Curriculum Vitae

New Haven, CT

Ph.D. Candidate in Statistics & Data Science

GPA 4.0/4.0

Philadelphia, PA

M.S.E. in Data Science

Outstanding Academic Award

Tianjin, China

B.S. in Applied Physics

Top 5%, Yan Jici Physics Elites Class

National Scholarship

Applied Scientist Intern, Buyer Risk Prevention | Agent Planning & Tool Orchestration

Seattle, WA

• Developed DAGENT, an end-to-end agentic planning and execution framework for autonomous business metric investigations. Designed a schema-constrained procedural knowledge graph (PKG) to transform past investigations into reusable memory. Implemented multi-case retrieval and objective decomposition with graph-grounded planning to construct dependency-aware execution DAGs for parallel tool orchestration and evidence-guided replanning.
• PKG-grounded planning achieved 2.17× the accuracy of RAG on SOP-Bench, while DAGENT outperformed the strongest graph-planning baseline by 12.6 percentage points. Integrated into an Amazon deep-dive product.

Machine Learning and Fire Modeling Intern

Neural Operators & Radiative Transfer Surrogate

Norwood, MA

• Developed nested Fourier-MIONet surrogates for the 5D radiative transfer equation in 2D and 3D fire simulations, achieving ∼3% relative L₂ error and faster inference than FireFOAM, with generalization across 14–58 kW heat-release rates.
• Implemented PCA-, KAN-, and Fourier-enhanced MIONets on 2D pool fires and scaled to large 3D McCaffrey fires (∼2.7k grid points).
• Engineered an HPC-to-cloud workflow spanning HPC clusters, Amazon S3, and AWS SageMaker for efficient data transfer, processing, and training on FireFOAM simulation datasets.

Graduate Research Assistant

Advisor: Prof. Lu Lu

New Haven, CT

• One-shot neural operator learning

  Proposed a data-efficient method for PDE solution operators that learns from only one data pair by leveraging the locality of derivatives, enabling generalization to varied new inputs and geometries.

• Symmetry-constrained scientific law discovery

  Developed a multi-agent reinforcement learning framework with structure-aware LSTMs and physical symmetry constraints to discover interpretable constitutive laws in crystal plasticity across 120 crystal orientations and 12 interacting slip systems. Identified nonlinear pairwise coupling beyond conventional additive hardening, reducing shear-rate relative L₂ error from 18.1% to 9.7%.

• Physics-informed neural operators on unknown manifolds

  Developed physics-informed DeepONet for PDEs on unknown manifolds, using RBF kernels and diffusion maps to approximate differential operators. Integrated the method into a Bayesian MCMC framework to estimate diffusion coefficients from sparse, noisy point-cloud observations.

• Physics-informed ML for chromatin inference

  Developed a parallel physics-informed neural network using mixed formulation and domain decomposition to infer spatially varying biophysical parameters (methylation rates and chromatin–lamina affinity) from single-cell STORM microscopy data, demonstrating robust recovery under up to 5% observational noise in synthetic benchmarks.

• Time-adaptive neural operator learning

  Developed TANTE, a time-adaptive Transformer-based operator-learning framework using neural Taylor expansion for continuous-time PDE prediction, achieving 60–80% performance improvements and 30–40% inference speedups over baselines.

Research Assistant

Lu group, Scientific Machine Learning

Philadelphia, PA

• Out-of-distribution operator adaptation

  Implemented DeepONet fine-tuning with new observations to improve out-of-distribution generalization of learned PDE solution operators.

• Open-source scientific ML

  Contributed to DeepXDE, an open-source library for scientific ML and AI for PDEs. Implemented physics-informed neural networks (PINNs) to solve complex inverse problems and parameter estimation tasks in heterogeneous domains, including photonic nanostructures, quantitative finance, and chromatin organization.