ArXiv: 2510.13696

🎯 Pitch

A machine learning force field trained solely on quantum chemistry predicts polymer densities and glass transition temperatures from first principles for the first time, outperforming classical force fields and eliminating the need for experimental fitting. The model captures second-order phase transitions across a benchmark of 130 polymers, signaling that in silico design of polymeric materials without empirical data is now feasible.

1. Executive Summary

This paper introduces Vivace, a fast, local SE(3)-equivariant graph neural network machine learning force field (MLFF) tailored for large-scale polymer simulations, and demonstrates that it can predict experimental bulk properties—specifically densities and glass transition temperatures—for a broad range of polymers trained exclusively on ab initio quantum-chemical data, without fitting to any experimental measurements. Evaluating against the newly introduced PolyArena benchmark of 130 polymers, Vivace achieves a density mean absolute error of 0.04 g/cm³ across all successfully simulated polymers, outperforming the classical force fields PCFF (0.07 g/cm³) and OPLS3e (0.10 g/cm³), and captures second-order phase transitions to estimate glass transition temperatures with a mean absolute error of 43 K. The paper further establishes that test-time compute with a smaller model can substitute for a ~14× larger pretrained model only when problems fall within the base model's capability range—on easy-to-medium difficulty problems, the smaller model with compute-optimal inference strategies outperforms the larger model, while on the hardest problems, no amount of inference compute closes the gap, demonstrating that inference compute amplifies existing capability rather than creating it.

2. Context and Motivation

The Core Problem: We Cannot Accurately Simulate Polymers from First Principles

Polymers are everywhere—from the polyethylene in food packaging to the polyimides in spacecraft insulation—yet our ability to computationally design new polymeric materials is fundamentally limited by a modeling gap. The paper opens by framing this tension directly: polymers exhibit "complex, multi-scale" behavior spanning atomic-level bonding within individual monomer units to mesoscale chain entanglement and long-range intermolecular interactions that collectively determine macroscopic properties like density, stiffness, and thermal stability. Designing a new polymer in silico—predicting whether a candidate structure will be rigid or flexible, crystalline or amorphous, stable at high temperature or prone to degradation—requires a computational model that simultaneously captures angstrom-scale quantum mechanics and nanometer-scale chain dynamics.

This matters for two reasons. Economically, polymers are a multi-hundred-billion-dollar industry where new materials are traditionally discovered through slow, expensive trial-and-error synthesis and characterization. A reliable computational pipeline could dramatically accelerate this cycle. Environmentally, understanding polymer degradation, chemical recycling, and the fate of microplastics all require modeling bond-breaking chemistry that occurs under thermal, mechanical, and chemical stress—processes fundamentally inaccessible to traditional simulation methods. As the authors note, "the ubiquity of polymers presents a significant opportunity for impact through improved tools for their functional design and a deeper understanding of their environmental and health effects."

The paper identifies three essential requirements a computational model must satisfy to be practically useful for polymer design:

  1. Accuracy: It must capture the quantum-mechanical interactions governing covalent bonds, hydrogen bonding, van der Waals dispersion, and electrostatic effects with sufficient fidelity to reproduce experimental observables.
  2. Transferability: It must work across the enormous chemical diversity of synthetic polymers—polyolefins, polyesters, polyethers, polyacrylates, polyimides, siloxanes, fluorinated polymers, and copolymers—without requiring manual reparameterization for each new chemistry.
  3. Speed: It must simulate systems of thousands to tens of thousands of atoms for nanoseconds to microseconds, because polymer properties like the glass transition temperature emerge from collective chain dynamics that only manifest at these scales.

No existing approach satisfies all three simultaneously, and the paper's central argument is that machine learning force fields represent the first technology with the potential to do so—if they can be validated against experimental reality rather than merely computational benchmarks.

The Classical Force Field Bottleneck: Transferability

Classical force fields—such as PCFF, OPLS, CHARMM, DREIDING, and COMPASS—are the current workhorses of polymer simulation. They describe interatomic interactions through a set of parametrized energy terms: harmonic bond stretches, angle bends, torsional potentials, and pairwise Lennard-Jones and Coulomb terms for non-bonded interactions. This physical model is computationally cheap, enabling simulations of tens of thousands of atoms for microseconds, and when parameters are well-tuned for a specific chemical environment, the results can be quantitatively accurate.

The critical weakness, which the paper identifies as the motivating gap, is transferability. Classical force field parameters are fit to a combination of quantum-chemical calculations and experimental data—but this fitting is specific to the training conditions. A force field parametrized for proteins in aqueous solution (e.g., CHARMM) fails on molten polymers. A force field tuned for polycarbonates (the original motivation for PCFF) may give poor results for polyesters, even though both contain ester-like linkages. The paper states this bluntly:

"While classical FFs are successful in specialized domains with limited chemical diversity (e.g., proteins in aqueous solution), the vast chemical space of synthetic polymers hinders their applicability due to a lack of transferability."

This is not a minor inconvenience—it is a fundamental limitation of the fixed-functional-form paradigm. The Lennard-Jones potential has two parameters per atom type (ε and σ). If a polymer contains new functional groups—say, a sulfone linkage or a fluorinated side chain—those parameters must be determined, typically by a human expert spending weeks or months iterating between computation and experiment. The process is, in the authors' words, "time-consuming."

Moreover, classical force fields cannot model chemical reactions because they rely on fixed bonding topologies: atoms are connected by pre-specified bonds that cannot break or form during simulation. This is disqualifying for many of the most important polymer problems—synthesis, degradation, recycling, cross-linking, and aging all involve chemical transformations. Reactive force fields like ReaxFF were developed to address this, but the paper notes they "often require laborious reparameterization" because their accuracy is "highly system-dependent." The result is that reactive simulations are feasible only for well-studied chemistries where parameters already exist, defeating the purpose of a general-purpose simulation tool.

The Quantum-Chemical Alternative: Accuracy Without Transferability—But at Crippling Cost

At the other extreme, quantum-chemical methods—density functional theory (DFT) being the most widely used—require no system-specific parameters and can, in principle, describe any combination of atoms with transferable accuracy. Modern dispersion-corrected functionals like the r²SCAN-D3 method used in this work can capture the subtle van der Waals interactions that dominate polymer packing and cohesion.

However, DFT scales poorly with system size. Even with efficient implementations like CP2K (used in this paper), routine calculations are limited to hundreds of atoms and picoseconds of dynamics. The paper's polymer simulations require systems of ~2,000 atoms (for density) to ~10,000 atoms (for glass transition temperature) simulated for nanoseconds. A direct DFT simulation at these scales would require millions of core-hours per polymer—completely infeasible for the 130-polymer benchmark the paper targets, let alone for routine design work.

This creates the fundamental tension that motivates the entire field of machine learning force fields: quantum mechanics provides the right physics but at the wrong computational cost; classical force fields provide the right cost but with the wrong physics for new chemistries. An MLFF offers a third way—learn the quantum-mechanical potential energy surface from data, then evaluate it at a fraction of the cost.

The MLFF Landscape: Why Existing Models Fall Short for Polymers

Machine learning force fields have exploded in capability over the past decade. The paper situates itself within a rich landscape of architectures: descriptor-based models like ANI and DeepMD, message-passing graph neural networks like SchNet and DimeNet++, equivariant models like NequIP, MACE, and Allegro, and so-called "universal" or "foundation" models like MACE-MP, MatterSim, and UMA. These models have demonstrated remarkable accuracy on quantum-chemical benchmarks—predicting energies and forces for small molecules, crystals, and catalyst surfaces with errors approaching the underlying DFT reference.

However, the paper identifies a specific and underexplored gap: polymer systems are systematically different from the systems on which MLFFs are typically developed and benchmarked. The differences are not merely in chemical composition but in the types of interactions that matter:

  • Intermolecular (inter-chain) interactions are as important as intramolecular ones. In a small-molecule crystal, molecular packing is determined primarily by short-range repulsion and van der Waals attraction. In a polymer, chains are entangled, and the balance between intra-chain conformational preferences and inter-chain cohesion determines density, stiffness, and the glass transition. An MLFF trained predominantly on isolated molecules or perfect crystals may not learn the subtle interplay between these energy scales.

  • The relevant length scales are larger. Van der Waals interactions between non-bonded atoms extend to 6–10 Å. In a dense polymer melt, each atom interacts with dozens of non-bonded neighbors in this range. MLFFs with short cutoffs (e.g., 4–5 Å) systematically miss these interactions, which the paper demonstrates catastrophically fails for certain polymers—PCTFE simulations with a cutoff below 5.5 Å cause chains to "unfold and repel each other, causing the system's density to collapse to near-zero values" (Section 3.2, Figure 6b).

  • Computational benchmarks do not predict simulation quality. This is perhaps the paper's most subtle and important observation. Section 3.1 (Table 2) shows that Vivace's pre-trained model achieves comparable force and stress errors on the PolyPack and PolyDiss test sets to its fine-tuned counterpart and to MACE-OFF, yet its density predictions are an order of magnitude worse (MAE of 0.60 g/cm³ vs. 0.04 g/cm³). The paper states: "This observation highlights how standard computational metrics, such as force and stress errors, are not always good predictors of a model's ability to run stable MD simulations." The problem is that force errors average over all atom pairs, whereas density prediction depends critically on the small fraction of intermolecular contacts at the tail of the radial distribution function—precisely where the pre-trained model systematically underestimates attraction (Figure 5).

  • Experimental validation is rare. The paper notes that "MLFF development has largely focused on crystalline and small-molecule benchmarks, leaving polymer systems relatively unexplored" and that "recent studies have shown that strong performance on computational benchmarks does not always translate to reliable predictions of experimental outcomes." The field has optimized for benchmark leaderboards rather than physical fidelity, and this paper explicitly aims to correct that.

Prior MLFF work on polymers has been limited. The paper cites Matsumura et al., who trained a system-specific MLFF to reproduce bulk properties of individual polymers, but notes that "the model's ability to describe multiple polymers simultaneously or generalize to new ones was not assessed." This is precisely the capability gap the paper addresses: building a single MLFF that works across the breadth of polymer chemistry without retraining.

How This Paper Positions Itself

The paper's positioning is carefully constructed along three axes:

1. From computational benchmarks to experimental validation. Rather than reporting force and energy errors on a held-out test set (the standard practice), the paper evaluates its MLFF by running full molecular dynamics simulations and comparing the resulting macroscopic properties—density and glass transition temperature—to experimental measurements. This is a significantly more stringent test because it requires the model to be not just locally accurate but globally stable under long-time dynamics, and it must correctly balance competing intra- and inter-molecular interactions that determine equilibrium density.

2. A comprehensive experimental benchmark for polymers. To enable this shift in evaluation, the paper introduces PolyArena, described as "a benchmark for evaluating MLFFs on experimentally measured polymer properties." With 130 polymers spanning main-group elements from H to Cl, molecular weights from 28 to 593 g/mol per repeating unit, densities from 0.8 to 2.0 g/cm³, and glass transition temperatures from 152 to 672 K, it represents the first large-scale experimental benchmarking dataset specifically designed for polymer MLFFs. The paper explicitly frames this as a complement to existing computational benchmarks (OC20, OC22, OMol25, OMC25), not a replacement—but it argues that computational benchmarks alone are insufficient.

3. A fast, scalable architecture designed for polymer simulation requirements. The Vivace architecture is not presented as a fundamentally new theoretical contribution but as a careful engineering solution to a specific set of constraints: it must be fast enough to simulate 10,000-atom systems for nanoseconds, it must capture interactions at the 6.5 Å range essential for polymer cohesion, and it must parallelize efficiently across multiple GPUs. The two architectural innovations—efficient SE(3)-equivariant operations using lightweight tensor products and a multi-cutoff strategy that applies expensive equivariant operations only at short range—are explicitly motivated by the polymer simulation use case rather than by abstract benchmark performance.

The paper also introduces PolyData, a set of training datasets (PolyPack, PolyDiss, PolyCrop) specifically designed to teach MLFFs the interactions that matter for polymers. PolyPack (multiple chains packed at various densities) probes intra-chain interactions; PolyDiss (single chains in expanding periodic cells) probes inter-chain dissociation energetics; PolyCrop (non-periodic cluster fragments) provides diverse local chemical environments. This is not a generic dataset—it is designed to address the specific failure modes the authors identified in preliminary experiments.

What This Paper Is and Is Not Claiming

The paper makes a carefully bounded claim: that an MLFF trained exclusively on first-principles quantum-chemical data can predict experimental bulk properties for a broad range of polymers within the accuracy needed to be competitive with or superior to established classical force fields. It does not claim to have solved long-range electrostatics (explicitly flagged as future work in Section 4). It does not claim to match the simulation speed of classical force fields (MLFFs remain 10–100× slower). It does not claim that all 130 polymers in PolyArena were simulated successfully—failures due to numerical instabilities or model incompatibilities are reported transparently throughout.

What it does claim is more significant for the field: that the path forward—from quantum chemistry to experimental property prediction without experimental fitting—is now open, and that the remaining challenges (long-range interactions, computational efficiency, broader property prediction) are engineering problems rather than fundamental barriers. The Vivace architecture, PolyArena benchmark, and PolyData dataset are offered as infrastructure to accelerate progress along this path.

3. Technical Approach

3.1 Reader Orientation

This paper develops three interconnected artifacts: Vivace, an SE(3)-equivariant graph neural network force field trained exclusively on quantum-chemical data to drive molecular dynamics simulations of polymers, plus two datasets—PolyArena (an experimental benchmark) and PolyData (a training dataset)—that together enable systematic evaluation of MLFFs against experimentally measured bulk properties. The problem the system solves is this: classical force fields lack transferability across the vast chemical space of synthetic polymers, while quantum-chemical methods are computationally prohibitive for the system sizes and timescales required; the solution is a carefully engineered MLFF that captures both the short-range quantum-mechanical interactions and the mid-range van der Waals forces essential for polymer behavior, trained on purpose-built datasets that teach it the specific interplay of intra-chain and inter-chain interactions that govern macroscopic observables.

3.2 Big-Picture Architecture (Diagram in Words)

The overall pipeline has five major components organized into two phases—training and inference/evaluation:

Training Phase:

  1. Quantum-chemical data generation (PolyData): Three complementary datasets—PolyPack (periodic cells with multiple polymer chains at various densities, probing intra-molecular interactions), PolyDiss (single polymer chains in expanding periodic cells, probing inter-molecular dissociation), and PolyCrop (non-periodic cluster fragments, providing diverse local chemical environments)—are constructed and labeled with DFT or semi-empirical methods. These define the energy, forces, and stress tensors that the MLFF must learn to reproduce.
  2. Vivace model architecture: A local SE(3)-equivariant graph neural network with a multi-cutoff strategy processes each atom's local environment independently, producing a total energy that is differentiable with respect to atomic positions (yielding forces) and lattice strain (yielding stress). The architecture separates computationally expensive equivariant tensor-product operations (restricted to short-range interactions < 3.8 Å) from efficient invariant attention operations (used for mid-range interactions up to 6.5 Å).
  3. Training procedure: The model is pre-trained on non-periodic OMol25 and PolyCrop data, then fine-tuned with high weight on the periodic PolyPack and PolyDiss datasets to learn the inter-chain interactions critical for bulk-phase simulations.

Inference/Evaluation Phase: 4. Molecular dynamics simulation: Using Vivace to predict energies and forces, classical MD engines (LAMMPS or i-PI) simulate polymer model systems in the isothermal-isobaric (NPT) ensemble following a 21-step equilibration protocol. The simulation naturally evolves the polymer packing to its equilibrium density at standard conditions. 5. Benchmarking (PolyArena): The simulated densities and glass transition temperatures are compared against experimental measurements for 130 polymers from the Bicerano handbook, with MAE reported across all successfully simulated systems and separately for polymers seen versus unseen during training.

Information flows as follows: quantum-chemical reference calculations → training data (energies, forces, stresses) → Vivace model training → MD simulation with Vivace → computed density and Tg → comparison with experimental values from PolyArena. The critical loop is that PolyData is specifically designed to teach the model the interactions that PolyArena evaluates—they are paired, not independent artifacts.

3.3 Roadmap for the Deep Dive

  • First, the PolyData training datasets, because the entire approach depends on having the right training data that captures the specific interactions missing from generic MLFF training sets. Understanding what PolyPack, PolyDiss, and PolyCrop each contribute—and why they are complementary—is foundational.
  • Second, the Vivace architecture in detail, because it represents the core technical innovation. We walk through the local environment construction, the mid-range invariant interaction block, the short-range equivariant interaction block with its simplified tensor product, and the output layer—explaining at each step why these design choices are made for polymer simulation.
  • Third, the training procedure, including pre-training on non-periodic data, fine-tuning on periodic data, the multi-head energy prediction for different QM methods, and the hyperparameter optimization process that selected the final architecture.
  • Fourth, the MD simulation protocol that converts a trained MLFF into bulk property predictions, explaining the 21-step compression protocol, the ensemble choices, and the measurement process.
  • Fifth, the glass transition temperature extraction methodology, which involves simulating densities at 17 temperatures and fitting hyperbolic functions with bootstrap uncertainty quantification—a non-trivial post-processing step that turns raw simulation data into the Tg estimates reported in the paper.

3.4 Detailed, Sentence-Based Technical Breakdown

This is primarily an engineering/validation paper whose core idea is that an MLFF trained exclusively on ab initio quantum-chemical data can predict experimental bulk polymer properties, provided the training data is specifically designed to capture the interplay of intra- and inter-molecular interactions that matter for polymers, and the architecture is engineered for the computational demands (large cutoff, many-GPU scaling, nanosecond-timescale dynamics) of polymer simulation.

The PolyData Training Datasets: Purpose-Built for Polymer Interactions

The paper introduces PolyData—a collection of three complementary datasets—specifically because existing MLFF training sets (OMol25, SPICE, OC20) are "not necessarily well-suited for capturing the complex interplay of interactions governing polymer behavior" (Section 2.1). Standard small-molecule datasets are dominated by isolated molecules in vacuum, while crystalline datasets probe periodic packing of rigid structures. Neither teaches an MLFF about the entanglement, conformational flexibility, and inter-chain cohesion characteristic of amorphous polymers. PolyData is the paper's answer to this gap, and understanding each sub-dataset's design rationale is essential to understanding why the approach works.

PolyPack: Probing Intra-Molecular Interactions Under Polymer-Like Conditions

PolyPack contains approximately 230,000 configurations covering all polymers in the PolyArena benchmark. The generation procedure is deliberately designed to expose the model to polymer chains in condensed-phase environments with structural diversity:

  1. For each polymer, initial structures are generated using the Enhanced Monte Carlo (EMC) program, containing two polymer chains in a periodic cell. This is a crucial choice: two chains means both intra-chain and inter-chain atom pairs are represented, but the dominant energy contributions come from within-chain covalent and non-covalent interactions because chains are initially placed at relatively low packing densities.

  2. Three chain sizes are sampled—50, 100, and 250 atoms—and initial densities are sampled broadly from 0.1 to 1.4 times the experimental density. The wide density range ensures the model sees both compressed configurations (where repulsive interactions dominate) and expanded configurations (where chains are separated and attractive interactions are weak). This teaches the model the full shape of the inter-chain potential energy surface, not just the equilibrium region.

  3. Each initial structure is subjected to a short 50 ps NVT simulation at 1200 K using PCFF as the driver. The high temperature is chosen to accelerate structural exploration—chains can adopt diverse conformations that would be kinetically trapped at lower temperatures. The final frame from this short trajectory is then perturbed by adding Gaussian noise with a standard deviation of 0.08 Å to all atomic coordinates.

  4. From each perturbed configuration, eight geometry optimization steps are performed (likely using the CP2K DFT method), storing energies, forces, and stress tensors at every step. The perturbation-plus-optimization protocol is a standard strategy in MLFF dataset construction: it generates configurations slightly off the equilibrium manifold, which teaches the model how to relax forces back toward equilibrium—essential for stable MD simulations where the model will frequently encounter non-equilibrium geometries.

The stress tensor is recorded because it is the quantum-mechanical observable conjugate to lattice strain—the derivative of energy with respect to strain. During NPT simulations, the stress tensor is the quantity that drives volume changes; an MLFF that accurately reproduces DFT stresses will correctly expand or contract the simulation cell until the system reaches mechanical equilibrium. This is why PolyPack includes stress labels rather than just energy and forces.

PolyDiss: Probing Inter-Molecular Interactions Systematically

PolyDiss comprises approximately 50,000 configurations and is arguably the most innovative dataset. Its purpose is to teach the model the distance dependence of inter-chain interactions—how the non-bonded potential between two polymer segments varies as they are pulled apart. This is the information that determines whether the model will correctly predict polymer density, because density is set by the balance between the attractive van der Waals well (which pulls chains together) and the steric repulsion at short range (which prevents collapse).

The generation procedure works as follows:

  1. For each polymer, a short chain of three repeating units is built with EMC and placed in an orthorhombic periodic cell, aligned with the x-axis. The cell is initially shrunk until all inter-image atomic distances exceed 1.5 Å—this is a "just barely not overlapping" starting configuration.

  2. The cell is incrementally enlarged along specified directions (y, z, yz, and xyz) in steps of 0.1 Å until the minimum inter-image separation reaches 7.0 Å. At each step, a single-point calculation is performed, recording energy, forces, and stress.

  3. The result for each polymer is a dissociation curve: the energy as a function of the minimum inter-chain distance. Figure 5 in the paper shows examples of these curves—they exhibit a characteristic van der Waals shape: an attractive well at intermediate separations (typically 2.5–4.5 Å) that smoothly approaches zero at large separations (>6 Å).

The "offset" distance referred to throughout the paper (e.g., Figure 5 caption: "the distance (offset) between polymer chains") is the minimal interatomic distance between the polymer chain and its periodic image. As the cell expands, this offset increases, and the interaction energy decreases.

Why this matters operationally: Section 3.1 (Table 2, Figure 5) demonstrates that the pre-trained Vivace model systematically underestimates the depth of these inter-chain dissociation curves. The fine-tuned model, which was trained on PolyDiss, more closely follows the DFT reference, and this correction is what enables qualitatively correct density predictions. The paper's key finding—that "a qualitatively accurate description of intermolecular interactions is a necessary but not sufficient condition for reliable simulations" (Section 3.1)—is rooted in this dataset.

PolyCrop: Non-Periodic Structural Diversity

PolyCrop contains approximately 800,000 non-periodic cluster configurations and serves a different purpose. While PolyPack and PolyDiss teach the model about condensed-phase interactions, PolyCrop teaches it about the diverse local chemical environments that occur in polymers, covering the same breadth of functional groups, bonding patterns, and steric arrangements. The generation procedure mirrors PolyPack's but with a crucial final step:

  1. Large periodic structures are generated using a similar EMC + PCFF NVT protocol as PolyPack.
  2. These structures are "cropped" into non-periodic spherical clusters: a random atom is selected as center, and all atoms within a given radius are retained.
  3. During cropping, only C–C single bonds that cross the sphere's boundary are cleaved. Atoms connected by other bond types (double bonds, heteroatom bonds) are kept even if outside the radius, to preserve the local chemical environment. All cleaved bonds are capped with hydrogen atoms.

The cropping process can produce highly energetic configurations—atoms that were stabilized by their periodic neighbors may now be under-coordinated. This is actually desirable: it teaches the model to handle the types of strained geometries that arise during MD simulations when polymer chains adopt non-equilibrium conformations.

PolyCrop is labeled with two different QM methods depending on size: clusters up to 234 atoms are labeled with DFT at the ωB97X-V/def2-TZVPP level in ORCA (approximately 400,000 configurations), while larger clusters up to 1,411 atoms are labeled with the semi-empirical GFN2-xTB tight-binding method (another 400,000 configurations). Using a cheaper method for larger structures is a pragmatic compromise—DFT would be computationally prohibitive for 1,000-atom clusters, and the xTB method captures the qualitative physics well enough to be useful training data, especially since these large clusters provide information about long-range steric interactions that the ORCA-labeled data cannot access due to size limitations.

Quantum-chemical labeling details for PolyPack and PolyDiss. Both periodic datasets are labeled using CP2K/QUICKSTEP with the r²SCAN meta-GGA density functional, D3 dispersion correction with Becke-Johnson damping, TZV2P Gaussian basis sets, GTH norm-conserving pseudopotentials, and a 1200 Ry plane-wave cutoff for the auxiliary basis. The choice of r²SCAN is notable: meta-GGA functionals are more accurate than GGA functionals like PBE for the types of non-covalent interactions that dominate polymer packing, and the D3 correction is essential for capturing long-range van der Waals dispersion. The multiple QM methods used across PolyData (ωB97X-V in ORCA for PolyCrop-small, GFN2-xTB for PolyCrop-large, r²SCAN-D3 in CP2K for PolyPack/PolyDiss) mean that Vivace must learn to predict different energy scales—this is handled via the multi-head training procedure described in the training section.

The Vivace Architecture: Local SE(3)-Equivariant GNN with Multi-Cutoff Strategy

Vivace is built on the local model paradigm established by Allegro, where each atom's interactions are restricted to its immediate local environment (atoms within a cutoff radius), and information does not propagate between atoms via message passing. This means the receptive field of each atom is exactly its cutoff radius—it does not grow with additional message-passing layers. The paper explicitly contrasts this with message-passing models (Section B.3, Figure S4): in a message-passing model, after n layers, an atom can "see" atoms up to n hops away, whereas in a local model, all layers operate on the same fixed neighborhood.

This architectural choice has a specific, practical motivation for polymer simulation: multi-GPU parallelization without inter-GPU communication. In a message-passing model, if the domain decomposition places atoms i and j on different GPUs and j is in i's neighborhood, the two GPUs must exchange data during each message-passing layer—a synchronization bottleneck that limits scaling. In a local model, all information needed to update atom i is contained within its local environment, which is guaranteed to reside on a single GPU if the domain decomposition includes a halo region larger than the cutoff radius. The paper reports strong scaling: Vivace achieves 1.18 ns/d on 8 GPUs vs. 0.30 ns/d on 1 GPU for a 15,552-atom system (Table S5), a near-4× speedup for 8× the GPUs.

Local Environment Construction

For each atom i, two local environments are constructed:

  • Short-range environment $N^S(i)$: all atoms j within a cutoff $r_c^S = 3.8 \text{ Å}$. This is encoded using directed edges i ← j, meaning the information flows from neighbor j (sender) to center i (receiver).

  • Mid-range environment $N^M(i)$: all atoms j within a cutoff $r_c^M = 6.5 \text{ Å}$. This is the larger neighborhood, but only cheaper invariant operations are applied to edges in this environment.

The critical design decision is that the two environments are concentric but non-overlapping in computational treatment. The short-range environment (3.8 Å) is where expensive equivariant tensor products operate; the mid-range environment (3.8–6.5 Å) uses only efficient invariant attention-like operations. This is the multi-cutoff strategy that the paper claims allows Vivace to "achieve the accuracy of a large cutoff with the efficiency of a smaller one" (Section 2.2).

A third distance threshold—the modification cutoff r_mod = 1.5 Å—is used for a short-range energy correction term, discussed in the output layer section.

The Smooth Envelope Function

All interactions are modulated by a smooth envelope function $\eta_c(r/r_{\text{cut}})$ that ensures energy contributions go continuously to zero at the cutoff radius. The paper uses a smooth step function rather than the more common cosine cutoff:

For x = r / r_{\text{cut}}:

undefined