Projects

A curated dataset of polymer molecular dynamics simulations compiled from 823 peer-reviewed papers (1995–2025). Built through automated literature mining, classification, and information extraction, the PolyDatabase provides structured simulation metadata for data-driven polymer informatics and molecular modeling.

Smart Ant Monitoring is an integrated AI- and IoT-based pest surveillance system developed for citrus orchards. It combines real-time computer vision and environmental sensing to monitor ant activity and predict mealybug infestations—two pests that are closely correlated in orchard ecosystems. The system uses a Raspberry Pi 5 and a Google Coral USB Accelerator running YOLOv8-OBB models enhanced with SAHI slicing to achieve over 90% detection precision even under challenging field conditions. The device automatically detects and counts ants every hour (or as scheduled) and simultaneously records temperature, humidity, irrigation status, and soil moisture to enhance predictive modeling of mealybug presence. All collected data are automatically transmitted to a centralized database every 10 minutes, where they can be visualized and analyzed through a web portal that provides live tracking, analytics, and insights for effective data-driven pest management.

Piano Music Generation employs an auto-regressive transformer architecture composed of stacked TransformerEncoder layers to model symbolic piano music. MIDI data are converted into discrete event tokens encoding note on/off, time shifts, and velocity changes. Using multi-head self-attention and positional encoding, the model captures long-range temporal and harmonic dependencies, enabling the generation of musically coherent and expressive piano compositions.

Our ActiveLearningDeepMD framework is designed to iteratively select informative data points from first-principles molecular dynamics simulations in high-dimensional chemical space. Rather than exhaustively sampling all configurations, it uses uncertainty estimation and query strategies to focus computational effort where the model is least confident. The resulting subset of data is then used to train or refine deep potential models, leading to accurate atomistic representations with far fewer expensive simulation runs.