Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “molecular data”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 199 records · Page 11
All sourcesNASA NTRSDOE OSTI
DOE OSTI2026

Optimizing Cryo-Focused Pyrolysis GC/MS for Tracing Soil Organic Matter Across Diverse Ecosystems

The cycling of organic matter in terrestrial soils and sediments is central to a range of biogeochemical processes that regulate nutrient cycling, crop productivity, trace gas emissions, and contaminant transport. Pyrolysis-gas chromatography/mass spectrometry (py-GC/MS) is a powerful tool for characterizing bulk soil organic matter (SOM) at the molecular level. In this study, we used a cryo-focused py-GC/MS system to analyze soil samples from seven diverse ecosystems: vernal pool, prairie pothole, temperate forest, tropical forest, tundra, wildfire-affected boreal forest, and grassland. We addressed a key bottleneck in molecular-level SOM characterization by developing an automated data analysis pipeline to optimize py-GC/MS and complementary evolved gas analysis/mass spectrometry (EGA/MS) methods, incorporating advanced tools for peak deconvolution, developing a custom compound class library, and implementing fragmentation spectrum-based molecular networking for the first time. This improved workflow was applied to soil samples from all seven ecosystems, including multiple depths and density fractions. Our findings demonstrate that ecosystem type plays a dominant role in shaping compositional differences in SOM. We also identified trends in the source of SOM compounds (e.g., microbial vs plantderived) across soil depth and density fractions, which are critical for understanding persistence and turnover of SOM. Our molecular networking analysis indicated that although many compounds are widespread across ecosystems, others are restricted to specific environments, such as wetlands. This underscores the utility of molecular-level data in elucidating the complexity of SOM composition and the environmental drivers that shape it. Such molecular-level insights can deepen our knowledge of biogeochemical SOM cycles.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2026

Neural network potentials with effective charge separation for non-equilibrium dynamics of ionic solids: a ZnO case study

Developing neural network potentials (NNPs) accurate under non-equilibrium dynamics is challenging, as such systems require extensive sampling beyond equilibrium phases. Here we construct high-fidelity NNPs for zinc oxide (ZnO), a polymorphic ionic solid, using density functional theory (DFT) reference data. To efficiently capture transitional configurations, we combine enhanced-sampling molecular dynamics with empirical potentials, data distillation, and pretraining on short-range atomic energies (A-Train), followed by transfer learning with DFT-relabeled datasets. This hierarchical approach improves transferability across polymorphs and stress states. We further introduce effective charge separation, treating long-range Coulombic terms analytically while short-range residual interactions are learned by the NNP. The optimal effective charges fall in the range 0.5–1.0 q e , consistent with dielectric-screened values derived from formal charges but distinct from Bader estimates. Motivated by this observation, we propose a simple data-driven protocol in which effective charges are optimized by comparing DFT reference energies with explicit Coulomb calculations, without additional NNP training. This strategy improves accuracy and transferability in DFT-level predictions of energies, forces, and stress. Together, these results provide a practical charge-selection framework for robust NNP development in ionic solids, enabling reliable simulation of polymorphic phase transformations and non-equilibrium dynamics.

Chemistry↗
DOE OSTI2024

Viscosity bounds in liquids with different structure and bonding types

Recently, it was realized that liquid viscosity has a lower bound which is nearly constant for all liquids and is governed by fundamental physical constants. This was supported by experimental data in noble and molecular liquids. Here, we perform large-scale molecular dynamics simulations to ascertain this bound in two other important liquid types: the ionic molten salt system LiF and metallic Pb. We find that these ionic and metallic systems similarly have lower viscosity bounds corresponding to the minimum of kinematic viscosity of ∼ 10 − 7 m 2 / s . We show that this agrees with experimental data in other systems with different structures and bonding types, including noble, molecular, metallic, and covalent liquids. This expands the universality of viscosity bounds into the main system types known. Published by the American Physical Society 2024

Materials Science↗
DOE OSTI2022

Post-Marcus Theory and Simulation of Interfacial Charge Transfer Dynamics in Organic Semiconducting Materials (Final Report)

The research program developed, validated and applies predictive computational protocols for calculating charge transfer (CT) rates in complex molecular systems, including molecular dyads and triads in liquid solution and solid-state organic semiconducting (OSC) materials. We have established a transformative computational scheme that goes beyond widely used simplifications, to achieve realistic descriptions of CT processes. The approach properly addresses the contribution of molecular environment at ambient conditions to CT processes. Our approach achieves unique insight on investigated CT processes in relevant experimental efforts. The collaborative team included three principal investigators (PIs), with complimentary expertise in classical molecular dynamics simulations and data science (Cheung), state-of-the-art electronic structure calculations (Dunietz), and cutting-edge theory and simulation techniques for modeling energy, charge and coherence transfer dynamics in molecular systems (Geva).

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2023

Solution Structure and Hydration Forces between Mica and Hydrophilic Versus Hydrophobic Surfaces

Solid-liquid interfaces are central to a range of interesting phenomena including catalysis, heterogeneous nucleation, water desalination, and biomolecular assembly. While three-dimensional Fast Force Mapping (3D FFM) has emerged as a technique for resolving interfacial solution structure at the molecular scale, key challenges for data interpretation persist, most notably regarding the influence of the probe on the measured structure. Using the mica-water system as a case study, we investigate the effect of hydrophilic and hydrophobic probes on interfacial solution structure measured by 3D FFM. Data from hydrophilic silicon-based probes are in good agreement with molecular dynamics simulations, wherein the innermost water molecules adsorb preferentially at the surface ditrigonal cavity sites followed by two subsequent hydration layers. In contrast, the hydrophobic carbon-based probes detect vertical oscillatory features, but do not show lateral patterning that matches the underlying mica lattice. At high ionic strength, up to six of these oscillatory features are observed extending 2 nm into the solution phase with an average spacing of 0.29 ± (0.04) nm. Further, we also determine that the repulsive hydration force between mica and the hydrophilic probe depends on the nature and concentration of ions in solution. Specifically, solutions with stronger ion-water and ion-ion interactions increase the interfacial viscosity, which results in a stronger repulsive hydration force as the probe approaches the surface. Based on these observations, we present a scheme for controlling the outcomes of particle attachment and aggregation by varying the solution conditions to tune the hydration force.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2023

Molecular dynamics simulation of Fe-Si alloys using a neural network machine learning potential

Interatomic potential development using machine learning (ML) approaches has attracted a lot of attention in recent years because these potentials can effectively describe the structural and dynamical properties of complex materials at the atomistic level. Here, in this work, we present the development of a neural network (NN) deep ML interatomic potential for Fe-Si alloys, and we demonstrate the effectiveness of the NN-ML potential in predicting the structures and energies of liquid and crystalline phases of Fe-Si alloys in comparison with the results from ab initio molecular dynamics simulations or experimental data. The developed NN-ML potential is also used to perform molecular dynamics simulations to study the structures of Fe-Si alloys with various compositions under rapid solidification conditions. The short-ranged orders in the rapidly solidified Fe-Si alloys are also analyzed by a cluster alignment method.

36 MATERIALS SCIENCE↗
DOE OSTI2023

Data and scripts associated with a manuscript investigating impacts of solid phase extraction on freshwater organic matter optical signatures and mass spectrometry pairing

This data package is associated with the publication “Investigating the impacts of solid phase extraction on dissolved organic matter optical signatures and the pairing with high-resolution mass spectrometry data in a freshwater system” submitted to “Limnology and Oceanography: Methods.” This data is an extension of the River Corridor and Watershed Biogeochemistry SFA’s Spatial Study 2021 (https://doi.org/10.15485/1898914). Other associated data and field metadata can be found at the link provided. The goal of this manuscript is to assess the impact of solid phase extraction (SPE) on the ability to pair ultra-high resolution mass spectrometry data collected from SPE extracts with optical properties collected on ambient stream samples. Forty-seven samples collected from within the Yakima River Basin, Washington were analyzed dissolved organic carbon (DOC, measured as non-purgeable organic carbon, NPOC), absorbance, and fluorescence. Samples were subsequently concentrated with SPE and reanalyzed for each measurement. The extraction efficiency for the DOC and common optical indices were calculated. In addition, SPE samples were subject to ultra-high resolution mass spectrometry and compared with the ambient and SPE generated optical data. Finally, in addition to this cross-platform inter-comparison, we further performed and intra-comparison among the high-resolution mass spectrometry data to determine the impact of sample preparation on the interpretability of results. Here, the SPE samples were prepared at 40 milligrams per liter (mg/L) based on the known DOC extraction efficiency of the samples (ranging from ~30 to ~75%) compared to the common practice of assuming the DOC extraction efficiency of freshwater samples at 60%. This data package folder consists of one main data folder with one subfolder (Data_Input). The main data folder contains (1) readme; (2) data dictionary (dd); (3) file-level metadata (flmd); (4) final data summary output from processing script; and (5) the processing script. The R-markdown processing script (SPE_Manuscript_Rmarkdown_Data_Package.rmd) contains all code needed to reproduce manuscript statistics and figures (with the exception of that stated below). The Data_Input folder has two subfolders: (1) FTICR and (2) Optics. Additionally, the Data_Input folder contains dissolved organic carbon (DOC, measured as non-purgeable organic carbon, NPOC) data (SPS_NPOC_Summary.csv) and relevant supporting Solid Phase Extraction Volume information (SPS_SPE_Volumes.csv). Methods information for the optical and FTICR data is embedded in the header rows of SPS_EEMs_Methods.csv and SPS_FTICR_Methods.csv, respectively. In addition, the data dictionary (SPS_SPE_dd.csv), file level metadata (SPS_SPE_flmd.csv), and methods codes (SPS_SPE_Methods_codes.csv) are provided. The FTICR subfolder contains all raw FTICR data as well as instructions for processing. In addition, post processed FTICR molecular information (Processed_FTICRMS_Mol.csv) and sample data (Processed_FTICRMS_Data.csv) is provided that can be directly read into R with the associated R-markdown file. The Optics subfolder contains all Absorbance and Fluorescence Spectra. Fluorescence spectra have been blank corrected, inner filter corrected, and undergone scatter removal. In addition, this folder contains Matlab code used to make a portion of Figure 1 within the manuscript, derive various spectral parameters used within the manuscript, and used for parallel factor analysis (PARAFAC) modeling. Spectral indices (SPS_SpectralIndices.csv) and PARAFAC outputs (SPS_PARAFAC_Model_Loadings.csv and SPS_PARAFAC_Sample_Scores.csv) are directly read into the associated R-markdown file. We acknowledge the Yakama Nation as owners and caretakers of the lands where we collected some of these data. We thank the Confederated Tribes and Bands of the Yakama Nation Tribal Council and Yakama Nation Fisheries for working with us to facilitate sample collection and optimization of data usage according to their values and worldview.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2023

MOFX-DB: An Online Database of Computational Adsorption Data for Nanoporous Materials

Machine learning and data mining coupled with molecular modeling have become powerful tools for materials discovery. Metal-organic frameworks (MOFs) are a rich area for this due to their modular construction and numerous applications. Here, we make data from several previous large-scale studies in MOFs and zeolites from our groups (and new data for N 2 and Ar adsorption in MOFs) easily accessible in one place. The database includes over 3 million simulated adsorption data points for H 2 , CH 4 , CO 2 , Xe, Kr, Ar, and N 2 in over 160 000 MOFs and zeolites, textural properties like pore sizes and surface areas, and the structure file for each material. We include metadata about the Monte Carlo simulations to enable reproducibility. The database is searchable by MOF properties, and the data are stored in a standardized JSON format that that is interoperable with the NIST adsorption database. We also identify several MOFs that meet high performance targets for multiple applications, such as high storage capacity for both hydrogen and methane or high CO 2 capacity plus good Xe/Kr selectivity. Here, by providing this data publicly, we hope to facilitate machine learning studies on these materials, leading to new insights on adsorption in MOFs and zeolites.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2020

Non-LTE modelling of cyanoacetylene: evidence for isomer-specific excitation

ABSTRACT Cyanoacetylene molecules are widespread in the interstellar medium (ISM) and several of its isomers have been detected in cold molecular clouds and circumstellar gas. Accurate estimates of the abundance ratio of cyanoacetylene isomers may provide deep insight into their environment. Such knowledge requires rigorous modelling of the emission spectra based on non-local thermodynamic equilibrium (LTE) radiative transfer calculations. To this end, we computed excitation cross-sections of HC2NC and HNC3 induced by collision with para- and ortho-H2, using a quantum mechanical close-coupling method. Then, by thermally averaging these data, we derived rate coefficients for the first 31 low-lying rotational levels of each isomer for temperatures up to 80 K. For the para-H2 collider, the propensity rules are in favour of rotational transitions involving Δj1 = 2 for both isomers, while for the ortho-H2 collider, Δj1 = 2 and Δj1 = 1 rotational transitions are favoured for HC2NC and HNC3, respectively. A comparison of rate coefficients for the HC3N isomers shows differences up to an order of magnitude, especially at low temperatures. Finally, we performed non-LTE radiative transfer calculations to assess the impact of such variations in the analysis of observations. Our simulation suggests that the lack of collisional data specific to each isomer could lead to errors up to a factor of 2–3 in the excitation temperatures. We expect that these data could help in better understanding the cyanoacetylene chemistry and constraining the nitrogen chemistry in the ISM.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2023

Exploring NaCl-PuCl 3 molten salts with machine learning interatomic potentials and graph theory

Actinide molten salts are the basis of the liquid fuels used in molten salt reactors. Due to the inherent difficulties associated with high temperature and hazardous conditions, experimental investigations of fundamental properties of these materials are usually challenging. In this work, we describe the structure and transport of NaCl-PuCl 3 mixtures using computational techniques. Three compositions were considered (16, 25, and 36 mol% PuCl 3 ) over a temperature range (730 – 1257K) using ab initio molecular dynamics, which provided the necessary data sets for training machine learned interatomic potentials. Further, molecular dynamics simulations based on these potentials were then used to determine structure and transport properties. A substantial change was noted in the structure factor when increasing the PuCl 3 content from 25 to 36 mol%. This change is linked to the aggregation of larger Pu 3+ clusters. In addition, the similarity of the atomic environments of metal cations in molten salt systems to their solid states counterparts was investigated using an unsupervised learning technique. Finally, graph theory was employed to explore the structure and size of actinide networks. Consistent with the structure factor, a dense Pu 3+ intermolecular structure is observed within the 36 mol% PuCl 3 mixture. The structure of cation-cation inter-junctions is also discussed. In all cases, the diffusion of Pu 3+ is significantly lower than that of Na + and Cl - .

36 MATERIALS SCIENCE↗
DOE OSTI2021

Adaptive Spatially Aware I/O for Multiresolution Particle Data Layouts

Large-scale simulations on nonuniform particle distributions that evolve over time are widely used in cosmology, molecular dynamics, and engineering. Such data are often saved in an unstructured format that neither preserves spatial locality nor provides metadata for accelerating spatial or attribute subset queries, leading to poor performance of visualization tasks. Furthermore, the parallel I/O strategy used typically writes a file per process or a single shared file, neither of which is portable or scalable across different HPC systems. We present a portable technique for scalable, spatially aware adaptive aggregation that preserves spatial locality in the output. We evaluate our approach on two supercomputers, Stampede2 and Summit, and demonstrate that it outperforms prior approaches at scale, achieving up to 2.5× faster writes and reads for nonuniform distributions. Furthermore, the layout written by our method is directly suitable for visual analytics, supporting low-latency reads and attribute-based filtering with little overhead.

Usher, Will↗
DOE OSTI2022

Intermolecular correlations of liquid and glassy CS 2 studied by synchrotron radiation x-ray diffraction

How is the orientation of molecular liquids ordered on cooling? What are the basic structures of molecular glasses, e.g., close to the crystalline structure or some special structures such as icosahedral cluster? These are long-standing questions in liquid and glass physics. Here, we have constructed a novel cryostat to prepare simple molecular glasses by vapor deposition and performed in situ synchrotron radiation x-ray diffraction experiments. The glassy state of a simple molecule CS 2 , which cannot be vitrified by normal liquid quenching, was successfully prepared with this instrument, and its diffraction data were collected in a wide Q-range of 0.16–25.7 Å -1 with a high-energy diffractometer at BL04B2, SPring-8. The diffraction data of liquid CS 2 were also recorded in a wide temperature range of 160–300 K. These diffraction data were analyzed with molecular dynamics simulations and reverse Monte Carlo modelings to investigate orientational correlation. From the obtained 3D structure models, the orientational correlation between neighboring CS 2 molecules was investigated quantitatively as a function of temperature. At room temperature, the parallel and T-shaped arrangements are preferred for the nearest neighbor correlation. On cooling, these arrangements are developed gradually, and its rate became prominent below the melting temperature (162 K). In the glassy state, the slipped-parallel arrangement is dominant as well as the T-shaped arrangement. Both arrangements appear in the CS 2 crystal, indicating that the structure of glassy CS 2 is close to that of crystalline CS 2 .

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2024

HDBind: encoding of molecular structure with hyperdimensional binary representations

Traditional methods for identifying “hit” molecules from a large collection of potential drug-like candidates rely on biophysical theory to compute approximations to the Gibbs free energy of the binding interaction between the drug and its protein target. These approaches have a significant limitation in that they require exceptional computing capabilities for even relatively small collections of molecules. Increasingly large and complex state-of-the-art deep learning approaches have gained popularity with the promise to improve the productivity of drug design, notorious for its numerous failures. However, as deep learning models increase in their size and complexity, their acceleration at the hardware level becomes more challenging. Hyperdimensional Computing (HDC) has recently gained attention in the computer hardware community due to its algorithmic simplicity relative to deep learning approaches. The HDC learning paradigm, which represents data with high-dimension binary vectors, allows the use of low-precision binary vector arithmetic to create models of the data that can be learned without the need for the gradient-based optimization required in many conventional machine learning and deep learning methods. This algorithmic simplicity allows for acceleration in hardware that has been previously demonstrated in a range of application areas (computer vision, bioinformatics, mass spectrometery, remote sensing, edge devices, etc.). To the best of our knowledge, our work is the first to consider HDC for the task of fast and efficient screening of modern drug-like compound libraries. We also propose the first HDC graph-based encoding methods for molecular data, demonstrating consistent and substantial improvement over previous work. We compare our approaches to alternative approaches on the well-studied MoleculeNet dataset and the recently proposed LIT-PCBA dataset derived from high quality PubChem assays. We demonstrate our methods on multiple target hardware platforms, including Graphics Processing Units (GPUs) and Field Programmable Gate Arrays (FPGAs), showing at least an order of magnitude improvement in energy efficiency versus even our smallest neural network baseline model with a single hidden layer. Our work thus motivates further investigation into molecular representation learning to develop ultra-efficient pre-screening tools. We make our code publicly available at https://github.com/LLNL/hdbind.

59 BASIC BIOLOGICAL SCIENCES↗
DOE OSTI2024

Steric and Electronic Influence of Excited-State Decay in Cu(I) MLCT Chromophores

For the past eleven years, a dedicated effort in our research group focused on fundamentally advancing the photophysical properties of cuprous bis-phenanthroline-based metal-to-ligand charge transfer (MLCT) excited states. We rationalized that by gaining control over the numerous factors limiting the more widespread use of Cu I MLCT photosensitizers, they would be readily adopted in numerous light-activated applications given the earth-abundance of copper and the extensive library of 1,10-phenanthrolines developed over the last century. Significant progress has been achieved by recognizing valuable structure-property concepts developed by other researchers in tandem with detailed ultrafast and conventional time-scale investigations, in-silico-inspired molecular designs to predict spectroscopic properties, and applying novel synthetic methodologies. Ultimately, we achieved a plateau in exerting cooperative steric influence to control Cu I MLCT excited state decay. This led to combining sterics with π-conjugation and/or inductive electronic effects to further exert control over molecular photophysical properties. The lessons gleaned from our studies of homoleptic complexes were recently extended to heteroleptic bis(phenanthrolines) featuring enhanced visible light absorption properties and long-lived room-temperature photoluminescence. Furthermore, this Account navigates the reader through our intellectual journey of decision-making, molecular and experimental design, and data interpretation in parallel with appropriate background information related to the quantitative characterization of molecular photophysics using Cu I MLCT chromophores as prototypical examples.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2021

Hubble WFC3 Spectroscopy of the Habitable-zone Super-Earth LHS 1140 b

Atmospheric characterization of temperate, rocky planets is the holy grail of exoplanet studies. These worlds are at the limits of our capabilities with current instrumentation in transmission spectroscopy and challenge our state-of-the-art statistical techniques. Here we present the transmission spectrum of the temperate super-Earth LHS 1140b using the Hubble Space Telescope (HST). The Wide Field Camera 3 (WFC3) G141 grism data of this habitable-zone (T {sub eq} = 235 K) super-Earth (R = 1.7 R {sub ⊕}) shows tentative evidence of water. However, the signal-to-noise ratio, and thus the significance of the detection, is low and stellar contamination models can cause modulation over the spectral band probed. We attempt to correct for contamination using these models and find that, while many still lead to evidence for water, some could provide reasonable fits to the data without the need for molecular absorption although most of these cause features in the visible ground-based data which are nonphysical. Future observations with the James Webb Space Telescope would be capable of confirming, or refuting, this atmospheric detection.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗
DOE OSTI2020

New H2O–H2O collisional rate coefficients for cometary applications

ABSTRACT We re-introduce a semiclassical methodology based on theories developed for the determination of broadening coefficients. We show that this simple and extremely fast methodology provides results that are in good agreement with results obtained using the more sophisticate MQCT approach. This semiclassical methodology could be an alternative approach which allows to provide large sets of collisional data for very complex molecular systems. It saves time both on the determination of potential energy surfaces and on the collisional dynamical calculations. In addition, this paper provides more complete sets of rotational de-excitation cross-sections and rate coefficients of H2O perturbed by a thermal average of water molecules. Those data can be used in the radiative transfer modelling of cometary atmospheres.

Boursier, C.↗
DOE OSTI2022

Scalable training of graph convolutional neural networks for fast and accurate predictions of HOMO-LUMO gap in molecules

Abstract Graph Convolutional Neural Network (GCNN) is a popular class of deep learning (DL) models in material science to predict material properties from the graph representation of molecular structures. Training an accurate and comprehensive GCNN surrogate for molecular design requires large-scale graph datasets and is usually a time-consuming process. Recent advances in GPUs and distributed computing open a path to reduce the computational cost for GCNN training effectively. However, efficient utilization of high performance computing (HPC) resources for training requires simultaneously optimizing large-scale data management and scalable stochastic batched optimization techniques. In this work, we focus on building GCNN models on HPC systems to predict material properties of millions of molecules. We use HydraGNN, our in-house library for large-scale GCNN training, leveraging distributed data parallelism in PyTorch. We use ADIOS, a high-performance data management framework for efficient storage and reading of large molecular graph data. We perform parallel training on two open-source large-scale graph datasets to build a GCNN predictor for an important quantum property known as the HOMO-LUMO gap. We measure the scalability, accuracy, and convergence of our approach on two DOE supercomputers: the Summit supercomputer at the Oak Ridge Leadership Computing Facility (OLCF) and the Perlmutter system at the National Energy Research Scientific Computing Center (NERSC). We present our experimental results with HydraGNN showing (i) reduction of data loading time up to 4.2 times compared with a conventional method and (ii) linear scaling performance for training up to 1024 GPUs on both Summit and Perlmutter.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2023

Data efficiency and extrapolation trends in neural network interatomic potentials

Abstract Recently, key architectural advances have been proposed for neural network interatomic potentials (NNIPs), such as incorporating message-passing networks, equivariance, or many-body expansion terms. Although modern NNIP models exhibit small differences in test accuracy, this metric is still considered the main target when developing new NNIP architectures. In this work, we show how architectural and optimization choices influence the generalization of NNIPs, revealing trends in molecular dynamics (MD) stability, data efficiency, and loss landscapes. Using the 3BPA dataset, we uncover trends in NNIP errors and robustness to noise, showing these metrics are insufficient to predict MD stability in the high-accuracy regime. With a large-scale study on NequIP, MACE, and their optimizers, we show that our metric of loss entropy predicts out-of-distribution error and data efficiency despite being computed only on the training set. This work provides a deep learning justification for probing extrapolation and can inform the development of next-generation NNIPs.

36 MATERIALS SCIENCE↗