arXiv subjects

Kin Long Kelvin Lee

Publications and source records attributed to Kin Long Kelvin Lee.

At least 19 recordsLinked to original sources

arXiv2026

UniFFBench: Evaluating Universal Machine Learning Force Fields Against Experimental Measurements

Universal machine learning force fields (UMLFFs) promise to revolutionize materials science by enabling rapid atomistic simulations across the periodic table. However, their evaluation has been limited to computational benchmarks that may not reflect real-world performance. We introduce UniFFBench, a comprehensive evaluation framework featuring the MinX dataset -- a diverse collection of 1,500+ mineral systems spanning 85 elements, extreme thermodynamic conditions (0--5000 K, 0--1000 GPa), and structural complexity, including partial occupancy and disorder. This diversity, combined with experimental reference values for validation, enables assessment of UMLFF generalization across chemical space and conditions substantially beyond typical training scenarios. Our systematic evaluation of six state-of-the-art UMLFFs reveals a substantial ``reality gap'': models achieving impressive performance on computational benchmarks often fail when confronted with experimental complexity. Even the best-performing models exhibit higher density prediction error than the threshold required for practical applications. We observe disconnects between simulation stability and mechanical property accuracy, with prediction errors correlating with training data representation rather than the modeling method.

cond-mat.mtrl-sci↗

arXiv2026

OXtal: An All-Atom Diffusion Model for Organic Crystal Structure Prediction

Accurately predicting experimentally realizable 3D molecular crystal structures from their 2D chemical graphs is a long-standing open challenge in computational chemistry called crystal structure prediction (CSP). Efficiently solving this problem has implications ranging from pharmaceuticals to organic semiconductors, as crystal packing directly governs the physical and chemical properties of organic solids. In this paper, we introduce OXtal, a large-scale 100M parameter all-atom diffusion model that directly learns the conditional joint distribution over intramolecular conformations and periodic packing. To efficiently scale OXtal, we abandon explicit equivariant architectures imposing inductive bias arising from crystal symmetries in favor of data augmentation strategies. We further propose a novel crystallization-inspired lattice-free training scheme, Stoichiometric Stochastic Shell Sampling ($S^4$), that efficiently captures long-range interactions while sidestepping explicit lattice parametrization -- thus enabling more scalable architectural choices at all-atom resolution. By leveraging a large dataset of 600K experimentally validated crystal structures (including rigid and flexible molecules, co-crystals, and solvates), OXtal achieves orders-of-magnitude improvements over prior ab initio machine learning CSP methods, while remaining orders of magnitude cheaper than traditional quantum-chemical approaches. Specifically, OXtal recovers experimental structures with conformer $\text{RMSD}_1<0.5$ Å and attains over 80\% packing similarity rate, demonstrating its ability to model both thermodynamic and kinetic regularities of molecular crystallization.

cs.LG↗

arXiv2025

Benchmarking Dimensionality Reduction Methods for High-Dimensional ALMA Image Cubes

High-dimensional astronomical data cubes provide a wealth of spectral and structural information that can be used to study astrophysical and chemical processes. The complexity and sheer size of these datasets pose significant challenges in their efficient analysis, visualization, and interpretation. In specific astronomical use cases, a number of dimensionality reduction techniques, including traditional linear (e.g. principal component analysis) and modern nonlinear methods (e.g. convolutional autoencoders) have been used to tackle this high-dimensional problem. In this study, we assess the strengths, weaknesses, and nuances of various methods in their ability to capture and preserve astronomically-relevant features at lower dimensions. We provide recommendations to guide users in identifying and incorporating these treatments to their data, and provide insights into the computational scalability of these methods for observatory level data processing. This benchmark study uses publicly available archival ALMA data from a diverse sampling of source morphologies and observing setups to assess the performance and trade-offs between computational cost, image reconstruction accuracy, and scalability. Finally, we discuss the generalizability of these techniques in regard to data segmentation and labeling algorithms and how they can be exploited for advanced data product generation and streamlined archival analysis as we prepare to enter the era of the ALMA Wideband Sensitivity Upgrade.

astro-ph.IM↗

arXiv2025

SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models

Generating novel crystalline materials has the potential to lead to advancements in fields such as electronics, energy storage, and catalysis. The defining characteristic of crystals is their symmetry, which plays a central role in determining their physical properties. However, existing crystal generation methods either fail to generate materials that display the symmetries of real-world crystals, or simply replicate the symmetry information from examples in a database. To address this limitation, we propose SymmCD, a novel diffusion-based generative model that explicitly incorporates crystallographic symmetry into the generative process. We decompose crystals into two components and learn their joint distribution through diffusion: 1) the asymmetric unit, the smallest subset of the crystal which can generate the whole crystal through symmetry transformations, and; 2) the symmetry transformations needed to be applied to each atom in the asymmetric unit. We also use a novel and interpretable representation for these transformations, enabling generalization across different crystallographic symmetry groups. We showcase the competitive performance of SymmCD on a subset of the Materials Project, obtaining diverse and valid crystals with realistic symmetries and predicted properties.

cond-mat.mtrl-sci↗

arXiv2025

Stiefel Flow Matching for Moment-Constrained Structure Elucidation

Molecular structure elucidation is a fundamental step in understanding chemical phenomena, with applications in identifying molecules in natural products, lab syntheses, forensic samples, and the interstellar medium. We consider the task of predicting a molecule's all-atom 3D structure given only its molecular formula and moments of inertia, motivated by the ability of rotational spectroscopy to measure these moments. While existing generative models can conditionally sample 3D structures with approximately correct moments, this soft conditioning fails to leverage the many digits of precision afforded by experimental rotational spectroscopy. To address this, we first show that the space of $n$-atom point clouds with a fixed set of moments of inertia is embedded in the Stiefel manifold $\mathrm{St}(n, 4)$. We then propose Stiefel Flow Matching as a generative model for elucidating 3D structure under exact moment constraints. Additionally, we learn simpler and shorter flows by finding approximate solutions for equivariant optimal transport on the Stiefel manifold. Empirically, enforcing exact moment constraints allows Stiefel Flow Matching to achieve higher success rates and faster sampling than Euclidean diffusion models, even on high-dimensional manifolds corresponding to large molecules in the GEOM dataset.

cs.LG↗

arXiv2025

Energy & Force Regression on DFT Trajectories is Not Enough for Universal Machine Learning Interatomic Potentials

Universal Machine Learning Interactomic Potentials (MLIPs) enable accelerated simulations for materials discovery. However, current research efforts fail to impactfully utilize MLIPs due to: 1. Overreliance on Density Functional Theory (DFT) for MLIP training data creation; 2. MLIPs' inability to reliably and accurately perform large-scale molecular dynamics (MD) simulations for diverse materials; 3. Limited understanding of MLIPs' underlying capabilities. To address these shortcomings, we aargue that MLIP research efforts should prioritize: 1. Employing more accurate simulation methods for large-scale MLIP training data creation (e.g. Coupled Cluster Theory) that cover a wide range of materials design spaces; 2. Creating MLIP metrology tools that leverage large-scale benchmarking, visualization, and interpretability analyses to provide a deeper understanding of MLIPs' inner workings; 3. Developing computationally efficient MLIPs to execute MD simulations that accurately model a broad set of materials properties. Together, these interdisciplinary research directions can help further the real-world application of MLIPs to accurately model complex materials at device scale.

cond-mat.mtrl-sci↗

arXiv2024

Accelerating Quantum Emitter Characterization with Latent Neural Ordinary Differential Equations

Deep neural network models can be used to learn complex dynamics from data and reconstruct sparse or noisy signals, thereby accelerating and augmenting experimental measurements. Evaluating the quantum optical properties of solid-state single-photon emitters is a time-consuming task that typically requires interferometric photon correlation experiments, such as Photon correlation Fourier spectroscopy (PCFS) which measures time-resolved single emitter lineshapes. Here, we demonstrate a latent neural ordinary differential equation model that can forecast a complete and noise-free PCFS experiment from a small subset of noisy correlation functions. By encoding measured photon correlations into an initial value problem, the NODE can be propagated to an arbitrary number of interferometer delay times. We demonstrate this with 10 noisy photon correlation functions that are used to extrapolate an entire de-noised interferograms of up to 200 stage positions, enabling up to a 20-fold speedup in experimental acquisition time from $\sim$3 hours to 10 minutes. Our work presents a new approach to greatly accelerate the experimental characterization of novel quantum emitter materials using deep learning.

quant-ph↗

arXiv2024

Deconstructing equivariant representations in molecular systems

Recent equivariant models have shown significant progress in not just chemical property prediction, but as surrogates for dynamical simulations of molecules and materials. Many of the top performing models in this category are built within the framework of tensor products, which preserves equivariance by restricting interactions and transformations to those that are allowed by symmetry selection rules. Despite being a core part of the modeling process, there has not yet been much attention into understanding what information persists in these equivariant representations, and their general behavior outside of benchmark metrics. In this work, we report on a set of experiments using a simple equivariant graph convolution model on the QM9 dataset, focusing on correlating quantitative performance with the resulting molecular graph embeddings. Our key finding is that, for a scalar prediction task, many of the irreducible representations are simply ignored during training -- specifically those pertaining to vector ($l=1$) and tensor quantities ($l=2$) -- an issue that does not necessarily make itself evident in the test metric. We empirically show that removing some unused orders of spherical harmonics improves model performance, correlating with improved latent space structure. We provide a number of recommendations for future experiments to try and improve efficiency and utilization of equivariant features based on these observations.

cond-mat.mtrl-sci↗

arXiv2024

An Automated Chemical Exploration of NGC 6334I at 340 au Resolution

Much of the information gleaned from observations of star-forming regions comes from the analysis of their molecular emission spectra, particularly in the radio regime. The time-consuming nature of fitting synthetic spectra to observations interactively for such line-rich sources, however, often results in such analysis being limited to data extracted from a single-dish observation or a handful of pixels from an interferometric observation. Yet, star-forming regions display a wide variety of physical conditions that are difficult, if not impossible, to accurately characterize with such a limited number of spectra. We have developed an automated fitting routine that visits every pixel in the field of view of an ALMA data cube and determines the best-fit physical parameters, including excitation temperature and column densities, for a given list of molecules. In this proof-of-concept work, we provide an overview of the fitting routine and apply it to 0".26, 1.1 km s$^{-1}$ resolution ALMA observations of two sites of massive star-formation in NGC 6334I. Parameters were found for 21 distinct molecules by generating synthetic spectra across 7.48 GHz of spectral bandwidth between 280 and 351 GHz. Spatial images of the derived parameters for each of the > 8000 pixels are presented with special attention paid to the C$_2$H$_4$O$_2$ isomers and their relative variations. We highlight the greater scientific utility of the column density and velocity images of individual molecules compared to traditional moment maps of single transitions.

astro-ph.GA↗

arXiv2023

$ρ$-Diffusion: A diffusion-based density estimation framework for computational physics

In physics, density $ρ(\cdot)$ is a fundamentally important scalar function to model, since it describes a scalar field or a probability density function that governs a physical process. Modeling $ρ(\cdot)$ typically scales poorly with parameter space, however, and quickly becomes prohibitively difficult and computationally expensive. One promising avenue to bypass this is to leverage the capabilities of denoising diffusion models often used in high-fidelity image generation to parameterize $ρ(\cdot)$ from existing scientific data, from which new samples can be trivially sampled from. In this paper, we propose $ρ$-Diffusion, an implementation of denoising diffusion probabilistic models for multidimensional density estimation in physics, which is currently in active development and, from our results, performs well on physically motivated 2D and 3D density functions. Moreover, we propose a novel hashing technique that allows $ρ$-Diffusion to be conditioned by arbitrary amounts of physical parameters of interest.

physics.comp-ph↗

arXiv2023

Towards Foundation Models for Materials Science: The Open MatSci ML Toolkit

Artificial intelligence and machine learning have shown great promise in their ability to accelerate novel materials discovery. As researchers and domain scientists seek to unify and consolidate chemical knowledge, the case for models with potential to generalize across different tasks within materials science - so-called "foundation models" - grows with ambitions. This manuscript reviews our recent progress with development of Open MatSci ML Toolkit, and details experiments that lay the groundwork for foundation model research and development with our framework. First, we describe and characterize a new pretraining task that uses synthetic data generated from symmetry operations, and reveal complex training dynamics at large scales. Using the pretrained model, we discuss a number of use cases relevant to foundation model development: semantic architecture of datasets, and fine-tuning for property prediction and classification. Our key results show that for simple applications, pretraining appears to provide worse modeling performance than training models from random initialization. However, for more complex instances, such as when a model is required to learn across multiple datasets and types of targets simultaneously, the inductive bias from pretraining provides significantly better performance. This insight will hopefully inform subsequent efforts into creating foundation models for materials science applications.

cond-mat.mtrl-sci↗

arXiv2023

Explaining the Chemical Inventory of Orion KL through Machine Learning

The interplay of the chemistry and physics that exists within astrochemically relevant sources can only be fully appreciated if we can gain a holistic understanding of their chemical inventories. Previous work by Lee et al. (2021) demonstrated the capabilities of simple regression models to reproduce the abundances of the chemical inventory of the Taurus Molecular Cloud 1 (TMC-1), as well as provide abundance predictions for new candidate molecules. It remains to be seen, however, to what degree TMC-1 is a ``unicorn'' in astrochemistry, where the simplicity of its chemistry and physics readily facilitates characterization with simple machine learning models. Here we present an extension in chemical complexity to a heavily studied high-mass star forming region: the Orion Kleinmann-Low (Orion KL) nebula. Unlike TMC-1, Orion KL is composed of several structurally distinct environments that differ chemically and kinematically, wherein the column densities of molecules between these components can have non-linear correlations that cause the unexpected appearance or even lack of likely species in various environments. This proof-of-concept study used similar regression models sampled by Lee et al. (2021) to accurately reproduce the column densities from the XCLASS fitting program presented in Crockett et al. (2014).

astro-ph.GA↗

arXiv2023

MatSciML: A Broad, Multi-Task Benchmark for Solid-State Materials Modeling

We propose MatSci ML, a novel benchmark for modeling MATerials SCIence using Machine Learning (MatSci ML) methods focused on solid-state materials with periodic crystal structures. Applying machine learning methods to solid-state materials is a nascent field with substantial fragmentation largely driven by the great variety of datasets used to develop machine learning models. This fragmentation makes comparing the performance and generalizability of different methods difficult, thereby hindering overall research progress in the field. Building on top of open-source datasets, including large-scale datasets like the OpenCatalyst, OQMD, NOMAD, the Carolina Materials Database, and Materials Project, the MatSci ML benchmark provides a diverse set of materials systems and properties data for model training and evaluation, including simulated energies, atomic forces, material bandgaps, as well as classification data for crystal symmetries via space groups. The diversity of properties in MatSci ML makes the implementation and evaluation of multi-task learning algorithms for solid-state materials possible, while the diversity of datasets facilitates the development of new, more generalized algorithms and methods across multiple datasets. In the multi-dataset learning setting, MatSci ML enables researchers to combine observations from multiple datasets to perform joint prediction of common properties, such as energy and forces. Using MatSci ML, we evaluate the performance of different graph neural networks and equivariant point cloud networks on several benchmark tasks spanning single task, multitask, and multi-data learning scenarios. Our open-source code is available at https://github.com/IntelLabs/matsciml.

cond-mat.mtrl-sci↗

arXiv2023

Implementation of Rare Isotopologues into Machine Learning of the Chemical Inventory of the Solar-Type Protostellar Source IRAS 16293-2422

Machine learning techniques have been previously used to model and predict column densities in the TMC-1 dark molecular cloud. In interstellar sources further along the path of star formation, such as those where a protostar itself has been formed, the chemistry is known to be drastically different from that of largely quiescent dark clouds. To that end, we have tested the ability of various machine learning models to fit the column densities of the molecules detected in source B of the Class 0 protostellar system IRAS 16293-2422. By including a simple encoding of isotopic composition in our molecular feature vectors, we also examine for the first time how well these models can replicate the isotopic ratios. Finally, we report the predicted column densities of the chemically relevant molecules that may be excellent targets for radioastronomical detection in IRAS 16293-2422B.

astro-ph.GA↗

arXiv2023

Detection of Interstellar $E$-1-cyano-1,3-butadiene in GOTHAM Observations of TMC-1

We report the detection of the lowest energy conformer of $E$-1-cyano-1,3-butadiene ($E$-1-C$_4$H$_5$CN), a linear isomer of pyridine, using the fourth data reduction of the GOTHAM deep spectral survey toward TMC-1 with the 100 m Green Bank Telescope. We performed velocity stacking and matched filter analyses using Markov chain Monte Carlo simulations and find evidence for the presence of this molecule at the 5.1$σ$ level. We derive a total column density of $3.8^{+1.0}_{-0.9}\times 10^{10}$ cm$^{-2}$, which is predominantly found toward two of the four velocity components we observe toward TMC-1. We use this molecule as a proxy for constraining the gas-phase abundance of the apolar hydrocarbon 1,3-butadiene. Based on the three-phase astrochemical modeling code NAUTILUS and an expanded chemical network, our model underestimates the abundance of cyano-1,3-butadiene by a factor of 19, with a peak column density of $2.34 \times 10^{10}\ \mathrm{cm}^{-2}$ for 1,3-butadiene. Compared to the modeling results obtained in previous GOTHAM analyses, the abundance of 1,3-butadiene is increased by about two orders of magnitude. Despite this increase, the modeled abundances of aromatic species do not appear to change and remain underestimated by 1--4 orders of magnitude. Meanwhile, the abundances of the five-membered ring molecules increase proportionally with 1,3-butadiene by two orders of magnitudes. We discuss implications for bottom-up formation routes to aromatic and polycyclic aromatic molecules.

astro-ph.GA↗

arXiv2023

Astronomical Detection of the Interstellar Anion C10H- towards TMC-1 from the GOTHAM Large Program on the GBT

Using data from the GOTHAM (GBT Observations of TMC-1: Hunting for Aromatic Molecules) survey, we report the first astronomical detection of the C10H- anion. The astronomical observations also provided the necessary data to refine the spectroscopic parameters of C10H-. From the velocity stacked data and the matched filter response, C10H- is detected at >9σ confidence level at a column density of 4.04e11 cm-2. A dedicated search for the C10H radical was also conducted towards TMC-1. In this case, the stacked molecular emission of C10H was detected at a ~3.2σ confidence interval at a column density of 2.02e11 cm-2. However, since the determined confidence level is currently <5σ, we consider the identification of C10H as tentative. The full GOTHAM dataset was also used to better characterize the physical parameters including column density, excitation temperature, linewidth, and source size for the C4H, C6H and C8H radicals and their respective anions, and the measured column densities were compared to the predictions from a gas/grain chemical formation model and from a machine learning analysis. Given the measured values, the C10H-/C10H column density ratio is ~2.0 - the highest value measured between an anion and neutral species to date. Such a high ratio is at odds with current theories for interstellar anion chemistry. For the radical species, both models can reproduce the measured abundances found from the survey; however, the machine learning analysis matches the detected anion abundances much better than the gas/grain chemical model, suggesting that the current understanding of the formation chemistry of molecular anions is still highly uncertain.

astro-ph.GA↗

arXiv2022

The Open MatSci ML Toolkit: A Flexible Framework for Machine Learning in Materials Science

We present the Open MatSci ML Toolkit: a flexible, self-contained, and scalable Python-based framework to apply deep learning models and methods on scientific data with a specific focus on materials science and the OpenCatalyst Dataset. Our toolkit provides: 1. A scalable machine learning workflow for materials science leveraging PyTorch Lightning, which enables seamless scaling across different computation capabilities (laptop, server, cluster) and hardware platforms (CPU, GPU, XPU). 2. Deep Graph Library (DGL) support for rapid graph neural network prototyping and development. By publishing and sharing this toolkit with the research community via open-source release, we hope to: 1. Lower the entry barrier for new machine learning researchers and practitioners that want to get started with the OpenCatalyst dataset, which presently comprises the largest computational materials science dataset. 2. Enable the scientific community to apply advanced machine learning tools to high-impact scientific challenges, such as modeling of materials behavior for clean energy applications. We demonstrate the capabilities of our framework by enabling three new equivariant neural network models for multiple OpenCatalyst tasks and arrive at promising results for compute scaling and model performance.

cs.LG↗

arXiv2022

Discovery of Interstellar 2-Cyanoindene (2-C$_9$H$_7$CN) in GOTHAM Observations of TMC-1

We present laboratory rotational spectroscopy of five isomers of cyanoindene (2-, 4-, 5-, 6-, and 7-cyanoindene) using a cavity Fourier-transform microwave spectrometer operating between 6-40 GHz. Based on these measurements, we report the detection of 2-cyanoindene (1H-indene-2-carbonitrile; 2-C$_9$H$_7$CN) in GOTHAM line survey observations of the dark molecular cloud TMC-1 using the Green Bank Telescope at centimeter wavelengths. Using a combination of Markov Chain Monte Carlo (MCMC), spectral stacking, and matched filtering techniques, we find evidence for the presence of this molecule at the 6.3$σ$ level. This provides the first direct observation of the ratio of a cyano-substituted polycyclic aromatic hydrocarbon (PAH) to its pure hydrocarbon counterpart, in this case indene, in the same source. We discuss the possible formation chemistry of this species, including why we have only detected one of the isomers in TMC-1. We then examine the overall hydrocarbon:CN-substituted ratio across this and other simpler species, as well as compare to those ratios predicted by astrochemical models. We conclude that while astrochemical models are not yet sufficiently accurate to reproduce absolute abundances of these species, they do a good job at predicting the ratios of hydrocarbon:CN-substituted species, further solidifying -CN tagged species as excellent proxies for their fully-symmetric counterparts.

astro-ph.GA↗