Selected work

Representative Papers

AMORE: Adaptive Multi-Output Operator Network for Stiff Chemical Kinetics

Oct 14, 2025

To address the computational efficiency bottleneck in CFD simulations of reactive systems—such as combustion and hypersonic flows—caused by stiff chemical kinetics, this work proposes a multi-output adaptive neural operator framework. Methodologically, it integrates DeepONet and Fourier Neural Operator (FNO) architectures, employs an adaptive weighted loss function to differentially penalize variable- and sample-wise errors, enforces physical consistency via an invertible analytical mapping that rigorously satisfies mass-fraction conservation and unity-sum constraints, and adopts a two-stage training strategy with dimensionality-reduction transformations to enhance generalizability. Experiments on syngas (12-species) and GRI-Mech 3.0 (24-species) mechanisms demonstrate that the model accelerates thermochemical state evolution prediction by one to two orders of magnitude over conventional implicit ODE solvers, while maintaining high fidelity. The framework serves as a robust, physics-informed acceleration module for turbulent combustion CFD simulations.

0 citationsRead paper

Geometric Feature Prompting of Image Segmentation Models

May 27, 2025

Automated segmentation of plant roots in rhizotron/minirhizotron images remains challenging due to structural complexity, while manual annotation is time-consuming and subjective. Method: We propose Geomprompt—a differential-geometry-based, semantic-aware prompt point generation method. Leveraging image gradient ridge detection, it automatically identifies geometrically salient points along slender structures (e.g., roots), providing only 1–3 precisely co-localized prompts to SAM/SAM2 to focus on local ridge-like features. Contribution/Results: This work pioneers the integration of differential geometry into prompt engineering, enhancing both physical interpretability and model robustness. Evaluated on root imagery, Geomprompt achieves a Dice score of 92.4%, significantly outperforming dense manual annotations and random-point baselines. The method is end-to-end compatible with the SAM ecosystem, and the open-source toolkit geomprompt is publicly released.

0 citationsRead paper

Recent publications

Latest Papers

AMORE: Adaptive Multi-Output Operator Network for Stiff Chemical Kinetics

Oct 14, 2025

To address the computational efficiency bottleneck in CFD simulations of reactive systems—such as combustion and hypersonic flows—caused by stiff chemical kinetics, this work proposes a multi-output adaptive neural operator framework. Methodologically, it integrates DeepONet and Fourier Neural Operator (FNO) architectures, employs an adaptive weighted loss function to differentially penalize variable- and sample-wise errors, enforces physical consistency via an invertible analytical mapping that rigorously satisfies mass-fraction conservation and unity-sum constraints, and adopts a two-stage training strategy with dimensionality-reduction transformations to enhance generalizability. Experiments on syngas (12-species) and GRI-Mech 3.0 (24-species) mechanisms demonstrate that the model accelerates thermochemical state evolution prediction by one to two orders of magnitude over conventional implicit ODE solvers, while maintaining high fidelity. The framework serves as a robust, physics-informed acceleration module for turbulent combustion CFD simulations.

0 citationsRead paper

Geometric Feature Prompting of Image Segmentation Models

May 27, 2025

Automated segmentation of plant roots in rhizotron/minirhizotron images remains challenging due to structural complexity, while manual annotation is time-consuming and subjective. Method: We propose Geomprompt—a differential-geometry-based, semantic-aware prompt point generation method. Leveraging image gradient ridge detection, it automatically identifies geometrically salient points along slender structures (e.g., roots), providing only 1–3 precisely co-localized prompts to SAM/SAM2 to focus on local ridge-like features. Contribution/Results: This work pioneers the integration of differential geometry into prompt engineering, enhancing both physical interpretability and model robustness. Evaluated on root imagery, Geomprompt achieves a Dice score of 92.4%, significantly outperforming dense manual annotations and random-point baselines. The method is end-to-end compatible with the SAM ecosystem, and the open-source toolkit geomprompt is publicly released.

0 citationsRead paper