Researchers are continually seeking to improve the efficiency of machine learning interatomic potentials (MLIP) for materials science applications. Emil Annevelink and Varun Shankar of Physics Inverted Materials, Inc., along with Emil Annevelink, demonstrate a new approach that utilizes triplet envelope functions to address the limitations of current MLIP techniques. This work is important because, despite the revolution that MLIP has brought to molecular dynamics, the computational cost of MLIP remains a barrier to widespread adoption. By introducing a higher-order envelope function that intelligently prunes local neighborhoods, this work doubles training and inference speeds, triples memory efficiency, improves simulation stability without compromising accuracy, and provides a path toward more efficient and scalable materials modeling.
Scientists have used machine learning interatomic potentials (MLIP) to revolutionize materials science over the past decade by providing the precision of density functional theory (DFT) with the computational cost of linear scaling in molecular dynamics workflows. However, MLIP still has a relatively high computational cost compared to empirical interatomic potentials, which prevents the conversion of most molecular dynamics simulations by MLIP.
Triplet envelope functions for sparse neighborhood definition and computational efficiency provide a new approach to point cloud processing
A 72-qubit superconducting processor forms the basis of this research and enables the development of higher-order envelope functions for machine learning interatomic potentials (MLIPs). This study addresses that despite offering the accuracy of density functional theory (DFT) with linear scaling, the high computational cost of MLIP currently limits its widespread adoption.
The researchers focused on improving the radial envelope function, which is a key factor determining the size of the local neighborhood and thus the computational cost. In this study, inspired by the geometric functions used in empirical interatomic potentials such as the modified embedded atom method (MEAM), we introduce a triplet envelope function as a method to eliminate local neighborhoods.
These functions reduce interactions between atomic pairs based on angular dependence and effectively dilute the neighbor list without disturbing energy conservation. To evaluate the performance of these new envelope functions, numerical experiments were performed on both solid and liquid materials using radial cutoffs of 5 Å and 8 Å.
Specifically, the methodology involved implementing a triplet envelope function in parallel with a standard radial envelope function and comparing training and inference speed, memory efficiency, and simulation stability. The researchers demonstrated that the triplet envelope function doubles training and inference speeds, triples memory efficiency, and maintains the accuracy of a typical 6 Å cutoff.
Additionally, experiments with a radius cutoff of 8 Å revealed the potential of the triplet envelope function to efficiently model open structures with large interatomic distances, paving the way to larger cutoff radii and improved MLIP performance. This approach is in contrast to K-Nearest Neighbor (KNN) graph sparsification, which reduces cost but compromises energy conservation, a critical element for accurate molecular dynamics simulations.
High-order envelope functions save energy during sparsified molecular dynamics simulations
Scientists are working on challenges within molecular dynamics workflows. The central issue is that machine learning interatomic potentials (MLIPs) use relatively large cutoff radii, converging to 6 Å over the past few years. Large cutoffs prioritize material accuracy over efficiency in certain regions of phase space, capturing dispersion effects and low-density materials at the expense of increased computational cost with high-density materials.
Previous work aimed to address this problem by sparsifying K-Nearest Neighbor (KNN) graphs, which significantly reduced cost but had the drawback of compromising energy conservation. In this work, researchers propose a high-order envelope function that prunes local atomic neighborhoods through physically inspired geometric functions to eliminate non-conservative energy dynamics while providing the memory and efficiency benefits of KNN graph sparsity.
Through numerical experiments on solids and liquids with a 5-8 Å cutoff, they showed that the triplet envelope function complements the radial envelope function by doubling training and inference speed, tripling memory efficiency, and improving simulation stability without affecting the accuracy or data efficiency of the most common 6 Å cutoff. Furthermore, experiments using a radius cutoff of 8 Å show that the triplet envelope function creates a pathway to larger cutoff radii to efficiently and accurately model open structures with large interatomic distances, indicating a promising new direction for engineering MLIP efficiency.
General-purpose MLIP is a new paradigm in computational materials science due to its reduced cost compared to density functional theory (DFT) calculations. Generic models trained on vast DFT datasets now routinely provide quantitative accuracy across the periodic table. Although this is transformative as a replacement for ab initio molecular dynamics, the goal of MLIP is to also replace empirical interatomic potentials.
However, current general-purpose models are still orders of magnitude more expensive than empirical interatomic potentials, and investigating the thermodynamics and kinetics of materials over nanosecond timescales using MLIP remains prohibitively expensive. These barriers have prevented widespread adoption across the computational materials science community and in industrial research settings in particular.
Locality is the main way MLIP reduces the simulation cost for DFT. The local neighborhood is defined by a radial cutoff Rc that limits the receptive field of the MLIP. The number of atoms in the local environment determines the amount of information that MLIP processes and the corresponding cost.
Importantly, locality allows the interatomic potential to scale linearly with system size to access much larger simulations than DFT. However, because the number of atoms in a neighborhood increases cubically with the cutoff radius, increasing the cutoff has a large impact on the cost of the model.
You need to balance cost and accuracy. Reducing the radial cutoff generally comes at the expense of accuracy, since the physical phenomena that the MLIP needs to capture must be within its receptive field. For isolated small molecules, a cutoff radius of 2–3 Å is sufficient to model covalent bonds.
For most metals, a cutoff radius of 4 to 5 Å is usually sufficient. A cutoff of 6 Å or greater is only required when modeling open inorganic structures or dispersion effects. Before general-purpose MLIP became popular, radial cutoff was a hyperparameter that balanced MLIP’s accuracy and cost for the particular physics being modeled.
To be accurate across the periodic table, MLIP employs a uniform cutoff of 6 Å to ensure accuracy across a wide range of materials. Due to the uniform cutoff radius and the variation in structure, different structures have very different numbers of neighbors. For example, a diamond cutoff distance of 5 Å creates a graph in which each carbon atom has five nearest shells in its neighbor list.
Compare this to Valium. For barium, the 5 Å cutoff spans only one nearest neighbor shell. This asymmetry requires a large cutoff of 6 Å for general-purpose models, but also increases the cost of evaluating materials that could otherwise be processed with smaller cutoffs. Previous attempts to deal with large numbers of neighborhoods have been done by sparsifying and pruning each atom’s neighborhood.
The simplest is K-Nearest Neighbors (KNN) sparsification. In this case, edges are sparsified by sorting them according to their length, and only the K nearest neighbors are kept. Each of these previous approaches has shown promising benefits from graph sparsification in terms of memory and cost improvements. However, the main drawback of each approach is that energy conservation is compromised.
Energy conservation is violated because neighbors are removed before the energy contribution goes smoothly to zero from the radial envelope function. Removing energy conservation in MLIP can have a significant impact on thermodynamic observations from molecular dynamics. As molecular dynamics is the primary use case for next-generation, low-cost MLIP, previous sparsification methods are untenable.
Furthermore, the isotropic nature of KNN sparsification can remove important anisotropic features, for example at the solid-liquid interface where the density of the material changes. Despite the tendency to remove the constraints of MLIP, incorporating physical priors has repeatedly proven to be an effective tool to improve the performance of MLIP.
Researchers aim to achieve the same benefits of graph sparsification, cost and memory reduction, without the drawbacks of loss of energy conservation and naive pruning. To achieve this objective, they heavily utilize functionals from empirical interatomic potentials and study the triplet envelope function as a principle method for graph sparsification.
For example, in the modified embedded atom method (MEAM), the local electron density is constructed as the sum of contributions from atoms in the local environment. In particular, the MEAM potential has improved the way local densities are calculated by screening the contributions of atoms within a neighborhood. The screening function distinguishes between atoms in the first and second nearest neighbor shells and decreases monotonically with angular overlap between triplets of atoms.
In this study, researchers investigate the feasibility of the MEAM screening function as a prototype higher-order envelope function. The triplet envelope function is a higher-order three-body function that complements the two-body radial envelope function. The triplet envelope function defines an angle-dependent way to reduce interactions between pairs of atoms.
When the envelope function becomes zero, the list of neighboring atoms is sparsified, reducing the cost of model inference without affecting energy conservation. Through numerical experiments, they showed that the triplet envelope function has a similar effect to KNN sparsification by significantly reducing the number of edges and creating a nearly uniform number of edges for different materials. Additionally, we show how the triplet envelope function improves the radial envelope function to reduce training and inference costs and reduce memory load.
Here we demonstrate the efficiency and stability improvements achieved by implementing higher-order envelope functions.
High-order envelope functions represent a new approach to improve the efficiency of machine learning interatomic potentials (MLIP) in materials science simulations. These functions use geometric principles to refine local neighborhoods and achieve benefits similar to KNN graph sparsification without compromising energy conservation during molecular dynamics workflows.
Numerical experiments conducted in both solid and liquid systems utilizing a cutoff of 5 to 8 angstroms demonstrate that the use of the triplet envelope function in conjunction with the radial envelope function doubles the training and inference speed and triples the memory efficiency. The implementation of the triplet envelope function ensures that the new screening method does not cause instability in the simulations and also improves the stability of the simulations, especially in water-like systems where radial cutoffs alone can be unstable.
This stability improvement is noteworthy since training loss alone is not sufficient to guarantee a stable MLIP. This suggests that learning which features to ignore is an important aspect of potential development. Furthermore, these features may provide a route to effectively utilize larger cutoff radii and enable more accurate modeling of open structures with large interatomic distances.
Future work may focus on exploring the design space of envelope functions and developing more efficient algorithms for computation on both CPUs and GPUs to address the computational overhead associated with triplet interactions. Collectively, these advances suggest promising new directions for engineering the performance of MLIP and extending its applicability to a broader range of materials science problems.
👉 More information
🗞 Triplet envelope function improves efficiency and stability of interatomic potentials in machine learning
🧠ArXiv: https://arxiv.org/abs/2602.02228
