GPA: 4.75 / 5.00

Relevant Coursework: Stochastic Processes, Eigenvalue of Random Matrices, Nonlinear Dynamics and Chaos, Fast Methods for Partial Differential Equations

GPA: 5.00 / 5.00

Relevant Coursework: Parallel Computing & Scientific Machine Learning, Optimization Methods, Numerical Methods for Partial Differential Equations, Introduction to Numerical Methods

GPA: 3.89 / 4.00, rank 1 / 137, honored as Weiming Bachelor (top 1%)

Relevant Coursework: Introduction to Computation, Data Structure and Algorithms, Computational Physics, Ordinary Differential Equations, Mathematical Method in Physics, Advanced Mathematics I, II, Advanced Algebra I, II

GPA: 4.00 / 4.00

Relevant Coursework: Introduction to Probability, Applied Numerical Methods

Advisor: Qilin He, Jiaqi Jiang (Platform Architecture)

• Fine-tuned pre-trained machine learning surrogate models for physical simulations, achieving 18× speedup over traditional methods while maintaining high accuracy
• Developed algorithms for differentiable physical solvers, enabling gradient-based optimization which is much more efficient than zeroth-order methods

Advisor: Yuan Lin (Deep Learning Compiler Team)

• Improved the heuristics of layer fusion (matrix multiplication or convolution combined with element-wise functions), for running transformer-based large language models (LLMs) efficiently on GPUs with TensorRT
• Designed a data-driven heuristic framework for fusion auto-tuning, enhancing search quality with 1.8× optimal tactic coverage.
• Reduced compilation time by 40% with caching, multi-threading and better auto-tuning strategies, based on profiling-driven analysis of layer fusion pipeline in TensorRT

Advisor: Christopher Rackauckas & Alan Edelman

• Developing advanced solvers for nonlinear equations and ordinary differential equations (ODEs) where higher-order derivatives can be used to enhance convergence and efficiency

Advisor: Christopher Rackauckas & Alan Edelman

• Developed higher-order forward-mode automatic differentiation algorithms that scale linearly with the order, suitable for physics-informed neural networks (PINNs) which require multiple order derivatives
• Generated efficient higher-order differentiation rules (aka primitives) automatically from first-order chain rules with symbolic computation and metaprogramming in Julia
• Composed the algorithm with existing first-order reverse-mode automatic differentiation libraries like Zygote.jl to establish a mixed-mode strategy for efficient PINN training

Advisor: Christopher Rackauckas & Alan Edelman

• Worked as a part of the Enzyme project, an automatic differentiation framework based on source code transformation at LLVM intermediate representation (IR) level, which can differentiate through all languages with a LLVM backend (e.g. Julia, C++, Fortran)
• Implemented algorithms that synthesize derivatives of BLAS/LAPACK kernels which are commonly used in scientific computing, and performed extensive optimizations based on linear algebra relations
• Outperformed other high-level AD frameworks in Julia with 1.3× speed on a linear algebra benchmark set

Advisor: Weinan E & Linfeng Zhang

• Modeled the exchange-correlation density functional in generalized Kohn-Sham theory with deep neural networks and descriptors from density matrices
• Established a method for using physical quantity labels that depend on the functional minimization result to train the functional model, in other words, addressed the "differentiate through argmin" problem
• Implemented the training process with multiple types of physical quantity label, such as energy levels of molecular orbitals and dipole moment
• Improved the accuracy and generalization performance of the model, obtained an average energy error of 0.06 kcal/mol on a test set that includes 1200 water molecule configurations labeled with SCAN0 functional (48% less than previous methods)

Advisor: Bin Zhang

• Applied noise contrastive learning for unsupervised training of the potential energy surface of coarse-grained molecular systems, which is a high-dimensional and unnormalized probability distribution (such that it cannot be learned with maximum likelihood estimation)
• Utilized normalizing flow-based neural network architecture to construct the "noise", enabling computationally efficient sampling, density evaluation, and gradient computation
• Obtained an efficient and systematic coarse-graining workflow, outperforming traditional force-mapping scheme which is inefficient to train and lacks transferability