Selected work

Representative Papers

Computer-aided shape features extraction and regression models for predicting the ascending aortic aneurysm growth rate

May 01, 2023Comput. Biol. Medicine

Clinical monitoring of ascending aortic aneurysms (AAoA) suffers from low accuracy in predicting aneurysm growth rate using conventional radial measurements. Method: We propose a quantitative, 3D morphology–driven prediction framework integrating local and global geometric features. Specifically, we construct a robust shape representation by jointly encoding multi-scale surface curvature and topological invariants, and design a growth-rate–sensitive dynamic feature-weighting regression scheme. Implicit surface reconstruction and differential-geometric feature extraction are performed directly from clinical CT volumes, followed by LASSO-regularized gradient-boosted decision tree (GBDT) regression. Contribution/Results: Validated on a multicenter cohort, our method achieves a mean absolute error of 0.18 mm/yr in growth-rate prediction—37% lower than radial metrics—and an AUC of 0.89 for binary classification of fast versus slow growth. This work is the first to incorporate topological invariants into AAoA growth modeling, substantially improving the reliability of noninvasive, patient-specific risk assessment.

13 citationsRead paper

Bayesian Optimization of Crossbar-Based Compute-In-Memory System Design for Efficient DNN Inference

May 08, 2026

This work addresses the challenge of efficiently optimizing deep neural network inference on compute-in-memory (CIM) crossbar arrays, where the high-dimensional, non-convex design space—shaped by model complexity and heterogeneous layer workloads—hinders effective exploration. To tackle this, we introduce, for the first time, a multi-objective Bayesian optimization framework for system-level co-design of CIM architectures, jointly tuning hardware configurations and per-layer network parameters. Our approach efficiently navigates a design space of up to 50 dimensions and ~10²⁷ possible configurations to identify Pareto-optimal trade-offs among accuracy, energy efficiency, and area. Integrating high-dimensional modeling, layer-granular resource allocation, and a CIM-aware simulator, the method achieves 91.72% accuracy on VGG8/CIFAR-10 and 57.2% on VGG16/Tiny-ImageNet-200, while reducing chip area by up to 65.52%, dynamic energy by 52.07%, and read latency by 13.27%, alongside significantly improved memory utilization.

0 citationsRead paper

Exposing LLM Safety Gaps Through Mathematical Encoding:New Attacks and Systematic Analysis

May 05, 2026

This work proposes a novel attack paradigm that systematically bypasses mainstream safety filters in large language models (LLMs) by recasting harmful requests as formal mathematical problems rooted in set theory, formal logic, and quantum mechanics. Current safety mechanisms, which rely primarily on semantic pattern matching, fail to recognize such mathematically encoded adversarial content. The approach leverages an auxiliary LLM to translate harmful semantics into rigorous formal expressions, demonstrating broad applicability and robustness across diverse mathematical frameworks. Experimental evaluation across eight target models and two benchmarks reveals average attack success rates of 46%–56%, with even advanced models like GPT-5 exhibiting significant vulnerabilities. These findings expose a fundamental gap in existing safety architectures: their inability to comprehend deep mathematical semantics, thereby leaving them susceptible to formally disguised harmful queries.

0 citationsRead paper

A Curated Literature Database for Monitoring More Than 30 Years of Ansys Granta Product Usage

Feb 20, 2026

This work addresses the challenge of quantifying the academic impact of commercial engineering software such as Ansys Granta, which is hindered by inconsistent citation practices and rapidly growing publication volumes. We propose the first reproducible, semi-automated framework that integrates DOI and citation parsing, expert annotation, and a relational database (Ansys Granta MI Enterprise) to transform heterogeneous usage evidence into a structured knowledge base. As of September 2025, the framework has compiled a multi-source literature repository comprising over 1,100 manually verified records, enabling rapid retrieval, systematic review reproduction, and technology landscape scanning. The resulting knowledge base reveals dominant application domains, key contributing institutions, and integration patterns within CAD/CAE/FEM environments, thereby facilitating systematic tracking and analysis of the long-term technical influence of commercial engineering software.

0 citationsRead paper

Hybrid Iterative Solvers with Geometry-Aware Neural Preconditioners for Parametric PDEs

Dec 16, 2025

To address the slow convergence and poor generalization of iterative solvers for parametric partial differential equations (PDEs) on arbitrary unstructured meshes, this paper proposes a geometry-aware hybrid preconditioning framework. The method integrates finite-element mesh encoding with a novel Geo-DeepONet architecture—yielding the first neural preconditioner capable of cross-geometry generalization without retraining. Coupled with Krylov subspace methods (e.g., GMRES) and multilevel relaxation strategies, it enables geometry-adaptive iterative acceleration. Evaluated on parametric PDEs from elasticity and heat conduction, the framework reduces generalization error on unseen geometries by over 60%, accelerates overall solution time by 3–5×, and significantly improves robustness and transferability across diverse geometric domains.

0 citationsRead paper

Recent publications

Latest Papers

Bayesian Optimization of Crossbar-Based Compute-In-Memory System Design for Efficient DNN Inference

May 08, 2026

This work addresses the challenge of efficiently optimizing deep neural network inference on compute-in-memory (CIM) crossbar arrays, where the high-dimensional, non-convex design space—shaped by model complexity and heterogeneous layer workloads—hinders effective exploration. To tackle this, we introduce, for the first time, a multi-objective Bayesian optimization framework for system-level co-design of CIM architectures, jointly tuning hardware configurations and per-layer network parameters. Our approach efficiently navigates a design space of up to 50 dimensions and ~10²⁷ possible configurations to identify Pareto-optimal trade-offs among accuracy, energy efficiency, and area. Integrating high-dimensional modeling, layer-granular resource allocation, and a CIM-aware simulator, the method achieves 91.72% accuracy on VGG8/CIFAR-10 and 57.2% on VGG16/Tiny-ImageNet-200, while reducing chip area by up to 65.52%, dynamic energy by 52.07%, and read latency by 13.27%, alongside significantly improved memory utilization.

0 citationsRead paper

Exposing LLM Safety Gaps Through Mathematical Encoding:New Attacks and Systematic Analysis

May 05, 2026

This work proposes a novel attack paradigm that systematically bypasses mainstream safety filters in large language models (LLMs) by recasting harmful requests as formal mathematical problems rooted in set theory, formal logic, and quantum mechanics. Current safety mechanisms, which rely primarily on semantic pattern matching, fail to recognize such mathematically encoded adversarial content. The approach leverages an auxiliary LLM to translate harmful semantics into rigorous formal expressions, demonstrating broad applicability and robustness across diverse mathematical frameworks. Experimental evaluation across eight target models and two benchmarks reveals average attack success rates of 46%–56%, with even advanced models like GPT-5 exhibiting significant vulnerabilities. These findings expose a fundamental gap in existing safety architectures: their inability to comprehend deep mathematical semantics, thereby leaving them susceptible to formally disguised harmful queries.

0 citationsRead paper

A Curated Literature Database for Monitoring More Than 30 Years of Ansys Granta Product Usage

Feb 20, 2026

This work addresses the challenge of quantifying the academic impact of commercial engineering software such as Ansys Granta, which is hindered by inconsistent citation practices and rapidly growing publication volumes. We propose the first reproducible, semi-automated framework that integrates DOI and citation parsing, expert annotation, and a relational database (Ansys Granta MI Enterprise) to transform heterogeneous usage evidence into a structured knowledge base. As of September 2025, the framework has compiled a multi-source literature repository comprising over 1,100 manually verified records, enabling rapid retrieval, systematic review reproduction, and technology landscape scanning. The resulting knowledge base reveals dominant application domains, key contributing institutions, and integration patterns within CAD/CAE/FEM environments, thereby facilitating systematic tracking and analysis of the long-term technical influence of commercial engineering software.

0 citationsRead paper

Hybrid Iterative Solvers with Geometry-Aware Neural Preconditioners for Parametric PDEs

Dec 16, 2025

To address the slow convergence and poor generalization of iterative solvers for parametric partial differential equations (PDEs) on arbitrary unstructured meshes, this paper proposes a geometry-aware hybrid preconditioning framework. The method integrates finite-element mesh encoding with a novel Geo-DeepONet architecture—yielding the first neural preconditioner capable of cross-geometry generalization without retraining. Coupled with Krylov subspace methods (e.g., GMRES) and multilevel relaxation strategies, it enables geometry-adaptive iterative acceleration. Evaluated on parametric PDEs from elasticity and heat conduction, the framework reduces generalization error on unseen geometries by over 60%, accelerates overall solution time by 3–5×, and significantly improves robustness and transferability across diverse geometric domains.

0 citationsRead paper

Crystallographic Texture-Generalizable Orientation-Aware Interaction-Based Deep Material Network for Polycrystal Modeling and Texture Evolution

Dec 07, 2025

Existing ODMN frameworks require separate training for each crystal texture, exhibiting poor generalizability. To address this, we propose TACS-GNN-ODMN—a novel framework enabling generalized prediction of polycrystalline mechanical response and texture evolution across diverse microstructures without retraining. Our method introduces Texture-Adaptive Clustering Sampling (TACS) to initialize model parameters, integrates a Graph Neural Network (GNN) to capture orientation-dependent intergranular interactions, and preserves the physics-driven architecture of ODMN. The resulting framework ensures both physical interpretability and cross-texture transferability. Quantitative evaluation demonstrates prediction accuracy comparable to direct numerical simulation across multiple textures, while achieving substantial computational speedup. This enables efficient multiscale simulation and accelerated design of novel metallic materials.

0 citationsRead paper