Research topics

My research connects machine learning with scientific computing. I am particularly interested in learning from partial observations, representing physical structure, and evaluating models through reproducible experiments.

Partial observations & neural operators

Many scientific systems are observed through sparse sensors or incomplete measurements. I study neural operators and inference methods that recover full fields from these observations, and how their predictions change across resolutions and observation patterns.

PDE-OBS: Controlled Evaluation Across Observation Patterns

A benchmark for physical-field reconstruction and forecasting under changing measurement density and spatial layout. PDE-OBS combines 560,000 fields and trajectories from seven PDE families with configurable observation operators and seven adapted baseline methods, keeping prediction targets and data splits fixed across observation patterns.

2025–present · PDE learning · Partial observations · Benchmarking

arXiv · PDF · Code · Benchmark resources

Related papers: Discretization mismatch in neural operators · Dynamic Schwartz–Fourier Neural Operator · Coordinate Transform FNO

Predictive representation learning

I investigate joint-embedding predictive architectures for scientific data, including student–teacher learning, latent full-field prediction, and representations across spatial resolutions.

  • SCOPE. Observation-conditioned full-target prediction and physical reconstruction for single-pass sparse PDE inference. [PDF] [Code]
  • JENO. Full-field latent prediction for sparse inverse PDE inference.
  • HP-JEPA. Hierarchical partitioning for multi-resolution graph joint-embedding predictive learning.

2026–present · JEPA · PDE inference · Graph representation learning

arXiv preprints · Submitted manuscripts

Physics-aware generation & optimization

I develop sampling and optimization methods that incorporate physical constraints, combining diffusion models, simulated annealing, and reinforcement learning.

  • APOD. Adaptive PDE-observation diffusion for physics-constrained sampling.
  • RL-QESA. Reinforcement-learning quasi-equilibrium simulated annealing.
  • Structure-aware dynamics. Hamiltonian and symplectic learning, including Kolmogorov–Arnold representations and learning dynamics from position-only observations.

APOD paper · RL-QESA paper · KAR-HNN paper · Position-only dynamics

Molecular generation & LLM-guided search

My work in molecular design combines generative modeling and optimization under structural and validity constraints.

ORACLE: LLM-guided molecular optimization

LLM-proposed molecular edits are combined with simulated annealing for multi-objective structure-based drug design. Evaluation considers docking, QED, synthetic accessibility, diversity, and Pareto trade-offs.

2025–present · LLMs · Simulated annealing · Molecular design

Related work: GAGA: 3D molecular generation · Generative backmapping of coarse-grained structures

Publications

See the complete publication list for arXiv preprints, conference, journal, and workshop papers, and submitted manuscripts.