Xiangjian Zhang

Xiangjian Zhang

Graduate Student, Applied Math and Computational science (AMCS), University of Pennsylvania

About Me

I am currently a master’s student in Applied Mathematics & Computational Science at the University of Pennsylvania. My research includes integrating numerical methods into deep learning architectures. I have strong experience in machine learning model design, numerical methods, and real-time data processing.

Research Interests

Keywords: Deep Learning, Differential Equations, Computational Mathematics

My academic interests lie in the fields of computational mathematics and machine learning. I am particularly passionate about developing algorithms and models that enhance our understanding and application of these technologies. My work includes research on finite difference methods to solve fluid dynamics problems and theoretical studies related to neural networks. I aim to continue exploring these areas and contribute to the advancement of artificial intelligence and its applications.

Work experience

Meituan

Duration: February 2025 – July 2025

Role: Algorithm Intern

Description: Designed a Transformer-based architecture for deep learning models in time series forecasting, enabling accurate and efficient prediction of sequential patterns in business data. Implemented an online learning training strategy, allowing the model to continuously adapt to newly incoming data streams and maintain high performance in dynamic, real-time environments.

Shanghai UniDT Technology

Duration: July 2023 – August 2023

Role: Algorithm Intern

Description: Designed a rule-based recognition method to accurately identify hierarchical structures such as chapters, paragraphs, and clauses by analyzing the tags in the text (such as title style, numbering, and specific key words) and combining it with regular expression matching technology. Introduced a context-aware segmentation algorithm, combining the context information of the sentence, such as the preceding and following words, sentence length, semantic relevance, etc., and solving the problem of fuzzy segmentation boundaries through a machine learning model.

Projects

Backward Reconstruction of Initial Vorticity Fields Using Frequency-Domain Neural Operators (University of Pennsylvania)

Duration: November 2025 – December 2025

Role: Project Leader, Department of Mathematics

Description: Investigated the backward reconstruction of initial vorticity fields for two-dimensional advection–diffusion systems, an intrinsically ill-posed inverse problem due to diffusion-induced information loss and advective mixing. Constructed a large-scale synthetic dataset consisting of 2000 high-resolution numerical trajectories generated from a discretized PDE solver. Designed and implemented a variable-step inverse operator learning framework to stabilize backward-time inference and mitigate long-horizon error accumulation. Conducted a systematic comparison between Fourier Neural Operators (FNO) and frequency-domain Vision Transformer architectures, analyzing reconstruction accuracy, spectral stability, and robustness to noise. Demonstrated that spectral inductive bias is critical for inverse PDE learning, with FNO producing stable, physically meaningful reconstructions while Transformer-based models suffer severe high-frequency amplification.

Runge-Kutta-Fehlberg Transformer for Robust Small-Target Feature Representation (Huazhong University of Science and Technology)

Duration: July 2024 – November 2024

Role: Research Intern, School of Computer Science and Technology

Description: Designed and implemented a Runge-Kutta-Fehlberg Transformer model that integrates the long-range dependency modeling capabilities of Transformer architectures with the local feature extraction strengths of convolutional neural networks, improving representation quality for small targets in complex environments. Introduced a random connection attention module to replace standard self-attention, enabling sparse attention learning that significantly reduced computational and memory overhead while maintaining or improving model accuracy. Further enhanced model robustness by stacking specially designed Runge-Kutta-Fehlberg Transformer modules, leveraging numerical integration principles to suppress noise, stabilize feature propagation, and improve performance under challenging conditions.

Optimizing the Marker-and-Cell Scheme for Accurate Fluid Dynamics in Lid-Driven Cavity Flow (Huazhong University of Science and Technology)

Duration: February 2024 – May 2024

Role: Research Intern

Description: Investigated the implementation of the Marker-and-Cell (MAC) scheme to effectively model fluid dynamics in Lid-driven cavity flow, focusing on accuracy and computational efficiency in numerical simulations. Analyzed the convergence behavior and stability of the MAC scheme, optimizing parameters to enhance performance and ensure reliable results in simulating complex fluid flow patterns within the cavity.Introduced the Kronecker product method to design matrix structures, effectively minimizing memory consumption and enhancing the overall efficiency of the numerical simulations for the Lid-driven cavity flow.

Integrating Eligibility Traces and Top-Down Signals for Improved Learning in Spiking Neural Networks (Shanghai Jiao Tong University)

Duration: January 2024 – February 2023

Role: Research Intern

Description: Introduced Eligibility Traces to record the contribution of synapses over a period of time to a multi-layer spiking neural network constructed using Leaky Integrate-and-Fire to simulate the spiking mechanism of biological neurons.Introduced Top-Down Learning Signals to carry target information or expected output, guide or adjust the learning process of lower-layer neurons, and helped the network learn more accurately based on feedback signals from higher-level or more complex neural network modules.

LIF Model: From Theoretical Construction to Application (University of Cambridge)

Duration: August 2022 – September 2022

Role: Research Intern

Description: Designed the simplified Leaky Integrate-and-Fire model to describe the change of neuron membrane potential over time, considering the different weights of the input currents of excitatory and inhibitory neurons respectively to construct neural network. Introduced numerical methods to discretize the differential equations and to calculate the update of the membrane potential at each time step, adjusting the ratio of excitatory neurons to inhibitory neurons, and recording the spike pulse firing frequency of the entire neural network.

Differential Manifolds and Convex Analysis (Huazhong University of Science and Technology)

Duration: January 2022 – April 2023

Role: Research Intern

Description: Analyzed the properties of epigraphs, such as closure and convexity, and their applications in optimization theory such as soft threshold function in L1 regularization and hinge loss in support vector machines to solve practical problems such as image reconstruction and data fitting. Introduced Fourier transform to transform the differential operation in partial differential equation into algebraic operation, simplifying the solution process of PDE and obtaining the solution infrequency domain.

Awards & Scholarship

Contact

Email: chkirin@sas.upenn.edu

Address: 209 S. 33rd Street, Philadelphia, Pennsylvania