⚛️ Physics & Scientific Computing¶

🤖 AAAI2026 · 15 paper notes

📌 Same area in other venues: 📷 CVPR2026 (2) · 🔬 ICLR2026 (69) · 🧪 ICML2026 (33) · 🧠 NeurIPS2025 (57) · 📹 ICCV2025 (2)

Adaptive Fidelity Estimation for Quantum Programs with Graph-Guided Noise Awareness: This paper proposes QuFid, a framework that models quantum circuits as directed acyclic graphs (DAGs), characterizes noise propagation via control-flow-aware random walks, quantifies circuit complexity through spectral features of the propagation operator, and achieves adaptive measurement budget allocation — significantly reducing the number of measurement shots while maintaining fidelity accuracy.
Catastrophic Forgetting in Kolmogorov-Arnold Networks: The first systematic study of catastrophic forgetting in Kolmogorov-Arnold Networks (KANs): establishes a theoretical framework linking forgetting to activation support overlap and intrinsic data dimensionality, and proposes KAN-LoRA for continual fine-tuning knowledge editing in language models.
Data Verification is the Future of Quantum Computing Copilots: This position paper argues that data verification must be elevated from a post-hoc filtering step to a foundational architectural principle in quantum computing AI copilots. Three positions are advanced: (1) verified data is a minimum requirement; (2) prior constraints outperform posterior filtering; (3) scientific domains governed by physical laws require verification-aware architectures. Experiments demonstrate that LLMs trained without verified data achieve at most 79% accuracy on circuit optimization tasks.
Fast 3D Surrogate Modeling for Data Center Thermal Management: This paper develops a vision-based 3D surrogate modeling framework for data centers. Server workloads, fan speeds, and air-conditioning temperature setpoints are encoded as 3D voxel representations, and architectures including 3D CNN U-Net, 3D Fourier Neural Operator, and 3D Vision Transformer are employed for real-time temperature field prediction. The proposed framework achieves inference speeds up to 20,000× faster than traditional CFD solvers while enabling a 7% reduction in energy consumption.
FlashKAT: Understanding and Addressing Performance Bottlenecks in the Kolmogorov-Arnold Transformer: This paper provides an in-depth analysis of the root cause behind KAT (Kolmogorov-Arnold Transformer) training being 123× slower than ViT. The bottleneck is identified not as FLOPs but as memory stalls caused by gradient accumulation during backpropagation (global memory contention from atomic add operations). The proposed FlashKAT restructures GPU kernels to achieve an 86.5× training speedup and reduces gradient rounding errors by nearly an order of magnitude.
Just Few States are Enough: Randomized Sparse Feedback for Stability of Dynamical Systems: This paper proposes a randomized sparse feedback control framework in which the controller accesses only a random subset of the state vector at each time step. Feedback gain matrices and Bernoulli sparsification parameters are jointly designed via LMIs to guarantee asymptotic mean-square stability (AMSS) while minimizing the required number of active sensors. Experiments demonstrate that as few as 0.3% of state components suffice to achieve performance comparable to full-state feedback.
Knowledge-Guided Masked Autoencoder with Linear Spectral Mixing and Spectral-Angle-Aware Reconstruction: This paper proposes KARMA, a framework that embeds the Linear Spectral Mixing Model (LSMM) as a physics constraint within the ViT-MAE decoder, combined with a Spectral Angle Mapper (SAM) loss, to improve reconstruction fidelity and downstream transfer performance for hyperspectral remote sensing imagery.
Learning Fair Representations with Kolmogorov-Arnold Networks: This paper proposes integrating Kolmogorov-Arnold Networks (KAN) into an adversarial debiasing framework, leveraging KAN's spline-based architecture to provide theoretical guarantees of Lipschitz continuity and smoothness. An adaptive \(\lambda\) update mechanism is introduced to dynamically balance fairness and accuracy. The approach achieves significant improvements on fairness metrics on the UCI college admissions dataset.
Phys-Liquid: A Physics-Informed Dataset for Estimating 3D Geometry and Volume of Transparent Deformable Liquids: This work introduces the Phys-Liquid dataset (97,200 physics-simulated images with 3D meshes), which models dynamic deformation of liquids inside transparent containers based on the Navier-Stokes equations, and proposes a four-stage reconstruction pipeline (segmentation → multi-view mask generation → 3D reconstruction → scaling) to achieve high-accuracy liquid geometry and volume estimation in both simulated and real-world scenes.
PhysicsCorrect: A Training-Free Approach for Stable Neural PDE Simulations: This paper proposes PhysicsCorrect, a training-free correction framework that models PDE residual correction as a linearized inverse problem and precomputes a cached pseudoinverse. At inference time, it achieves up to 100× error reduction with less than 5% computational overhead, and is applicable to arbitrary pretrained neural operators including FNO, UNet, and ViT.
PIMRL: Physics-Informed Multi-Scale Recurrent Learning for Burst-Sampled Spatiotemporal Dynamics: This paper proposes PIMRL, a framework for learning from burst-sampled (short high-frequency bursts followed by long intervals) sparse spatiotemporal data. It features a dual-module architecture combining macro-scale latent-space reasoning and micro-scale physics correction, integrated via cross-scale message passing, achieving up to 80% error reduction across 5 PDE benchmarks.
SAOT: An Enhanced Locality-Aware Spectral Transformer for Solving PDEs: This paper proposes SAOT (Spectral Attention Operator Transformer), which captures high-frequency local details via linear-complexity Wavelet Attention (WA) and complements it with the global receptive field of Fourier Attention (FA) through a gated fusion mechanism. SAOT achieves state-of-the-art performance on 6 operator learning benchmarks, reducing the relative error on Navier-Stokes by 22.3% compared to Transolver.
Scientific Knowledge-Guided Machine Learning for Vessel Power Prediction: A Comparative Study: A hybrid modeling framework combining a physics baseline with a data-driven residual is proposed. The sea trial power curve (propeller law \(P=cV^n\)) serves as the baseline, and XGBoost/NN/PINN models learn the residual correction, significantly improving extrapolation stability and physical consistency in sparse data regions.
SVD-NO: Learning PDE Solution Operators with SVD Integral Kernels: This paper proposes SVD-NO, a neural operator that explicitly parameterizes the SVD decomposition of integral kernels, achieving \(O(ndL)\) linear computational complexity while maintaining high expressiveness, and attaining new state-of-the-art performance on 5 PDE benchmarks.
Towards a Foundation Model for Partial Differential Equations Across Physics Domains: This paper proposes PDE-FM, a modular PDE foundation model combining spatial-spectral dual-modal tokenization, FiLM-based physics modulation, and a Mamba state-space backbone. It achieves an average 46% reduction in VRMSE across 12 heterogeneous physics-domain datasets from The Well benchmark.