Stencil Computation at the Intersection of AI and HPC
研究使用TinyTC和PyTorch/Triton在GPGPU上实现高效高阶模板计算，解决科学与工业应用中的性能瓶颈，通过硬件优化提升计算速度。
0 citationsRead paper
Selected work
Representative Papers
Contrastive Learning for Seismic Horizon Tracking with Domain-Specific Priors
Unsupervised 3D seismic horizon tracking often fails near faults: signal-driven methods offer high precision but poor robustness, while texture-driven approaches handle discontinuities yet rely on labels and suffer from limited local accuracy. This work proposes a self-supervised contrastive learning framework that integrates both signal and texture cues, introducing for the first time inter-trace flow derived from reflection dip estimation as a domain-specific prior. Positive sample pairs are constructed within high-confidence neighborhoods to propagate horizon identity consistently across faults. The method trains a texture-aware voxel embedding model by combining high-confidence region constraints with an optional fault mask. Evaluated on the public F3 dataset and synthetic data with faults, the approach achieves a mean absolute error (MAE) significantly better than unsupervised baselines and comparable to semi-supervised methods using only a single annotated slice.
0 citationsRead paper
Towards Fast GNN Surrogates for CO2 Migration in Complex Geological Formations
This study addresses the challenge of predicting CO₂ plume migration in complex geological formations by proposing an end-to-end graph neural network surrogate model. The approach represents geological grids as graph structures enriched with geometric attributes and introduces a novel geometrically conditioned edge embedding to drive anisotropic message passing, effectively capturing directional transport behaviors induced by grid topology, permeability contrasts, and geological heterogeneity. By integrating autoregressive residual modeling with multi-step supervised training, the model achieves high-fidelity spatiotemporal simulation of multiphase flow in porous media. Evaluated on the SPE11A benchmark, the method demonstrates significantly lower cumulative errors in long-term predictions of gas-phase saturation and liquid-phase density compared to existing approaches, highlighting its superior generalization capability and numerical stability.
0 citationsRead paper
Random Process Flow Matching: Generative Implicit Representations of Multivariate Random Fields
This work addresses the challenge of generating high-fidelity realizations and providing reliable uncertainty quantification for high-dimensional, high-frequency, or multi-dimensional random fields from only a single sparse observational sample. The authors propose RP Flow, a novel framework that uniquely integrates flow matching with Gaussian process posteriors. By leveraging neural implicit representations and random Fourier features, RP Flow learns a continuous field from sparse data that can be queried at arbitrary locations. Calibrated uncertainty estimates are achieved through ensemble sampling. This approach overcomes the conventional reliance of generative models on large-scale datasets, enabling accurate synthesis and traceable uncertainty quantification even under extreme sparsity, high frequencies, or high-dimensional settings.
0 citationsRead paper
GICC: A High-Performance Runtime for GPU-Initiated Communication and Coordination in Modern HPC Systems
This work addresses the inability of GPUs in current high-performance computing (HPC) systems to autonomously initiate cross-node communication, particularly the lack of an efficient, low-overhead GPU-driven communication mechanism on OFI-based interconnects such as Slingshot, alongside inefficient NIC resource reclamation. The authors propose GICC, a runtime system that, for the first time on OFI architectures, enables GPUs to directly trigger NIC operations without host intervention, facilitating fine-grained overlap of computation and communication. GICC also introduces an asynchronous, lock-free, lightweight resource reclamation mechanism. Experimental results demonstrate a 229× reduction in coordination latency and a 25% improvement in weak scaling efficiency on Slingshot; on InfiniBand, it achieves 1.95× lower Put latency compared to NVSHMEM. In an industrial-scale stencil application, GICC attains 42% parallel efficiency, significantly outperforming MPI’s 35.4%.
0 citationsRead paper
Recent publications
Latest Papers
Stencil Computation at the Intersection of AI and HPC
研究使用TinyTC和PyTorch/Triton在GPGPU上实现高效高阶模板计算，解决科学与工业应用中的性能瓶颈，通过硬件优化提升计算速度。
0 citationsRead paper
Contrastive Learning for Seismic Horizon Tracking with Domain-Specific Priors
Unsupervised 3D seismic horizon tracking often fails near faults: signal-driven methods offer high precision but poor robustness, while texture-driven approaches handle discontinuities yet rely on labels and suffer from limited local accuracy. This work proposes a self-supervised contrastive learning framework that integrates both signal and texture cues, introducing for the first time inter-trace flow derived from reflection dip estimation as a domain-specific prior. Positive sample pairs are constructed within high-confidence neighborhoods to propagate horizon identity consistently across faults. The method trains a texture-aware voxel embedding model by combining high-confidence region constraints with an optional fault mask. Evaluated on the public F3 dataset and synthetic data with faults, the approach achieves a mean absolute error (MAE) significantly better than unsupervised baselines and comparable to semi-supervised methods using only a single annotated slice.
0 citationsRead paper
Towards Fast GNN Surrogates for CO2 Migration in Complex Geological Formations
This study addresses the challenge of predicting CO₂ plume migration in complex geological formations by proposing an end-to-end graph neural network surrogate model. The approach represents geological grids as graph structures enriched with geometric attributes and introduces a novel geometrically conditioned edge embedding to drive anisotropic message passing, effectively capturing directional transport behaviors induced by grid topology, permeability contrasts, and geological heterogeneity. By integrating autoregressive residual modeling with multi-step supervised training, the model achieves high-fidelity spatiotemporal simulation of multiphase flow in porous media. Evaluated on the SPE11A benchmark, the method demonstrates significantly lower cumulative errors in long-term predictions of gas-phase saturation and liquid-phase density compared to existing approaches, highlighting its superior generalization capability and numerical stability.
0 citationsRead paper
Random Process Flow Matching: Generative Implicit Representations of Multivariate Random Fields
This work addresses the challenge of generating high-fidelity realizations and providing reliable uncertainty quantification for high-dimensional, high-frequency, or multi-dimensional random fields from only a single sparse observational sample. The authors propose RP Flow, a novel framework that uniquely integrates flow matching with Gaussian process posteriors. By leveraging neural implicit representations and random Fourier features, RP Flow learns a continuous field from sparse data that can be queried at arbitrary locations. Calibrated uncertainty estimates are achieved through ensemble sampling. This approach overcomes the conventional reliance of generative models on large-scale datasets, enabling accurate synthesis and traceable uncertainty quantification even under extreme sparsity, high frequencies, or high-dimensional settings.
0 citationsRead paper
GICC: A High-Performance Runtime for GPU-Initiated Communication and Coordination in Modern HPC Systems
This work addresses the inability of GPUs in current high-performance computing (HPC) systems to autonomously initiate cross-node communication, particularly the lack of an efficient, low-overhead GPU-driven communication mechanism on OFI-based interconnects such as Slingshot, alongside inefficient NIC resource reclamation. The authors propose GICC, a runtime system that, for the first time on OFI architectures, enables GPUs to directly trigger NIC operations without host intervention, facilitating fine-grained overlap of computation and communication. GICC also introduces an asynchronous, lock-free, lightweight resource reclamation mechanism. Experimental results demonstrate a 229× reduction in coordination latency and a 25% improvement in weak scaling efficiency on Slingshot; on InfiniBand, it achieves 1.95× lower Put latency compared to NVSHMEM. In an industrial-scale stencil application, GICC attains 42% parallel efficiency, significantly outperforming MPI’s 35.4%.
0 citationsRead paper