Jiapeng Chen AlexanderChan

Structural Engineering · Master's student

Chien-Shiung Wu College · Undergraduate studies

AI for CAE Intern

Worked on an AI for CAE system for smartphone impact response prediction, spanning finite-element data modeling, graph-network architecture, and training validation, with a focus on temporal surrogates for complex contact scenarios.

• Graph-based finite-element surrogatesBuilt a GraphGPS-based temporal model with Patch + Core/Halo partitioning. Unified node states, element topology, contact relations, and tie constraints in a graph representation to predict next-step displacement, equivalent plastic strain (PEEQ), and contact response.
• Physics-aware multi-relation architectureDesigned parallel branches for Node–Element–Node bipartite message passing, Tie Event MPNN, Contact Event MPNN, and global attention. Used gated residual fusion to combine local physical interactions with long-range dependencies.
• Causal history memoryCompressed historical finite-element frames into Contact / Part / Global tokens. Combined age encoding, chunked multi-head attention, FIFO memory, and gated fusion to use multi-frame history while controlling memory cost.
• Training engineering and validationImplemented multi-GPU training with core-only loss, field masks, and backpropagation through time (BPTT) for patch-level spatiotemporal learning. Evaluated convergence, generalization across phone models, and inference acceleration.

Connector mechanics simulation · Project lead

Lead connector mechanics simulation projects within a university–industry research collaboration. Study mating reliability under manufacturing tolerance combinations, combining finite-element automation, surrogate modeling, and model reduction with physical test comparisons.

• Tolerance modeling and surrogatesStructure geometry and mesh information for multi-part assemblies under different tolerance combinations. Build and train connector surrogate models to reduce the cost of repeated finite-element analyses through data-driven prediction.
• Simulation automationAutomate HyperMesh and Abaqus through Tcl and abaqpy, connecting Monte Carlo tolerance sampling, mesh checks, and batch job submission in a reusable process for repeated modeling and analysis tasks.
• Substructuring and multiple load casesDevelop and implement a substructuring-based reduction strategy for single-pair insertion and extraction, and coupled analysis of multiple connector pairs. Cover aligned, offset, and angularly misaligned conditions to reduce computational cost for complex assemblies.
• Physical test comparisonCoordinate HyperMesh–Abaqus simulation and independently conduct insertion and extraction tests under multiple conditions. Clean and compare experimental and simulated data to assess the engineering applicability of the reduction strategy and surrogate models.

Intern · Industry research & technology commercialization

Conducted deep-tech industry research and business interviews across simulation, quantum computing, large language models, and biomedicine. Contributed to an evaluation system for the commercialization potential of research technologies.

• Technology commercialization researchInvestigated technical approaches and applications in Physics AI and AI-accelerated industrial simulation. Combined industry research and interviews to understand deployment scenarios and support the assessment of early-stage technology ventures.
• Data preparation for technology evaluationHelped build a research commercialization evaluation system from the ground up. Organized and cleaned data from YC Companies, Google Scholar, and PubMed to support downstream analysis and modeling.
• Machine learning and technology-factor analysisTrained machine-learning models using industry research data and contributed to technology-asset factor analysis, connecting technical research with data-driven commercialization assessment.

Project lead

Led a warehouse slot-allocation project that incorporated structural stability into AGV decision-making, spanning reinforcement-learning environment design, policy training, and tests in a physical warehouse.

• Reinforcement-learning environmentIndependently built a Gymnasium environment that translated warehouse slot allocation and structural stability requirements into a reinforcement-learning task, enabling agents to learn warehouse policies in simulation.
• Coupled simulation and structural feedbackConnected CoppeliaSim, Isaac Sim, and SAP2000 in a simulation process. Extended SAP2000 through Python to automatically evaluate structural stability and safety indicators for different slot allocations and provide feedback to the policy.
• Policy training and physical testingTrained multi-agent policies with PPO and MAPPO in the custom environment, then deployed the models in a physical warehouse and completed on-site operational tests to assess transfer from simulation to practice.