arXiv subjects

Sungwoo Kang

Publications and source records attributed to Sungwoo Kang.

24 records · Page 2Linked to original sources

arXiv2025

An efficient forgetting-aware fine-tuning framework for pretrained universal machine-learning interatomic potentials

Pretrained universal machine-learning interatomic potentials (MLIPs) have revolutionized computational materials science by enabling rapid atomistic simulations as efficient alternatives to ab initio methods. Fine-tuning pretrained MLIPs offers a practical approach to improving accuracy for materials and properties where predictive performance is insufficient. However, this approach often induces catastrophic forgetting, undermining the generalizability that is a key advantage of pretrained MLIPs. Herein, we propose reEWC, an advanced fine-tuning strategy that integrates Experience Replay and Elastic Weight Consolidation (EWC) to effectively balance forgetting prevention with fine-tuning efficiency. Using Li$_6$PS$_5$Cl (LPSC), a sulfide-based Li solid-state electrolyte, as a fine-tuning target, we show that reEWC significantly improves the accuracy of a pretrained MLIP, resolving well-known issues of potential energy surface softening and overestimated Li diffusivities. Moreover, reEWC preserves the generalizability of the pretrained MLIP and enables knowledge transfer to chemically distinct systems, including other sulfide, oxide, nitride, and halide electrolytes. Compared to Experience Replay and EWC used individually, reEWC delivers clear synergistic benefits, mitigating their respective limitations while maintaining computational efficiency. These results establish reEWC as a robust and effective solution for continual learning in MLIPs, enabling universal models that can advance materials research through large-scale, high-throughput simulations across diverse chemistries.

cond-mat.mtrl-sci↗

arXiv2025

Stock Price Prediction Using Triple Barrier Labeling and Raw OHLCV Data: Evidence from Korean Markets

This paper demonstrates that deep learning models trained on raw OHLCV (open-high-low-close-volume) data can achieve comparable performance to traditional machine learning (ML) models using technical indicators for stock price prediction in Korean markets. While previous studies have emphasized the importance of technical indicators and feature engineering, we show that a simple LSTM network trained on raw OHLCV data alone can match the performance of sophisticated ML models that incorporate technical indicators. Using a dataset of Korean stocks from 2006 to 2024, we optimize the triple barrier labeling parameters to achieve balanced label proportions with a 29-day window and 9\% barriers. Our experiments reveal that LSTM networks achieve similar performance to traditional machine learning models like XGBoost, despite using only raw OHLCV data without any technical indicators. Furthermore, we identify that the optimal window size varies with model hidden size, with a configuration of window size 100 and hidden size 8 yielding the best performance. Additionally, our results confirm that using full OHLCV data provides better predictive accuracy compared to using only close price or close price with volume. These findings challenge conventional approaches to feature engineering in financial forecasting and suggest that simpler approaches focusing on raw data and appropriate model selection may be more effective than complex feature engineering strategies.

cs.CE↗

arXiv2024

How Graph Neural Network Interatomic Potentials Extrapolate: Role of the Message-Passing Algorithm

Graph neural network interatomic potentials (GNN-IPs) are gaining significant attention due to their capability of learning from large datasets. Specifically, universal interatomic potentials based on GNN, usually trained with crystalline geometries, often exhibit remarkable extrapolative behavior towards untrained domains, such as surfaces or amorphous configurations. However, the origin of this extrapolation capability is not well understood. This work provides a theoretical explanation of how GNN-IPs extrapolate to untrained geometries. First, we demonstrate that GNN-IPs can capture non-local electrostatic interactions through the message-passing algorithm, as evidenced by tests on toy models and DFT data. We find that GNN-IP models, SevenNet and MACE, accurately predict electrostatic forces in untrained domains, indicating that they have learned the exact functional form of the Coulomb interaction. Based on these results, we suggest that the ability to learn non-local electrostatic interactions, coupled with the embedding nature of GNN-IPs, explains their extrapolation ability. We find that the universal GNN-IP, SevenNet-0, effectively infers non-local Coulomb interactions in untrained domains but fails to extrapolate the non-local forces arising from the kinetic term, which supports the suggested theory. Finally, we address the impact of hyperparameters on the extrapolation performance of universal potentials, such as SevenNet-0 and MACE-MP-0, and discuss the limitations of the extrapolation capabilities.

cond-mat.mtrl-sci↗

arXiv2024

Using a Deep Learning Model to Simulate Human Stock Trader's Methods of Chart Analysis

Despite the efficient market hypothesis, many studies suggest the existence of inefficiencies in the stock market leading to the development of techniques to gain above-market returns. Systematic trading has undergone significant advances in recent decades with deep learning schemes emerging as a powerful tool for analyzing and predicting market behavior. In this paper, a method is proposed that is inspired by how professional technical analysts trade. This scheme looks at stock prices of the previous 600 days and predicts whether the stock price will rise or fall 10% or 20% within the next D days. The proposed method uses the Resnet's (a deep learning model) skip connections and logits to increase the probability of the prediction. The model was trained and tested using historical data from both the Korea and US stock markets. The backtest is done using the data from 2020 to 2022. Using the proposed method for the Korea market it gave return of 75.36% having Sharpe ratio of 1.57, which far exceeds the market return by 36% and 0.61, respectively. On the US market it gives total return of 27.17% with Sharpe ratio of 0.61, which outperforms other benchmarks such as NASDAQ, S&P500, DOW JONES index by 17.69% and 0.27, respectively.

q-fin.ST↗

arXiv2021

Accelerated identification of equilibrium structures of multicomponent inorganic crystals using machine learning potentials

The discovery of new multicomponent inorganic compounds can provide direct solutions to many scientific and engineering challenges, yet the vast size of the uncharted material space dwarfs current synthesis throughput. While the computational crystal structure prediction is expected to mitigate this frustration, the NP-hardness and steep costs of density functional theory (DFT) calculations prohibit material exploration at scale. Herein, we introduce SPINNER, a highly efficient and reliable structure-prediction framework based on exhaustive random searches and evolutionary algorithms, which is completely free from empiricism. Empowered by accurate neural network potentials, the program can navigate the configuration space faster than DFT by more than 10$^{2}$-fold. In blind tests on 60 ternary compositions diversely selected from the experimental database, SPINNER successfully identifies experimental (or theoretically more stable) phases for ~80% of materials within 5000 generations, entailing up to half a million structure evaluations for each composition. When benchmarked against previous data mining or DFT-based evolutionary predictions, SPINNER identifies more stable phases in the majority of cases. By developing a reliable and fast structure-prediction framework, this work opens the door to large-scale, unbounded computational exploration of undiscovered inorganic crystals.

physics.comp-ph↗

arXiv2020

Training machine-learning potentials for crystal structure prediction using disordered structures

Prediction of the stable crystal structure for multinary (ternary or higher) compounds with unexplored compositions demands fast and accurate evaluation of free energies in exploring the vast configurational space. The machine-learning potential such as the neural network potential (NNP) is poised to meet this requirement but a dearth of information on the crystal structure poses a challenge in choosing training sets. Herein we propose constructing the training set from densityfunctional-theory (DFT) based dynamical trajectories of liquid and quenched amorphous phases, which does not require any preceding information on material structures except for the chemical composition. To demonstrate suitability of the trained NNP in the crystal structure prediction, we compare NNP and DFT energies for Ba2AgSi3, Mg2SiO4, LiAlCl4, and InTe2O5F over experimental phases as well as low-energy crystal structures that are generated theoretically. For every material, we find strong correlations between DFT and NNP energies, ensuring that the NNPs can properly rank energies among low-energy crystalline structures. We also find that the evolutionary search using the NNPs can identify low-energy metastable phases more efficiently than the DFTbased approach. By proposing a way to developing reliable machine-learning potentials for the crystal structure prediction, this work will pave the way to identifying unexplored multinary phases efficiently.

physics.comp-ph↗