Selected work

Representative Papers

LithoFormer: A Robust Framework for Stratigraphic Inference via Transformers

Jul 24, 2026

This work addresses the limitations of conventional sliding-window approaches in capturing the global geological context of well-log data, which often leads to misaligned stratigraphic boundaries and violations of geological sequence. To overcome this, we propose LithoFormer—the first application of a full-sequence Seq2Seq Transformer architecture to stratigraphic inference. Our model integrates a channel-agnostic PatchTST backbone, rotary position embeddings (RoPE), and a decoupled multi-task head to jointly predict lithological zones and boundary probabilities. Furthermore, we incorporate geological prior constraints into the loss function to enforce stratigraphic consistency. Evaluated on three real-world datasets, LithoFormer reduces the median boundary error by 90%, entirely eliminates sequence violations, and cuts expert annotation effort by 80%, thereby substantially enhancing the reliability and scalability of large-scale subsurface modeling.

0 citationsRead paper

Vancomycert: A Certified Neuro-Symbolic Drug Delivery System (Case Study)

Jun 17, 2026

This work addresses the challenge of ensuring both adaptability and provable safety over an infinite time horizon for neural network controllers in clinical settings. The authors develop a clinically interpretable, simplified physiological model and train a neural network via supervised learning to emulate vancomycin dosing strategies. For the first time, they formally verify the controller’s infinite-time safety using the Rocq language and the Vehicle interactive theorem prover. The end-to-end proof guarantees that the controller’s outputs never exceed drug toxicity thresholds, thereby ensuring that blood concentration remains within the therapeutic window at all time steps. This approach effectively mitigates nephrotoxicity risk while accommodating personalized and diverse dosing regimens.

0 citationsRead paper

Grammar of the Wave: Towards Explainable Multivariate Time Series Event Detection via Neuro-Symbolic VLM Agents

Mar 11, 2026

This work addresses the challenge of accurately mapping natural language descriptions to temporal intervals in multivariate time series for semantic event detection under scarce annotation. To this end, the authors propose a knowledge-guided neuro-symbolic framework that introduces, for the first time, an Event Logic Tree (ELT) to structurally align linguistic semantics with temporal logic over time series. Integrating a vision-language model (VLM) agent, the framework iteratively instantiates signal visualization primitives under ELT constraints, enabling zero-shot and interpretable event detection. Evaluated on real-world data and a newly curated expert-annotated benchmark, the method substantially outperforms both supervised fine-tuning approaches and existing large language model (LLM)/VLM zero-shot baselines. Human evaluations further confirm its high detection accuracy and trustworthy interpretability.

0 citationsRead paper

On the workflow, opportunities and challenges of developing foundation model in geophysics

Apr 24, 2025

To address challenges in geophysical data—including heterogeneity, low signal-to-noise ratio, and physical inconsistency—this paper introduces the first end-to-end development framework for geophysical foundation models. Methodologically, it integrates physics-informed priors to establish a novel paradigm comprising differentiable physics-constrained embedding, few-shot transfer adaptation, and interpretability-enhanced training—unifying physics-informed neural networks (PINNs), contrastive learning, multi-scale time-frequency representation, and self-supervised pretraining, with support for Model-as-a-Service (MaaS) deployment. Contributions include: (1) systematic coverage of the full lifecycle—from data acquisition and physics-aware preprocessing to architecture design, constrained pretraining, and deployment optimization; (2) a 70% reduction in annotation dependency; and (3) over 35% improvement in physical consistency of inversion results. The framework provides a standardized, reproducible technical pathway for multimodal geophysical analysis, including seismic, electromagnetic, and gravity data.

0 citationsRead paper

Recent publications

Latest Papers

LithoFormer: A Robust Framework for Stratigraphic Inference via Transformers

Jul 24, 2026

This work addresses the limitations of conventional sliding-window approaches in capturing the global geological context of well-log data, which often leads to misaligned stratigraphic boundaries and violations of geological sequence. To overcome this, we propose LithoFormer—the first application of a full-sequence Seq2Seq Transformer architecture to stratigraphic inference. Our model integrates a channel-agnostic PatchTST backbone, rotary position embeddings (RoPE), and a decoupled multi-task head to jointly predict lithological zones and boundary probabilities. Furthermore, we incorporate geological prior constraints into the loss function to enforce stratigraphic consistency. Evaluated on three real-world datasets, LithoFormer reduces the median boundary error by 90%, entirely eliminates sequence violations, and cuts expert annotation effort by 80%, thereby substantially enhancing the reliability and scalability of large-scale subsurface modeling.

0 citationsRead paper

Vancomycert: A Certified Neuro-Symbolic Drug Delivery System (Case Study)

Jun 17, 2026

This work addresses the challenge of ensuring both adaptability and provable safety over an infinite time horizon for neural network controllers in clinical settings. The authors develop a clinically interpretable, simplified physiological model and train a neural network via supervised learning to emulate vancomycin dosing strategies. For the first time, they formally verify the controller’s infinite-time safety using the Rocq language and the Vehicle interactive theorem prover. The end-to-end proof guarantees that the controller’s outputs never exceed drug toxicity thresholds, thereby ensuring that blood concentration remains within the therapeutic window at all time steps. This approach effectively mitigates nephrotoxicity risk while accommodating personalized and diverse dosing regimens.

0 citationsRead paper

Grammar of the Wave: Towards Explainable Multivariate Time Series Event Detection via Neuro-Symbolic VLM Agents

Mar 11, 2026

This work addresses the challenge of accurately mapping natural language descriptions to temporal intervals in multivariate time series for semantic event detection under scarce annotation. To this end, the authors propose a knowledge-guided neuro-symbolic framework that introduces, for the first time, an Event Logic Tree (ELT) to structurally align linguistic semantics with temporal logic over time series. Integrating a vision-language model (VLM) agent, the framework iteratively instantiates signal visualization primitives under ELT constraints, enabling zero-shot and interpretable event detection. Evaluated on real-world data and a newly curated expert-annotated benchmark, the method substantially outperforms both supervised fine-tuning approaches and existing large language model (LLM)/VLM zero-shot baselines. Human evaluations further confirm its high detection accuracy and trustworthy interpretability.

0 citationsRead paper

On the workflow, opportunities and challenges of developing foundation model in geophysics

Apr 24, 2025

To address challenges in geophysical data—including heterogeneity, low signal-to-noise ratio, and physical inconsistency—this paper introduces the first end-to-end development framework for geophysical foundation models. Methodologically, it integrates physics-informed priors to establish a novel paradigm comprising differentiable physics-constrained embedding, few-shot transfer adaptation, and interpretability-enhanced training—unifying physics-informed neural networks (PINNs), contrastive learning, multi-scale time-frequency representation, and self-supervised pretraining, with support for Model-as-a-Service (MaaS) deployment. Contributions include: (1) systematic coverage of the full lifecycle—from data acquisition and physics-aware preprocessing to architecture design, constrained pretraining, and deployment optimization; (2) a 70% reduction in annotation dependency; and (3) over 35% improvement in physical consistency of inversion results. The framework provides a standardized, reproducible technical pathway for multimodal geophysical analysis, including seismic, electromagnetic, and gravity data.

0 citationsRead paper