information Figure <ref type="figure">S11</ref>. Beginning with &#8629; = 5, exceeds 0.66, resulting in more significant three-body interactions. This analysis is consistent with Figure <ref type="figure">5</ref>, where three-body interactions are weaker for &#8629; = 4 but prominent for &#8629; = 8. As previously noted, we did not observe a transition from FCC to BCC for PGNs with &#8629; = 4. In light of the value calculations, we attribute this behavior to the comparable stability of the FCC and BCC structures at &#8629; = 4. Specifically, the value of 0.64 is close to the transition threshold of 0.66, which quantitatively explains the FCC/BCC co-existing structure for &#8629; = 4 in the expanding box simulation described above.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.3">OS X-ChIMES Models for Inverse Design</head><p>We now focus on development of an extended OS-CG model (X-OS-CG), i.e., an X-ChIMES model that uni- fies the description of inter-PGN interactions with varying ligand length &#8629;. We note that, in this section, we will use the terms X-OS-CG and X-ChIMES interchangeably. These models are defined as:</p><p>where 2 E &#8629;ij and 3 E &#8629;ijk are two-and three-body X-ChIMES energies defined in Equation <ref type="formula">10</ref>and 11. The X-OS-CG model was generated by separately fitting two-and three-body PMFs, W X-OS-CG 2B and W X-OS-CG 3B , and taking</p><p>. The resulting model treats &#8629; as a continuous variable, e.g., where &#8629; = 2.5 can produce valid energy predictions. We remind the reader that &#8629; corresponds to the number of beads in each MS-CG PGN ligand. Therefore within the context of our &#8629;, these fractional bead counts can account for the fact that each bead is coarse-grained such that multiple repeat units are contained in a single bead. Moreover, the use of a continuous &#8629; dimension makes resulting models smoothly di&#8629;erentiable, allowing for use in general molecular simulation where forces are computed based on potential energy.</p><p>To evaluate the e&#8629;ectiveness of X-OS-CG, a model was trained on PMFs with &#8629; = 3, 4, 6, and 8 while &#8629; = 5 and 7 are held out for validation. Hyperparameter optimization was performed to identify model polynomial orders a&#8629;ording a balance in model accuracy and complexity. For this study, we found optimal polynomial orders to be O s,2B = 8 and O &#8629;,2B = 3 for two-body interactions and O s,3B = 3 and O &#8629;,3B = 2 for three-body interactions.</p><p>When fitting W X-OS-CG</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>2B</head><p>and W X-OS-CG</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head>3B</head><p>, a global inner radial cuto&#8629; of 4.85 nm and an outer radial cuto&#8629; of 12.0 nm was used. A Morse transformation was used on r ij with Morse = 6.0 nm, approximately corresponding to the most attractive inter-PGN distance. A direct transformation was used on &#8629;. Inner and outer cuto&#8629;s for the &#8629; dimension were set to 3 and 8, corresponding to the minimum and maximum number of beads per polymer chain over which models were trained. The complete list of fitting parameters is in Table <ref type="table">2</ref>. Figures 10 and 11 provide validation of the twoand three-body components of the X-OS-CG models in terms of parity plots corresponding to the validation dataset. Overall, we find excellent agreement with Pearson correlation values R 2 of &#8672; 0.9957 and &#8672; 0.9699, respectively. Training parity plots are given in Figure S12. Figures 10 and 11 also provide corresponding PMF heatmaps. We observe expected trends, with W X-OS-CG 2B exhibiting a continuously increasing PMF well depth, well width, and well location near r = 6 nm as &#8629; increases. Figure 11 bottom panel shows W X-OS-CG 3B for di&#8629;erent &#8629; values near the optimal bonding distance, i.e., r 12 = 6 nm, where &#8629; = 4 and 8 correspond to the training dataset and &#8629; = 5 and 7 correspond to the validation dataset. W X-OS-CG 3B  </p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.3.1">Digital Alchemy with X-ChIMES model</head><p>Here, we provide an example application of DA applied to our X-ChIMES model. To this point, DA has been applied to optimize physical attributes of building blocks, such as double well potential, 32 nanoparticle shape anisotropy <ref type="bibr">26,</ref><ref type="bibr">27</ref> and patch patterns <ref type="bibr">28</ref> to enable self-assembly into target SAS. These physical attributes are interpreted as thermodynamic variables known as alchemical parameters &#8629; and are paired with their conjugated quantities &#181; &#8629; , referred to as alchemical potentials. &#8629; is treated as part of the phase space dimensions along with particle position r and momentum p, which jointly fluctuate in the extended statistical ensemble. This application focuses on the Monte Carlo realization of DA, henceforth referred to as Alch-MC. The method works as follows: the extended isothermal ensemble (i.e., the NVT&#181; &#8629; ensemble) is sampled via the Alch-MC method. In this ensemble, the systems possess a biased Hamiltonian that eventually equilibrates toward the target SAS. The biased Hamiltonian is expressed as:</p><p>where K is the kinetic energy, U is the potential energy, and &#8676; (crystal) is the biased potential that is initially applied to constrain the particles onto the lattice sites of the target SAS. This bias is gradually turned o&#8629; during the simulation to allow a smooth transition to an unbiased equilibrium state. This process ensures the final configuration is representative of the true energy landscape of the system that is independent of the initial bias. The &#8676; (crystal) can be expressed as a generic harmonic potential as:</p><p>where the summation runs over all the particles in the system. r represents particle positions, and r (crystal) represents the lattice sites of target SAS. k(t) is the spring constant that linearly decreases to 0 when the simulation time t reaches t end , where the bias potential is completely turned o&#8629;, such that k(0) = k 0 , and k(t end ) = 0. Intuitively, the goal of applying DA via Alch-MC is to optimize &#8629;, which corresponds to a maximized likelihood of producing the specified SAS. To achieve this, we maximize the extended Boltzmann distribution &#8673; &#8629; :</p><p>where Z is the extended partition function that governs the extended statistical ensemble. We choose FCC and BCC crystals as the target SAS where the interactions between PGNs are modeled using the W X-OS-CG mB . To perform the Alch-MC simulation, we first packed PGNs onto FCC and BCC lattice sites using the biased Hamiltonian outlined in Equation <ref type="formula">18</ref>to constrain the PGNs onto lattice sites. We note that the potential energy term U (r, &#8629;) is now W X-OS-CG mB . Each Alch-MC sweep involves two types of MC moves: configurational moves and alchemical moves. Configurational moves attempt to randomly perturb every particle's position, while alchemical moves attempt to randomly adjust the &#8629; value. Both MC moves are based on the Metropolis acceptance criterion. We perform a total of 10000 Alch-MC sweeps. The biased potential was turned o&#8629; at t end = 7500 sweeps to ensure structural stability at the final optimized &#8629;. W X-OS-CG mB consists of attractive two-body interactions dominating at large inter-PGN distances and repulsive three-body interactions being prominent at shorter distances. Due to the interplay between the range of the two-and three-body interactions and the observation that both types of interactions intensify with increasing polymer length, we expect the optimized &#8629; value to change with changing lattice constant a 0 .</p><p>We first initialized FCC and BCC crystals with three di&#8629;erent lattice constants: a 0 = 8.5, 7.6, and 7.1 nm for FCC crystals and a 0 = 6.9, 6.2, and 5.8 nm for BCC crystals. Next, we conducted Alch-MC simulations with the initial &#8629; guesses of 3, 5, and 7, as shown in Figure <ref type="figure">12</ref>. For each pair of a 0 and initial &#8629;, we ran five replica simulations. The top row of Figure <ref type="figure">12</ref> illustrates &#8629; evolution during Alch-MC simulations for FCC crystals with various a 0 . We observe that at the largest a 0 tested, the optimized &#8629; is either 3 or 8, depending on the initial guess of &#8629;. Even though &#8629; = 8 corresponds to the global energy minimum, the system with an initial &#8629; of 3 is not able to overcome the energy barrier to reach &#8629; = 8, illustrating the possibility of multiple local optima. At the smaller a 0 tested, we observe a unique optimum equal to 3, regardless of the initial guess. This result is expected since these densely packed structures with small a 0 prefer smaller &#8629; whose repulsive three-body interactions are weaker. Furthermore, such a transition between optimal &#8629; due to the target structure density features a unique transition from a regime dominated by strong three-body interactions to two-body interactions. Additionally, we observe similar results for BCC systems in the bottom row of Figure <ref type="figure">12</ref>.</p><p>Our observations here are supported by the single-point energy calculation illustrated in Supplementary Figure <ref type="figure">S13</ref>.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="4">Conclusion</head><p>Inverse design algorithms like DA represent powerful tools for optimizing the self-assembly of PGNs for target applications. The predictive power of these tools can be significantly enhanced by building more accu-rate, e cient, and chemically aware design models to guide the exploration. To meet these needs, this work explores the use of the ChIMES ML-IAM 50 to generate highly coarse-grained yet highly accurate models for PGN interactions based on data generated at the MS-CG resolution. We show that these models yield predictions in excellent agreement with the ground truth data, yet are orders of magnitude more computationally e cient. Moreover, we present an e cient sampling strategy for generating PMF training data needed to train these PGN interaction models by integrating the FR method <ref type="bibr">49</ref> into the e cient many-body grid sampling scheme. <ref type="bibr">43</ref> We then demonstrate an extension of the ChIMES CG ML-IAM to describe how these interactions change as a function of PGN attributes -polymer length in this case -and demonstrate how these models can be used within DA simulations for inverse design of target SAS.</p><p>Our simulations of FCC and BCC PGN structures are found to exhibit interesting behaviors. Namely, for &#8629; = 8, our OS-CG model predicted that systems initialized as FCC crystals exhibited a tendency to undergo phase transition to BCC at 10 atm and 100 atm due to the strong repulsive three-body interactions in condensed FCC crystal structures. These results are not observed in the MS-CG simulation due to the limited timescales accessible to this resolution. Our observations are consistent with predictions of Travesset's OTM model, <ref type="bibr">20</ref> where PGNs with longer polymer ligands favor the BCC structure during the self-assembly process. These results demonstrate that the acceleration achieved through our highly coarse-grained OS-CG representation a&#8629;ords access to the longer simulation timescales that can be relevant to system relaxation. Furthermore, our findings on phase transition tendencies highlight the interplay between two-body and threebody interactions and their dependence on the inter-PGN distance.</p><p>Our X-ChIMES DA <ref type="bibr">26,</ref><ref type="bibr">27</ref> application demonstrates that our models preserve physical specificity that is often lost during the CG process. Using PGNs with distinct polymer lengths as a prototypical design parameter, we demonstrate the model's potential for nanoparticle design because the optimized &#8629; using DA corresponds to a specific chemical or physical feature. Our future work will focus on extending this framework by incorporating additional design parameters, such as polymer composition in grafting density, block copolymer, solvent conditions, and particle anisotropy. Ultimately, this work lays a foundation for further advancements in physics-informed CG ML-IAMs to facilitate the design and discovery of novel nanomaterials. als Research Award # DMR 2302470 (machine-learned interaction models, and digital alchemy inverse-design simulations). Computational resources and services were provided in part by the Delta system at the National Center for Supercomputing Applications through allocation DMR 140129 and Stampede3 at the University of Texas at Austin Texas Advanced Computing Center through allocations CHM 240010 and CHM 240082 as part of the Advanced Cyberinfrastructure Coordination Ecosystem: Services &amp; Support (AC-CESS) program, which is supported by National Science Foundation grants #2138259, #2138286, #2138307, #2137603, and #2138296. Additional computing resources were provided by Advanced Research Computing at the University of Michigan, Ann Arbor.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="6">Supporting information and Code Availability</head><p>&#8226; Additional details on methods used in the present work and validation data are available in Supplementary Information.pdf.</p><p>&#8226; Demonstration scripts and parameter files for fitting the ChIMES and X-ChIMES models are available in the the ChIMES-LSQ repository. <ref type="bibr">73,</ref><ref type="bibr">74</ref> &#8226; Example simulations using our ChIMES models in LAMMPS are available in the ChIMES-Calculator repository. <ref type="bibr">75</ref> &#8226; The corresponding training data for these model fits can be accessed through the University of Michigan's Deep Blue Data repository. <ref type="bibr">76</ref> &#8226; Additionally, a separate repository demonstrating integration of ChIMES with HOOMD-blue for conducting simulations and a set of examples of the forward-reverse method for PMF sampling are available at 77.</p><p>Table S1: interaction parameter interaction type parameter 1 parameter 2 parameter 3 U LJ,C1-C1 C1-C1 = 0.47 &#9999; C1-C1 = 0.81 -U SLJ,core-core = 0.47 &#9999; = 0.81 = 2.35 U SLJ,core-C1 = 0.47 &#9999; = 0.81 = 2.35 U bond,core-C1 k bond = 1998 r 0 = 2.35 -U bond,C1-C1 k bond = 1998 r 0 = 0.47 -U &#10003;,C1-C1-C1</p><p>k &#10003; = 2.98 &#10003; 0 = &#8673; -Units: nm, kcal/mol, kcal/mol-nm 2 , radian 3 Background on the Forward-Reverse (FR) Method</p><p>The FR method relies heavily on the sti&#8629; spring approximation, 2 which mitigates the challenges associated with directly applying the Jarzynski Equality (JE).</p><p>To revisit the idea of the sti&#8629; string approximation, we consider a one-dimensional system with a single particle, where the position x serves as the reaction coordinate. The particle resides in a thermal bath at temperature 1 and experiences an external, time-dependent harmonic guiding potential V (x, t) with spring constant k whose equilibrium point moves at a constant velocity v.</p><p>The e&#8629;ect of this harmonic guiding potential can be illustrated by visualizing the particle attached to a constantly moving spring, which is analogous to the harmonic guiding potential used in the SMD simulation.</p><p>Furthermore, we assume that the motion of this particle follows overdamped Langevin dynamics:</p><p>Where is the friction constant, and V(t) represents a delta-correlated white noise with a variance of 2/ that satisfies hV(t)i = 0 and hV(t), V(t 0 )i = 2 (t t 0 ). This white noise arises from random fluctuations due to thermal interactions with the surrounding environment. To describe the system evolution, we are interested in the work w done by the external potential on the particle over a time interval, starting from t = 0 up to an arbitrary time t. We solve for the work w or the power dw/dt as follows:</p><p>Using Equations 8 and 9, we ask the following question: Given that the particle position x and the work done on the particle w are joint random variables evolving in time, what is the corresponding joint probability distribution p(x, w, t) describing this statistical ensemble? To determine this joint probability distribution, we solve the corresponding Fokker-Planck equation, which describes the time evolution of the joint probability for our system governed by the overdamped Langevin dynamics. The corresponding Fokker-Planck equation can be derived as follows:</p><p>We substitute the particle position x with the reference position x = x vt by setting the equilibrium point of the external potential as the origin. Additionally, our system satisfies the initial condition that the particle starts at position x = xi without work being performed on it. Thus, the joint distribution of position and work at initial time step, p(x, w, t = 0|x i ), can be expressed as the product of two delta functions, (x xi ) and (w), which indicates that at t=0, the position x is exactly xi and no work has been performed yet.</p><p>p(x, w, t</p><p>Given this initial condition, we can solve Equation 10 and find that the conditional distribution p(x, w|x i ) is a multivariate Gaussian. <ref type="bibr">2</ref> Furthermore, we can obtain the distribution of position x and work w by marginalizing p, which also follow a Gaussian form:</p><p>Where hxi is the mean reference position at time t, h wi is the mean work at time t, 2</p><p>x is the variance of reference position at time t, and 2 w is the variance of work at time t. Without losing generality, we solve for the conditional distribution p by perturbing the system with small amounts of potential energy relative to the harmonic potential. This perturbation does not alter the fact that p has the form of a multivariate Gaussian. Furthermore, the Gaussian work distribution o&#8629;ers advantageous mathematical properties when deriving the PMFs. This approach is known as the sti&#8629;-spring approximation.</p><p>The FR method 3 combines the conclusions drawn from the sti&#8629;-spring approximation with the Crooks fluctuation theorem to calculate PMFs. The Crooks fluctuation theorem <ref type="bibr">4</ref> states that the ratio of the probability of the work done on the forward path to that on the reverse path is proportional to the total dissipative work done:</p><p>Where w represents the work done on the system, W MS-CG represents PMF, and (w W MS CG ) corresponds the dissipative work. We plug in the forward and reverse work distribution and derive the following equations for the W MS-CG :</p><p>This method accurately calculates the system's PMF using the time-reversal symmetry of non-equilibrium forward and reverse paths. It leverages the existence of a true mean of W MS-CG around which the Gaussian work distributions of the forward and reverse processes are centered.</p><p>Similarly, the conclusion derived from Equation 15 can also be obtained using JE through the cumulant expansion. JE is expressed as:</p><p>The accuracy of calculating W MS-CG crucially depends on the precision of the expected value calculated on the right-hand side of Equation <ref type="formula">16</ref>. Through the cumulant expansion, the right-hand side can be expressed as an infinite series in terms of the probability cumulants of w:</p><p>In general, this infinite series is intractable without knowledge of the form of the work distribution, due to countless possible integration paths to consider, making JE inapplicable in most scenarios. However, if the work distribution can be approximated by a Gaussian, in which only the first two cumulants (mean and variance) are significant, the higher-order terms can be neglected. The corresponding W MS-CG can then be expressed as:</p><p>where we obtain the cumulant approximation of W MS-CG . <ref type="bibr">5</ref> Similarly, along the reverse path, W CG becomes:</p><p>By combining Equations 18 and 19 and assuming that the variances of the forward and reverse work distributions are equal by symmetry, we recover Equation <ref type="formula">15</ref>:</p><p>It is worth noting that this FR method is robust in the sense that (1) it does not require a large number of samples, and (2) a simple linear integration path can be chosen to accurately sample W MS-CG .</p><p>To better illustrate the FR method, we performed the SMD simulations with the FR method with a pulling velocity of 10 6 nm/fs and compared the PMFs against the ones directly calculated from the forward SMD simulations.      13 <ref type="url">https://doi.org/10.26434/chemrxiv-2025-n4d3s</ref> ORCID: <ref type="url">https://orcid.org/0000-0002-3438-9064</ref> Content not peer-reviewed by ChemRxiv. License: CC BY 4.0 Here, we provide the results that support our observation in the main text Section 3.2.</p><p>In figure <ref type="figure">S8</ref>, we illustrate three snapshots during the expanding box simulations using OS-CG models for PGNs with &#8629; = 4 and 8, respectively. The particles are classified and colored as either FCC, BCC, HCP crystal, or others using the polyhedral template matching (PTM) method 6 implemented in Ovito. <ref type="bibr">7</ref> We use the root-mean-square deviation cuto&#8629; of 0.1 as a threshold to compare the particle's local environment with the ideal structural template. To investigate the ideal crystal structure assembled by our PGNs across low to high pressure, we estimate the phase diagram in Figure S9. We approximate the phase diagram by arranging the PGNs at ideal lattice sites of FCC and BCC at di&#8629;erent lattice constant a 0 and varying the &#8629; = 3 to 8. Then, we perform single-point energy calculations on the selected state points. Specifically, we calculate the corresponding potential energy U using the OS-CG model W OS-CG and adding the product of pressure To further verify our observation of the phase transition, we employ the Steinhardt order parameter, q l , with l=6, for probing FCC, BCC, and liquid structures, 11 as shown in Figure <ref type="figure">S10</ref>. For all five replicas at 10 atm, three remain at q 6 &#8673; 0.57, corresponding to stable FCC structures, while the remaining two transition towards q 6 &#8673; 0.38, which indicates existence of liquid structure, since q 6 for liquid is close to 0. This calculation aligns with our previous observation where two out of the five replicas transform from FCC to an FCC/BCC/liquid mixture. At 100 atm, q 6 for all replicas decrease to q 6 &#8673; 0.35, which is consistent with the observation of FCC/BCC/liquid composition upon equilibrium. At 1000 atm, all replicas remain q 6 &#8673; 0.57 throughout the simulation, which is consistent with the observation that all the systems are kinetically trapped in the initial FCC structures at 1000 atm.</p><p>To calculate the PGNs' capping layer thickness , we generated the normalized partial number integral N (r) between core nanoparticles and polymer beads for all of the PGNs studied in this work. is obtained by inspecting where the N (r) firstly reaches one, indicating the outermost distance between core particles and polymer beads. The result is shown in Figure <ref type="figure">S11</ref>.   To better understand the stability of FCC and BCC crystals assembled by PGNs, we conduct single- point energy calculations of crystals by placing the PGNs on perfect lattice sites using the many-body X-ChIMES model. We scan through three di&#8629;erent lattice constants a 0 for FCC and BCC lattices with &#8629; = 3 to 8, and the results are demonstrated in Figure <ref type="figure">S13</ref>. Furthermore, we use these results to approximate the phase diagram by comparing crystal energy at the same first nearest distance between PGNs (i.e., r near ) and the same &#8629;, shown in Figure <ref type="figure">S9</ref>. We note that the equilibrium distances of r near are near 6 nm across all the &#8629; in this work. 8 General Workflow for Coarse-grained (CG) ChIMES Model Generation Here, we describe the general workflow to generate CG ChIMES models using packages and modules we developed.</p><p>contribute to the calculation of PMFs. For simplicity, we showcase the PMF calculation of LJ particles in the repository. Due to the fact that the LJ particle system is extremely simple, one set of SMD simulations is a su cient approximation. Special treatment should be implemented when conducting three-body SMD simulations. In Pseudocode 2, we provide an example usage of our three-body SMD module. Intuitively, the three-body SMD simulation is performed on a system that contains three groups of particles at the fine-grained resolution.</p><p>Two of the groups are fixed, and the other one is pulled along a user-defined reaction coordinate toward point c, denoted as r c . In our paper, as illustrated in Figure <ref type="figure">4</ref>, r c is always on the axis connecting the system represented with group1 index and group2 index. The three-body SMD module is constructed such that the particle group labeled by group3 index is pulled toward r c . r c can be user-defined and by default, is the midpoint between the center-of-mass position of the two groups represented with group1 index and group2 index. During SMD pulling, the group being pulled experiences a static field exerted by the other two fixed groups. We note that the forward and reverse work are functions of the distance between r c and the center-of-mass of the particles represented by group3 index. The resulting work contains two-body e&#8629;ects and must be removed when generating pure three-body PMFs. Additionally, when conducting three-body SMD simulation, the center-of-mass position of the groups represented by group1 index and group2 index should be fixed using a strong spring with spring constant of k fix , while the other spring applied on the group labeled by group3 index have a spring constant of k pull . The three-body SMD module requires users to input two spring constants, k fix and k pull . Where t is the timestep, the three components of the pulling force are force x, force y, force z, respectively, total force is the projection of the pulling force along the pulling direction, spring length is the pulling spring length, d is the actual distance between particle group 1 and 2, and the work is the work done on particle group 2. Below, we showcase the output log file of a three-body SMD simulation. The structure of the log file is similar to that of a two-body SMD simulation, but it contains more information since the system is more complex. For illustration purposes, we split the log file into two sections: .96305 2.97999 2.97039 0.00172 -1.69557 2.96999 2.95310 -0.00148</p><p>Where t is the timestep, r1 deviation and r2 deviation are positional deviations of particle groups 1 and 2 from the desired distance r fix that is specified upon module initialization. The three components of the pulling force are pull force x, pull force y, pull force z, respectively. pull total force is the projection of pulling force along pulling direction, spring length is the pulling spring length, d3c is the actual distance between particle group 3 and r c , and work is the work done on particle group 3.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="8.2">ChIMES Model Fitting</head><p>After we parse the SMD output log file into the xyzf file format that ChIMES uses to formulate the training data, we generate ChIMES models for each PGN configuration using the general ChIMES workflow outlines in the ChIMES LSQ documentation (<ref type="url">https://chimes-lsq.readthedocs.io/en/latest/index.html</ref>).</p><p>Unlike the energy-, force-, and stress-matching approach commonly used in ChIMES model fitting, our models are generated through learning PMF training data. Therefore, we implemented di&#8629;erent bash commands in the job submission script to parse out PMF data during the model training process. After</p></div><note xmlns="http://www.tei-c.org/ns/1.0" place="foot" xml:id="foot_0"><p>https://doi.org/10.26434/chemrxiv-2025-n4d3s ORCID: https://orcid.org/0000-0002-3438-9064 Content not peer-reviewed by ChemRxiv. License: CC BY 4.0</p></note>
		</body>
		</text>
</TEI>
