Scientific Machine Learning × Graph Neural Networks × Inverse Problems

Learning complex physical systems with scientific machine learning.

I am Sanjian Zhang, a Ph.D. student in Interdisciplinary Engineering at Kennesaw State University. My research focuses on scientific machine learning, graph neural networks, and inverse problems for complex physical systems. My current work investigates inverse modeling for patient-specific cardiac digital twins. Given pressure-dependent 3D cardiac deformation, I study learning-based methods for estimating nonlinear anisotropic material properties and latent biomechanical states using geometric deep learning and physics-based simulation data. More broadly, I am interested in AI for Science, geometric deep learning, surrogate modeling, and spatiotemporal learning, with an emphasis on integrating data-driven methods with physics-based simulation.

View Projects Publications Contact Me Google Scholar

Deformation + Pressure → Inverse Learning → Material Properties

Research focus

My work asks how scientific machine learning can recover hidden biomechanical properties from observed physical behavior. I develop geometric learning methods that operate on unstructured cardiac meshes and connect inverse models with physics-based simulations and forward surrogates.

Inverse identification of biomechanical material properties from pressure-dependent cardiac deformation Graph neural networks and geometric deep learning on unstructured 3D cardiac meshes Physics-constrained surrogate modeling for patient-specific cardiac digital twins

News

Recent updates and milestones. Scroll to view earlier items.

Apr 2026

Began research as a Research Assistant at the University of Mississippi.

Feb 2026

Received Ph.D. offer in Interdisciplinary Engineering at Kennesaw State University.

Dec 2025

Received M.S. degree from the University of Washington.

Aug 2025

One paper accepted at Findings of EMNLP 2025.

Jun 2025

Awarded Best Paper Award at CVPR Workshop 2025.

May 2025

Invited Reviewer for ACM Transactions on the Web (TWEB).

Jan 2025

Won 5th Place in the Cell Behavior Video Classification Challenge (CBVCC), Lugano.

Selected Projects / Systems

Research systems that show both scientific direction and engineering depth.

Ongoing Direction

Inverse Modeling for Patient-Specific Cardiac Digital Twins

Investigating scientific machine learning methods for inferring biomechanical material properties from pressure-dependent 3D cardiac deformation.

Problem: Estimate nonlinear anisotropic myocardial material properties from observed cardiac deformation under known pressure conditions.
Methods: Graph neural networks, geometric deep learning, inverse modeling, physics-constrained learning, and forward surrogate models.
Data: Learning from approximately 5,000 finite-element simulation cases spanning multiple material configurations and pressure states.

Benchmark System

Safety-by-Design Visual De-identification Benchmark

Built a reproducible benchmark framework for evaluating privacy-preserving visual transformation under realistic adversarial and utility constraints.

Evaluates privacy leakage through face/plate re-identification and OCR text recovery.
Measures utility, perceptual quality, robustness, latency, and deployment trade-offs.
Designed for camera streams in smart sensing, transportation, and safety-critical environments.

Large-scale Modeling

Temporal Graph Modeling for Community Evolution

Constructed and analyzed large-scale dynamic transaction networks to quantify community evolution and decentralization patterns in blockchain games.

Processed 34M+ transactions and 1.3M+ nodes into monthly temporal graphs.
Applied Leiden community detection and overlap-based tracking for merge/split/birth/death patterns.
Produced CSCWD 2025 oral paper on governance tokens and community evolution.

Selected Publications

Publications connecting trustworthy AI, benchmark design, multimodal systems, and large-scale analysis.

Large Language Model Agents in Finance: A Survey Bridging Research, Practice, and Real-World Deployment Sanjian Zhang, et al. · Findings of EMNLP 2025

Paper →

Securing the Skies: A Comprehensive Survey on Anti-UAV Methods, Benchmarking, and Future Directions Co-first Author · CVPR Workshop 2025 · Best Paper Award

Paper →

Decentralization and Community Evolution in Blockchain Games: The Role of Governance Tokens Sanjian Zhang · CSCWD 2025 · Oral Presentation

IEEE Xplore →

Featured Research Project

Biomedical AI · Computational Hemodynamics · Digital Twin Systems

Physics-Informed Microvascular Digital Twin for Hemodynamic Analysis and Visualization Biomedical AI · Scientific Visualization · Three.js · WebGL

Developed an interactive microvascular digital twin platform for three-dimensional vascular network visualization, hemodynamic analysis, and red blood cell (RBC) flow simulation. The system integrates vascular topology reconstruction, real-time particle dynamics, hemodynamic parameter exploration, and scientific presentation tools to support biomedical research and computational vascular modeling.

Reconstructed complex microvascular networks from biomedical datasets and generated interactive 3D vascular models.
Developed real-time red blood cell (RBC) particle simulation with velocity-dependent motion and vessel-specific trajectory mapping.
Implemented bifurcation-aware RBC routing based on vascular connectivity and flow conservation principles to simulate physiologically realistic blood-flow distribution at vessel junctions.
Integrated vessel-level hemodynamic metrics including vessel diameter, pressure, flow velocity, volumetric flow rate, hematocrit, wall shear stress, and vascular resistance.
Developed interactive vessel inspection panels for real-time exploration of structural and hemodynamic properties.
Implemented camera trajectory recording, animation replay, and automated video export for reproducible scientific demonstrations and conference presentations.
Established visualization-driven validation workflows for identifying potential inconsistencies in computational hemodynamic simulations.

Biomedical AI Digital Twin Computational Hemodynamics Scientific Visualization Red Blood Cell Simulation Microvascular Modeling Three.js WebGL Python Vascular Network Analysis

Professional Experience

Production engineering experience supporting deployable, reliable, and measurable AI systems.

Feb 2022 – Aug 2024

Shandong Bee Intelligent Manufacturing Software Developer

Shipped microservice dashboards and exposed ML inference as REST APIs.
Scaled data pipelines to process 2.3 million rows; optimized PostgreSQL/MySQL queries to reduce hot-query P95 latency by 42%.
Improved deployment lead time by 46% by introducing Docker and GitHub Actions CI/CD.
Integrated multiple external data sources with idempotent loaders, retry mechanisms, and structured monitoring.

Aug 2021 – Feb 2022

Tencent (Shenzhen) Backend Developer

Designed search and pagination REST APIs for a knowledge system, reducing P95 latency by 41%.
Implemented JWT authentication, RBAC, and rate limiting for secure backend services.
Reduced post-release defects by 35% through unit and integration testing.
Instrumented structured logging to support offline evaluation of ML features.

Technical Skills

Scientific machine learning, geometric learning, and research engineering.

Machine Learning

PyTorchDeep LearningGraph Neural NetworksTransformersComputer VisionCUDA

Scientific ML / Research

Scientific Machine LearningGeometric Deep LearningInverse ProblemsSurrogate ModelingSpatiotemporal ModelingDigital Twins

Engineering

PythonC/C++LinuxDockerGitSQLFastAPI

Research Interests

Machine learning methods for complex physical systems and scientific discovery.

Scientific Machine Learning & Geometric Deep Learning

I study inverse problems, surrogate modeling, and physics-constrained learning, with an emphasis on graph neural networks that learn from unstructured 3D meshes and simulation data.

Inverse identification and parameter estimation.
Graph neural networks, 3D meshes, and structure-agnostic learning.
Reproducible ML systems that connect learning with physical simulation.

AI for Science & Biomedical Digital Twins

My current application focus is patient-specific cardiac mechanics, where learned inverse and forward models can make computational digital twins faster, more accessible, and more informative.

Learning complex physical systems from simulation and observational data.
Patient-specific cardiac mechanics and computational modeling.
Transferable scientific ML methods for broader physical systems.

Contact

Let's connect scientific machine learning with physical simulation and patient-specific digital twins.

Email GitHub Scholar LinkedIn