arXiv subjects

Gowoon Cheon

Publications and source records attributed to Gowoon Cheon.

9 recordsLinked to original sources

arXiv2025

CURIE: Evaluating LLMs On Multitask Scientific Long Context Understanding and Reasoning

Scientific problem-solving involves synthesizing information while applying expert knowledge. We introduce CURIE, a scientific long-Context Understanding,Reasoning and Information Extraction benchmark to measure the potential of Large Language Models (LLMs) in scientific problem-solving and assisting scientists in realistic workflows. This benchmark introduces ten challenging tasks with a total of 580 problems and solution pairs curated by experts in six disciplines - materials science, condensed matter physics, quantum computing, geospatial analysis, biodiversity, and proteins - covering both experimental and theoretical work-flows in science. We evaluate a range of closed and open LLMs on tasks in CURIE which requires domain expertise, comprehension of long in-context information,and multi-step reasoning. While Gemini Flash 2.0 and Claude-3 show consistent high comprehension across domains, the popular GPT-4o and command-R+ fail dramatically on protein sequencing tasks. With the best performance at 32% there is much room for improvement for all models. We hope that insights gained from CURIE can guide the future development of LLMs in sciences. Evaluation code and data are in https://github.com/google/curie

cs.CL↗

arXiv2023

Accurate Prediction of Experimental Band Gaps from Large Language Model-Based Data Extraction

Machine learning is transforming materials discovery by providing rapid predictions of material properties, which enables large-scale screening for target materials. However, such models require training data. While automated data extraction from scientific literature has potential, current auto-generated datasets often lack sufficient accuracy and critical structural and processing details of materials that influence the properties. Using band gap as an example, we demonstrate Large language model (LLM)-prompt-based extraction yields an order of magnitude lower error rate. Combined with additional prompts to select a subset of experimentally measured properties from pure, single-crystalline bulk materials, this results in an automatically extracted dataset that's larger and more diverse than the largest existing human-curated database of experimental band gaps. Compared to the existing human-curated database, we show the model trained on our extracted database achieves a 19% reduction in the mean absolute error of predicted band gaps. Finally, we demonstrate that LLMs are able to train models predicting band gap on the extracted data, achieving an automated pipeline of data extraction to materials property prediction.

cond-mat.mtrl-sci↗

arXiv2020

Dataset of Random Relaxations for Crystal Structure Search of Li-Si System

Crystal structure search is a long-standing challenge in materials design. We present a dataset of more than 100,000 structural relaxations of potential battery anode materials from randomized structures using density functional theory calculations. We illustrate the usage of the dataset by training graph neural networks to predict structural relaxations from randomly generated structures. Our models directly predict stresses in addition to forces, which allows them to accurately simulate relaxations of both ionic positions and lattice vectors. We show that models trained on the molecular dynamics simulations fail to simulate relaxations from random structures, while training on our data leads to up to two orders of magnitude decrease in error for the same task. Our model is able to find an experimentally verified structure of a stoichiometry held out from training. We find that randomly perturbing atomic positions during training improves both the accuracy and out of domain generalization of the models.

cond-mat.mtrl-sci↗

arXiv2020

Discovery of stable surfaces with extreme work functions by high-throughput density functional theory and machine learning

The work function is the key surface property that determines how much energy is required for an electron to escape the surface of a material. This property is crucial for thermionic energy conversion, band alignment in heterostructures, and electron emission devices. Here, we present a high-throughput workflow using density functional theory (DFT) to calculate the work function and cleavage energy of 33,631 slabs (58,332 work functions) that we created from 3,716 bulk materials, including up to ternary compounds. The number of materials for which we calculated surface properties surpasses the previously largest database, the Materials Project, by a factor of $\sim$27. On the tail ends of the work function distribution we identify 34 and 56 surfaces with an ultra-low (<2 eV) and ultra-high (>7 eV) work function, respectively. Further, we discover that the $(100)$-Ba-O surface of BaMoO$_3$ and the $(001)$-F surface of Ag$_2$F have record-low (1.25 eV) and record-high (9.06 eV) steady-state work functions without requiring coatings, respectively. Based on this database we develop a physics-based approach to featurize surfaces and use supervised machine learning to predict the work function. We find that physical choice of features improves prediction performance far more than choice of model. Our random forest model achieves a mean absolute test error of 0.09 eV, which is more than 6 times better than the baseline and comparable to the accuracy of DFT. This surrogate model enables rapid predictions of the work function ($\sim 10^5$ faster than DFT) across a vast chemical space and facilitates the discovery of material surfaces with extreme work functions for energy conversion, electronic applications, and contacts in 2-dimensional devices.

cond-mat.mtrl-sci↗

arXiv2020

A data-centric framework for crystal structure identification in atomistic simulations using machine learning

Atomic-level modeling performed at large scales enables the investigation of mesoscale materials properties with atom-by-atom resolution. The spatial complexity of such cross-scale simulations renders them unsuitable for simple human visual inspection. Instead, specialized structure characterization techniques are required to aid interpretation. These have historically been challenging to construct, requiring significant intuition and effort. Here we propose an alternative framework for a fundamental structural characterization task: classifying atoms according to the crystal structure to which they belong. Our approach is data-centric and favors the employment of Machine Learning over heuristic rules of classification. A group of data-science tools and simple local descriptors of atomic structure are employed together with an efficient synthetic training set. We also introduce the first standard and publicly available benchmark data set for evaluation of algorithms for crystal-structure classification. It is demonstrated that our data-centric framework outperforms all of the most popular heuristic methods -- especially at high temperatures when lattices are the most distorted -- while introducing a systematic route for generalization to new crystal structures. Moreover, through the use of outlier detection algorithms our approach is capable of discerning between amorphous atomic motifs (i.e., noncrystalline phases) and unknown crystal structures, making it uniquely suited for exploratory materials synthesis simulations.

cond-mat.mtrl-sci↗

arXiv2020

The Joint Automated Repository for Various Integrated Simulations (JARVIS) for data-driven materials design

The Joint Automated Repository for Various Integrated Simulations (JARVIS) is an integrated infrastructure to accelerate materials discovery and design using density functional theory (DFT), classical force-fields (FF), and machine learning (ML) techniques. JARVIS is motivated by the Materials Genome Initiative (MGI) principles of developing open-access databases and tools to reduce the cost and development time of materials discovery, optimization, and deployment. The major features of JARVIS are: JARVIS-DFT, JARVIS-FF, JARVIS-ML, and JARVIS-Tools. To date, JARVIS consists of 40,000 materials and 1 million calculated properties in JARVIS-DFT, 1,500 materials and 110 force-fields in JARVIS-FF, and 25 ML models for material-property predictions in JARVIS-ML, all of which are continuously expanding. JARVIS-Tools provides scripts and workflows for running and analyzing various simulations. We compare our computational data to experiments or high-fidelity computational methods wherever applicable to evaluate error/uncertainty in predictions. In addition to the existing workflows, the infrastructure can support a wide variety of other technologically important applications as part of the data-driven materials design paradigm. The JARVIS datasets and tools are publicly available at the website: https://jarvis.nist.gov .

cond-mat.mtrl-sci↗

arXiv2019

Quantifying the search for solid Li-ion electrolyte materials by anion: a data-driven perspective

We compile data and machine learned models of solid Li-ion electrolyte performance to assess the state of materials discovery efforts and build new insights for future efforts. Candidate electrolyte materials must satisfy several requirements, chief among them fast ionic conductivity and robust electrochemical stability. Considering these two requirements, we find new evidence to suggest that optimization of the sulfides for fast ionic conductivity and wide electrochemical stability may be more likely than optimization of the oxides, and that the oft-overlooked chlorides and bromides may be particularly promising families for Li-ion electrolytes. We also find that the nitrides and phosphides appear to be the most promising material families for electrolytes stable against Li-metal anodes. Furthermore, the spread of the existing data in performance space suggests that fast conducting materials that are stable against both Li metal and a >4V cathode are exceedingly rare, and that a multiple-electrolyte architecture is a more likely path to successfully realizing a solid-state Li metal battery by approximately an order of magnitude or more. Our model is validated by its reproduction of well-known trends that have emerged from the limited existing data in recent years, namely that the electronegativity of the lattice anion correlates with ionic conductivity and electrochemical stability. In this work, we leverage the existing data to make solid electrolyte performance trends quantitative for the first time, building a roadmap to complement material discovery efforts around desired material performance.

cond-mat.mtrl-sci↗

arXiv2018

Machine learning-assisted discovery of many new solid Li-ion conducting materials

We discover many new crystalline solid materials with fast single crystal Li ion conductivity at room temperature, discovered through density functional theory simulations guided by machine learning-based methods. The discovery of new solid Li superionic conductors is of critical importance to the development of safe all-solid-state Li-ion batteries. With a predictive universal structure-property relationship for fast ion conduction not well understood, the search for new solid Li ion conductors has relied largely on trial-and-error computational and experimental searches over the last several decades. In this work, we perform a guided search of materials space with a machine learning (ML)-based prediction model for material selection and density functional theory molecular dynamics (DFT-MD) simulations for calculating ionic conductivity. These materials are screened from over 12,000 experimentally synthesized and characterized candidates with very diverse structures and compositions. When compared to a random search of materials space, we find that the ML-guided search is 2.7 times more likely to identify fast Li ion conductors, with at least a 45x improvement in the log-average of room temperature Li ion conductivity. The F1 score of the ML-based model is 0.50, 3.5 times better than the F1 score expected from completely random guesswork. In a head-to-head competition against six Ph.D. students working in the field, we find that the ML-based model doubles the F1 score of human experts in its ability to identify fast Li-ion conductors from atomistic structure with a thousand- fold increase in speed, clearly demonstrating the utility of this model for the research community. All conducting materials reported here lack transition metals and are predicted to exhibit low electronic conduction, high stability against oxidation, and high thermodynamic stability.

cond-mat.mtrl-sci↗

arXiv2018

Elastic properties of bulk and low-dimensional materials using Van der Waals density functional

In this work, we present a high-throughput first-principles study of elastic properties of bulk and monolayer materials mainly using the vdW-DF-optB88 functional. We discuss the trends on the elastic response with respect to changes in dimensionality. We identify a relation between exfoliation energy and elastic constants for layered materials that can help to guide the search for vdW bonding in materials. We also predicted a few novel materials with auxetic behavior. The uncertainty in structural and elastic properties due to the inclusion of vdW interactions is discussed. We investigated 11,067 bulk and 257 monolayer materials. Lastly, we found that the trends in elastic constants for bulk and their monolayer counterparts can be very different. All the computational results are made publicly available at easy-to-use websites: https://www.ctcms.nist.gov/~knc6/JVASP.html and https://jarvis.nist.gov/ . Our dataset can be used to identify stiff and flexible materials for industrial applications.

cond-mat.mtrl-sci↗