Computational Physics

Shi Qiu, Junyi Deng, Yiwei Deng et al.

This benchmark provides a rigorous, expert-validated test for evaluating AI agents' capability to autonomously reproduce scientific research, revealing systematic failures that must be addressed for progress in AI-driven science.

Bowen Deng, Bohan Li, Matthew Cox et al.

For computational materials science and chemistry researchers, AtomisticSkills provides a modular infrastructure to automate complex atomistic workflows, addressing the challenge of scaling monolithic agents in fragmented software ecosystems.

Yadi Cao, Sicheng Lai, Jiahe Huang et al.

This addresses the gap in cost-aware evaluation for physics simulations, providing a practical benchmark for researchers and practitioners, though it is incremental in extending existing benchmarking approaches.

Zijian Zhang, Aiwei Yin, Amaan Baweja et al.

This work addresses the bottleneck of static, hand-curated toolsets in scientific agentic systems, enabling autonomous tool creation and reuse for quantum chemistry and dynamics tasks.

Nithin Somasekharan, Youssef Hassan, Shiyao Lin et al.

For researchers building scientific AI assistants, this benchmark addresses the overlooked problem of refining ill-posed user requests through dialogue, but the results show current models are far from reliable.

Yuanzhao Zhang, William Gilpin

For researchers developing time-series foundation models, this work reveals a critical performance gap and failure mode, providing a simple baseline that should be used to validate whether models are truly learning beyond trivial copying.

Xingyang Yu, Yinghuan Zhang, Yufei Zhang et al.

It provides a sensitive evaluation lens for epistemic limits of LLMs in highly abstract theoretical physics, but the dataset is small and the findings are incremental.

Yeping Hu, Ruben Glatt, Shusen Liu

For users of graph-based CFD surrogates in digital twins and closed-loop control, this work provides a method to correct phase drift post hoc, extending latent-space steering to time-dependent physical systems.

Shyam Sankaran, Hanwen Wang, Paris Perdikaris

For researchers building neural PDE solvers, this shows that small models with strong priors can rival huge foundation models, challenging the scaling trend in this domain.

Luca Maria Del Bono, Giulio Biroli, Patrick Charbonneau et al.

Provides theoretical insight into the limitations of diffusion models near criticality and demonstrates how architectural design can overcome these bottlenecks, relevant for statistical physics and generative modeling.

Yi-Lun Liao, Alexander J. Hoffman, Sabrina C. Shen et al.

This work addresses the need for scalable and physically consistent models in computational chemistry and materials science, representing an incremental advancement over previous versions.

Arthur Kosmala, Stephan Günnemann, Meng Gao et al.

For computational chemists and physicists, LSD addresses the serial bottleneck in molecular dynamics to increase single-system throughput without introducing error.

Dong-Xiao Zhang, Hu Lou, Jun-Jie Zhang et al.

It addresses reliability issues in AI systems by providing a unified framework for diagnosing and mitigating anomalies across perception and generation tasks, though it is incremental in applying existing uncertainty concepts to new contexts.

Hadi Sotoudeh, Payel Mukhopadhyay, Ruben Ohana et al.

This provides practical guidance for training efficient physics emulators in data-limited settings, though it is incremental as it focuses on optimizing an existing approach.

Chaoqian Yuan, Chinnapat Panwisawas, Ye Lu

For researchers simulating microstructure evolution in additive manufacturing, this framework offers a way to drastically reduce computational cost without sacrificing stability.

Nicholas Gao, Till Grutschus, Frank Noé et al.

This work addresses the scalability challenge for quantum chemistry simulations, offering a more efficient method for modeling excited states across molecules, though it is incremental in improving existing neural-network wave function approaches.

Paul Garnier, Vincent Lannelongue, Elie Hachem

For researchers in geometric deep learning and computational physics, this work addresses a critical bottleneck in training physics surrogates by incorporating numerical analysis principles, leading to more accurate and stable simulations.

Yuhao Zhang, Ti John, Matthias Stosiek et al.

This work provides a practical BO framework for mixed-variable optimization problems in the natural sciences, particularly useful in autonomous laboratory settings with noise and limited data, but it is incremental as it builds on existing PR methods.

Aleksander Krasowski, René P. Klausen, Aycan Celik et al.

This addresses the trust issue for researchers and practitioners using PINNs in scientific computing by providing interpretable error explanations, though it is incremental as it builds on existing PINN frameworks.

Jiajun He, Zijing Ou, Francisco Vargas et al.

This work provides a more general and efficient method for free energy estimation, which is a fundamental problem across physics and statistics.