arXiv subjects

Heng Ge

Publications and source records attributed to Heng Ge.

4 recordsLinked to original sources

arXiv2026

High-Pressure Crystal Structure Database

High-pressure research is a productive route to new structures and emergent properties. However, crucial high-pressure structural information remains highly fragmented across individual publications and heterogeneous computational repositories. This fragmentation creates a major bottleneck for data-driven materials design. To bridge this gap, we introduce the High-Pressure Crystal Structure Database (HPCSD), a traceable, pressure-resolved repository that integrates experimental and theoretical high-pressure structures. HPCSD is constructed from two complementary data streams: elemental high-pressure phases and a searchable configuration space of stable and metastable phases generated via CALYPSO crystal structure prediction. To ensure rigorous comparability, all retained structures underwent re-optimization under a unified density functional theory (DFT) framework , with continuous enthalpy curves systematically generated specifically for the elemental phases across their stability fields. The initial release encompasses 77,346 consistently evaluated structural entries spanning 89 elements. An analysis reveals that pressure-induced polymorphism is ubiquitous and exhibits pronounced family-dependent trends. Structural diversity is strongly influenced by an element's electronic adaptability , with the greatest structural complexity emerging at intermediate rather than highest pressures. By providing standardized, reusable, and rigorously evaluated high-pressure structural data, HPCSD establishes a robust infrastructure to accelerate experimental phase identification, facilitate cross-study thermodynamic comparisons, and support the development of machine-learning interatomic potentials and generative models for high-pressure systems.

cond-mat.mtrl-sci↗

arXiv2025

DiffKnock: Diffusion-based Knockoff Statistics for Neural Networks Inference

We introduce DiffKnock, a diffusion-based knockoff framework for high-dimensional feature selection with finite-sample false discovery rate (FDR) control. DiffKnock addresses two key limitations of existing knockoff methods: preserving complex feature dependencies and detecting non-linear associations. Our approach trains diffusion models to generate valid knockoffs and uses neural network--based gradient and filter statistics to construct antisymmetric feature importance measures. Through simulations, we showed that DiffKnock achieved higher power than autoencoder-based knockoffs while maintaining target FDR, indicating its superior performance in scenarios involving complex non-linear architectures. Applied to murine single-cell RNA-seq data of LPS-stimulated macrophages, DiffKnock identifies canonical NF-$\kappa$B target genes (Ccl3, Hmox1) and regulators (Fosb, Pdgfb). These results highlight that, by combining the flexibility of deep generative models with rigorous statistical guarantees, DiffKnock is a powerful and reliable tool for analyzing single-cell RNA-seq data, as well as high-dimensional and structured data in other domains.

stat.ME↗

arXiv2025

Neural Tangent Kernels for Complex Genetic Risk Prediction: Bridging Deep Learning and Kernel Methods in Genomics

Given the complexity of genetic risk prediction, there is a critical need for the development of novel methodologies that can effectively capture intricate genotype--phenotype relationships (e.g., nonlinear) while remaining statistically interpretable and computationally tractable. We develop a Neural Tangent Kernel (NTK) framework to integrate kernel methods into deep neural networks for genetic risk prediction analysis. We consider two approaches: NTK-LMM, which embeds the empirical NTK in a linear mixed model with variance components estimated via minimum quadratic unbiased estimator (MINQUE), and NTK-KRR, which performs kernel ridge regression with cross-validated regularization. Through simulation studies, we show that NTK-based models outperform the traditional neural network models and linear mixed models. By applying NTK to endophenotypes (e.g., hippocampal volume) and AD-related genes (e.g., APOE) from Alzheimer's Disease Neuroimaging Initiative (ADNI), we found that NTK achieved higher accuracy than existing methods for hippocampal volume and entorhinal cortex thickness. In addition to its accuracy performance, NTK has favorable optimization properties (i.e., having a closed-form or convex training) and generates interpretable results due to its connection to variance components and heritability. Overall, our results indicate that by integrating the strengths of both deep neural networks and kernel methods, NTK offers competitive performance for genetic risk prediction analysis while having the advantages of interpretability and computational efficiency.

stat.AP↗

arXiv2019

Infinite-horizon Off-Policy Policy Evaluation with Multiple Behavior Policies

We consider off-policy policy evaluation when the trajectory data are generated by multiple behavior policies. Recent work has shown the key role played by the state or state-action stationary distribution corrections in the infinite horizon context for off-policy policy evaluation. We propose estimated mixture policy (EMP), a novel class of partially policy-agnostic methods to accurately estimate those quantities. With careful analysis, we show that EMP gives rise to estimates with reduced variance for estimating the state stationary distribution correction while it also offers a useful induction bias for estimating the state-action stationary distribution correction. In extensive experiments with both continuous and discrete environments, we demonstrate that our algorithm offers significantly improved accuracy compared to the state-of-the-art methods.

cs.LG↗