⚛️ Physics & Scientific Computing¶
🧠 NeurIPS2026 · 61 paper notes
📌 Same area in other venues: 🎞️ ECCV2026 (9) · 📷 CVPR2026 (10) · 🔬 ICLR2026 (76) · 🧪 ICML2026 (36) · 🤖 AAAI2026 (16) · 🧠 NeurIPS2025 (64)
🔥 Top topics: Diffusion Models ×2
- APPSolver: Adaptive Patch Partitioning for Point-Wise Ship Flow Prediction on Unstructured Meshes
-
To alleviate the prohibitive attention cost on large, non-uniform unstructured ship CFD meshes, APPSolver introduces deterministic Adaptive Patch Partitioning (APP) guided by hull distance priors and spatial downsampling, enabling ultra-fast point-wise flow prediction (1.3 ms / 1.8 GFLOPs per forward pass) followed by distance-weighted \(k\)-NN full-reference recovery.
- AtomMOF: All-Atom Flow Matching for MOF-Adsorbate Structure Prediction
-
This paper introduces AtomMOF, the first all-atom variational flow matching model for metal-organic frameworks (MOFs) and host-guest adsorption complexes, which maps 2D molecular graphs of building blocks and adsorbates directly to unconstrained 3D atomic coordinates and lattice parameters and incorporates Feynman-Kac steering guided by machine-learned interatomic potentials (MLIPs) to boost match rate by 35.00% and reduce RMSD by 32.68% on the BW benchmark while outperforming Grand Canonical Monte Carlo (GCMC) by orders of magnitude in sampling speed on ODAC25.
- Breakeven Complexity: A New Perspective on Neural Partial Differential Equation Solvers
-
Addressing the fundamental blind spot where standard accuracy metrics ignore both the substantial up-front training costs of neural PDE surrogates and the low computational cost of low-fidelity numerical solvers, this paper introduces a compute-aware "breakeven complexity" evaluation framework optimized via scaling laws, accompanied by the GPU-native PyFR multi-obstacle benchmark BreakFlow, demonstrating that neural surrogates achieve practical amortization advantages primarily on computationally demanding, high-difficulty physical regimes.
- ClothTransformer: Unified Latent-Space Transformers for Scalable Cloth Simulation
-
ClothTransformer reformulates cloth simulation as autoregressive sequence modeling in a learned latent space, compressing variable-resolution meshes into fixed-size latent tokens via cross-attention and integrating differentiable Continuous Collision Detection (CCD) to deliver unified, scalable, and penetration-free simulation across body-driven garments, robotic manipulation, and free-fall collisions.
- CME–SpectrumBench: Can LLMs Analyze Condensed Matter Spectral Data?
-
Researchers from Stanford and UC Berkeley introduce CME–SpectrumBench, the first expert-designed benchmark evaluating frontier LLMs on research-level ARPES/RIXS spectral data inverse analysis across 190 questions, demonstrating severe degradation under direct dense context reasoning and showing that code-generating agents with Python interpreter access achieve parity with graduate-level handwritten code.
- Compute-Optimal Data Scaling for Neural Surrogates via Multi-Fidelity Training
-
This paper formalizes neural surrogate training as an explicit function of data quantity and fidelity under fixed generation compute budgets, demonstrating that continuous multi-fidelity conditioning yields Pareto-optimal scaling that matches high-fidelity accuracy with up to 6.3× less simulation compute.
- Coupled Integral PINN for Discontinuity
-
Addressing the optimization barrier where strong-form PINNs exhibit diverging residuals and collapse into overly smooth non-physical states at discontinuities, CI-PINN couples a primitive variable network with an auxiliary integral potential network under entropy and adaptive mask constraints, achieving robust mesh-free shock and rarefaction capturing.
- Deep-Koopman-KANDy: Dictionary Discovery for Deep-Koopman Operators with Kolmogorov-Arnold Networks for Dynamics
-
By replacing Deep-Koopman autoencoders with two-layer Kolmogorov-Arnold Networks (KANs) and deriving a chain-rule gradient projection for level-set manifold analysis, this paper decouples representation learning from symbolic description and enables post-hoc dictionary discovery without prior library commitment.
- diffusion-warm sampling of the xy model enables fast thermalization at scale
-
To address the exponential explosion of thermalization times in traditional MCMC for large-scale, low-temperature 2D continuous-spin XY models, this paper proposes using a temperature-conditioned diffusion model trained on small lattices as a warm start, requiring only tens of Wolff MCMC steps to calibrate observables and achieving a tenfold acceleration in thermalization time.
- Discovering Phase Space Structure in Learned Hamiltonian Systems
-
SPS-GAN introduces a conditional generative framework combining port-Hamiltonian dynamics with an extended cyclic-coordinate regularization loss, enabling joint physical trajectory generation and the unsupervised discovery of system degrees of freedom and canonical mechanical coordinates from raw Cartesian observations and video streams without requiring pre-specified phase space dimensions.
- Efficient, Accurate and Stable Gradients for Neural ODEs
-
Introduces a general class of algebraically reversible single-step ODE solvers that achieve exact gradients, high-order convergence, and non-zero linear stability, training Neural ODEs in \(\mathcal{O}(n)\) time and \(\mathcal{O}(1)\) memory to strictly outperform recursive checkpointing.
- Efficient Neural Field Learning via Adaptive Coverage and Focused Sampling
-
Addressing the training inefficiency of implicit neural representations on heterogeneous continuous fields where uniform sampling is wasteful and pointwise adaptive methods cause redundant localized over-clustering while compromising global coverage, this paper presents ACES (Adaptive Coverage-aware Efficient Sampling), a structured framework combining adaptive recursive spatial partitioning with controlled-bias region-level weighting.
- End-to-End Differentiable Diffusion Conditioning for Physics-Informed Optimization
-
This paper introduces Differentiable Diffusion Conditioning (D2C), which turns a frozen conditional diffusion model into an end-to-end gradient-based optimizer by differentiating through the deterministic DDIM sampling chain back to the conditioning variables, optimizing complex physical systems under strict physical constraints.
- Exactness Matters for Physical Rule Enforcement
-
This paper identifies an operator-exactness regime split in autoregressive scientific forecasting: while exact on-manifold projection substantially suppresses rollout error in periodic flows, approximate boundary-preserving cleanup can cause severe target distortion and distribution shift, meaning enforcement strength must be validated via target distortion rather than linear-system residuals.
- Fast Learning Rates for Physics-Informed Kernel Methods
-
Focusing on learning target functions from noisy value and linear differential observations, this paper establishes finite-sample prediction error bounds for physics-informed kernel methods (PIKS) via a novel value-derivative capacity decomposition, proving a two-regime phase transition from a differential-limited rate to an oracle saturation rate while retaining physically consistent convergence in stronger norms.
- Fast Reconstruction of Exact Maxwell Dynamics from Sparse Data
-
Proposes FLASH-MAX, an exact-by-construction shallow neural network whose hidden neurons analytically satisfy the homogeneous Maxwell equations, reaching sub-1% relative validation error within seconds from sparse data while preserving a zero PDE residual.
- Few-Step Boltzmann Generators via Scalable Likelihood Flow Maps
-
Addressing the computational bottleneck where continuous flow likelihood estimation relies on high-variance Hutchinson trace estimators or expensive Jacobian determinants, this paper proposes SCALLOP, a trace-free, vectorizable conditional divergence matching framework that enables molecular Boltzmann generators to perform fast sampling and exact likelihood evaluation within a few function evaluations.
- GAIA: Geometry-Adaptive Operator Learning for Forward and Inverse Problems
-
Overcoming the structural limitation where existing geometry-adaptive neural operators require shared input-output spatial discretizations and fail to solve inverse or boundary value problems in a single pass, GAIA introduces dual-pathway geometry tokenization and spatially adaptive cross-attention kernel conditioning, enabling end-to-end single-pass solutions across varying irregular geometries.
- Generalization Error Bounds for Picard-Type Operator Learning in Nonlinear Parabolic PDEs
-
The paper derives implementation-agnostic generalization error bounds for Duhamel-Picard operator learning in nonlinear parabolic PDEs, proving that increasing Picard depth geometrically reduces truncation error without enlarging the entropy-based Rademacher complexity, thereby circumventing the curse of parametric complexity.
- Geometry-Aware Post-Hoc Uncertainty Quantification in Operator Learning
-
REEF-GP is a post-hoc uncertainty quantification framework that fits a Gaussian process to the residuals of a frozen neural operator using its internal multiscale geometry representations as a deep kernel feature space, matching Deep Ensembles in calibration quality without sacrificing predictive accuracy.
- Geometry Meets Physics: Data-Efficient Pre-Training for Unstructured Neural PDE Solvers
-
Addressing the severe computational bottleneck of generating massive simulation datasets for unstructured 3D neural PDE solvers, this paper introduces GMP (Geometry Meets Physics), the first entirely disk-data-free online pre-training framework that decouples steady-state differential geometry reconstruction from transient synthetic PDE masked autoencoding, achieving faster convergence and up to 47% relative error reduction in low-data regimes on the modular GMP-Net operator.
- GeoTransolver: Learning Physics on Irregular Domains using Multi-scale Geometry Aware Physics Attention Transformer
-
To address geometry dilution across depth, boundary-layer to far-field scale mismatch, and single-shot regime conditioning in industrial CAE, GeoTransolver equips Transolver with Geometry-Aware Latent Embeddings (GALE) and multi-scale context projectors, cutting flow and structural crash errors across massive irregular domains.
- GIST: Gauge-Invariant Spectral Transformers for Scalable Graph Neural Operators
-
Addressing the conflict between cubic eigendecomposition overhead and broken gauge invariance that degrades cross-resolution generalization in mesh neural operators, GIST restricts self-attention to pairwise inner products of randomized spectral embeddings, achieving end-to-end \(\mathcal{O}(N)\) scaling with the first provable discretization-mismatch bound for graph neural operators.
- How deep is your network? Deep vs. shallow learning of transfer operators
-
This paper introduces RaNNDy, a randomized neural network framework combined with variational principles for learning transfer operators in dynamical systems; by freezing random hidden layer weights and training only the output layer via closed-form generalized eigenvalue solutions, it achieves competitive accuracy with deep VAMPnets while speeding up training by two to three orders of magnitude.
- In-Context Multi-Operator Learning with DeepOSets
-
DeepOSets introduces a permutation-invariant, non-autoregressive architecture combining DeepSets and DeepONet for in-context multi-operator learning, establishing the first universal approximation theorem for compact operator sets and enabling minute-level training with robust generalization across differential equations.
- Inpainting physics: self-supervised learning for context-driven fluid simulation
-
Steady computational fluid dynamics (CFD) prediction is reformulated as self-supervised spatial context inpainting; by tokenising high-resolution velocity fields into compact spatial latents, latent flow matching and masked autoencoders enable robust flow completion and efficient local geometry editing.
- Learning Discrete Riemannian Metrics for Physical Fields with Cochain-Frame Equivariance
-
By freezing topological cellular coboundaries to strictly preserve exact differential identities and learning symmetric positive-definite discrete Riemannian cochain metrics with cochain-frame equivariance, RHMP achieves an exact architectural separation between topological conservation laws and learnable geometry-dependent propagation.
- Learning Where to Simulate: Generative Active Sampling for Online PDE Surrogate Training
-
Addressing the severe under-representation of extreme physical dynamics caused by uniform sampling in PDE surrogate training, this paper introduces Online Generative Active Sampling (OGAS), an asynchronous framework pairing a conditional diffusion model with density-ratio debiasing to adaptively steer numerical simulations toward challenging physical regimes, substantially compressing worst-case and tail prediction errors.
- LEIA: Learned Environment for Interactive Architected Materials
-
LEIA introduces an interactive, action-conditioned world model for architected metamaterials and nonlinear solid mechanics, decoupling mesh resolution via Perceiver latent compression and utilizing a lightweight stress head to directly predict Cauchy stress tensors, achieving high-fidelity real-time autoregressive stress-displacement rollout on large 3D unstructured meshes up to 442k nodes.
- LEVDA: Latent Ensemble Variational Data Assimilation via Differentiable Dynamics
-
To tackle long-range forecast degradation in geophysical systems under chaos and severe observation sparsity, LEVDA conducts four-dimensional ensemble-variational (4DEnVar) smoothing within the low-dimensional latent subspace of a pretrained differentiable neural dynamics surrogate, enabling adjoint-free, encoder-free joint state-parameter assimilation with arbitrary irregular spatiotemporal sampling.
- Lie Generator Networks for Nonlinear Partial Differential Equations
-
LGN-KM lifts nonlinear PDEs into a linear latent space and decomposes the continuous-time Koopman generator into a shared skew-symmetric conservative coupling matrix and a positive-definite diagonal modal dissipation matrix (\(L_k = S - D_k\)), providing mathematically guaranteed long-horizon stability and extracting multi-branch dispersion relations directly from trajectory data.
- M\(^3\): Reframing Training Measures for Discretized Physical Simulations
-
Addressing the measure-induced optimization bias caused by non-uniform discretization grids in neural physical simulation surrogates, this paper introduces M\(^3\) (Multi-scale Morton Measure), an empirical measure construction framework based on variation-aware Morton octree partitioning and capacity-constrained water-filling allocation, achieving substantial physical consistency and fidelity gains on industrial CFD benchmarks with up to 10× less data.
- Mechanism-Aware Ensemble Conditioning for Data-Limited Emulation of Extreme Events
-
The paper compresses a reference-nudged stochastic coarse ensemble into dynamical-sensitivity context and uses FiLM to modulate existing temporal correctors, improving QG extreme-event area and frequency statistics with limited high-resolution training data without guaranteeing better global distributions or every tail diagnostic than a long-data baseline.
- Mechanisms of Misgeneralization in Physical Sequence Modeling
-
Reveals why generative sequence models trained on trajectory marginals produce locally plausible rollouts yet systematically distort aggregate physical distributions (physical misgeneralization), formulates data deviation kernels to anticipate mass drift a priori, and proposes a kernel-informed coordinate transform representation that rectifies drift without conditioning.
- Mesh-Free Convolution: Learned Spectral Attenuation and Transport
-
The paper introduces Mesh-Free Convolutions (MFCs), a continuous-space operator learning framework parameterizing convolutional filters as solution operators of anisotropic parabolic diffusion and advection PDEs, replacing hard Fourier mode truncation with smooth learned spectral attenuation and directional transport to achieve stable long-horizon autoregressive rollouts and robust zero-shot cross-resolution generalization.
- NAGO: Noise-Aware Generative Operator via Flow Matching
-
To tackle spectral stiffness and cross-resolution distortion in sampling invariant measures of stochastic PDEs, NAGO treats the source distribution as an active design variable, establishing an admissibility-constrained spectral-matching rule that utilizes tailored Matérn noise to smooth marginal velocity fields.
- Neural Harmonic Measure Operator
-
NHMO learns a geometry-only harmonic-measure density, integrates boundary data against its normalized kernel, and uses a learned lift for source contributions and approximation error, outperforming four compared baselines across all five 3D MCB-B Poisson categories while reducing repeated solves on one geometry to a cached matrix-vector product and a lift forward pass.
- Neural Modal Decomposition: Architectural Priors from Observables
-
Addressing the \(\mathcal{O}(N^2)\) output dimensionality explosion and ill-posed rational fitting pathologies in multi-port linear time-invariant surrogate modeling, Neural Modal Decomposition (NMD) integrates the canonical pole-residue solution as an architectural prior trained end-to-end purely on system-level observables, enabling zero-shot generalization to unseen port counts and emergent physical eigenmode discovery.
- Neural Quantum Spectral Operator Learning for Solving Partial Differential Equations
-
Addressing the sign ambiguity and instance-specific optimization bottlenecks of standard VQLS, NVQLS integrates the Legendre-Galerkin weak formulation with a phase-aware state overlap objective and classical neural parameter embedding, realizing label-free unsupervised quantum spectral operator learning with theoretical scaling advantages.
- NOFE – Neural Operator Function Embedding
-
Addressing the fundamental limitations of classical dimensionality reduction techniques that treat physical fields as isolated discrete point clouds while ignoring underlying continuous domain topology and sampling independence, NOFE reformulates continuous dimensionality reduction as a sheaf morphism via Graph Kernel Operators, enabling mesh-free function-to-function mapping and local topological preservation across arbitrary query locations.
- Optimal Algorithmic Complexity of Inference in Quantum Kernel Methods
-
This paper systematically characterizes the algorithmic complexity of inference in quantum kernel methods, achieving the query-optimal complexity of \(\mathcal{O}(\|\alpha\|_1 / \varepsilon)\) by encoding the full inference sum into a single observable expectation value combined with quantum amplitude estimation.
- PDE-PFN: Prior-Data Fitted Neural PDE Solver
-
Addressing the reliance of conventional neural operators on regular grids, symbolic equations, and clean data, PDE-PFN uses a prior-data fitted network with Transformer self- and cross-attention to approximate the posterior predictive mean over noisy PINN solution priors, enabling zero-gradient-update in-context PDE prediction on arbitrary coordinate observations without explicit equations.
- Per-Loss Adapters for Gradient Conflict in Physics-Informed Neural Networks
-
Addressing optimization stagnation caused by multi-loss gradient conflict in physics-informed neural networks, this work demonstrates that gradient conflict exhibits distinct regime dependence, introduces a 1000-step diagnostic-first workflow, and proposes per-loss low-rank adapters that provide explicit loss-indexed parameter subspaces while strictly preserving a single physical output.
- Phaedra: Learning High-Fidelity Discrete Tokenization for the Physical Sciences
-
Phaedra separates physical-field latents into multidimensional morphology tokens and high-precision one-dimensional amplitude tokens, then learns to recombine and decode them, reducing in-distribution PDE reconstruction nMAE from FSQ's 2.603 to 1.522 and compressing 65.0 GB of scientific data to 3.44 GB.
- PR-Smoother: Simulator-Preserving Non-Gaussian Smoothing for Data Assimilation
-
PR-Smoother learns a non-Gaussian initial-state posterior and future-conditioned stepwise corrections in physical space while retaining the prescribed simulator, using an observation-only variational objective to jointly estimate states, physical parameters, and sensor biases, with validation on nonlinear Lorenz–96 and a 16,384-dimensional fluid system.
- QMaxCal: Path-Space Regularization for Open Quantum Control via Girsanov's Theorem
-
By applying Girsanov's theorem to continuous homodyne measurement records of open quantum systems, this work derives a closed-form differentiable trajectory-space KL divergence that regularizes optimal control pulses toward decoherence-free manifolds without manual reward engineering.
- Random Neural Network Expressivity for Non-Linear Partial Differential Equations
-
Addressing the approximation of non-linear evolutionary partial differential equations in mixed spatiotemporal Sobolev spaces, this paper establishes a dimension-free \(\mathcal{O}(N^{-1/2})\) approximation error bound for random neural networks (RaNNs) via higher-order Ridgelet analysis and provides quantitative convergence guarantees for PINN training residuals to zero on porous medium and compressible Navier-Stokes equations.
- Reformulating Neural Operators in d+1 Dimensions for Embedding Evolution
-
This paper reformulates neural operators on a \((d+1)\)-dimensional product domain by introducing an auxiliary embedding coordinate, proposing a basis-diversified auxiliary evolution module (SKNO) that evolves embeddings via structured dual-basis operators instead of brute-force width scaling, achieving state-of-the-art relative \(L_2\) errors across over ten PDE benchmarks.
- SBNO : Schrödinger Bridge Neural Operator for the Forward and Inverse Problems under Missing Information
-
SBNO establishes an equation-free, paired-data diffusion bridge matching operator framework that unifies forward and inverse partial differential equation (PDE) solving under partial observations into a single drift network pair, delivering high-fidelity deterministic point reconstructions via probability-flow ODE integration without requiring observation guidance tuning or governing equation constraints.
- Smooth Piecewise Cutting for Neural Operator to Handle Discontinuities and Sharp Transitions
-
Cut-DeepONet proposes a two-stage operator learning framework that combines an input-dependent Cutting Net with higher-dimensional coordinate lifting to decompose discontinuous physical fields into smooth sub-pieces, enabling the explicit generation of true step discontinuities without directly fitting steep gradients.
- Soft Geometric Inductive Bias for Object Centric Dynamics
-
The paper encodes provided object states as Clifford multivectors and learns next-timestep states with geometric-product layers and a block-causal Transformer that do not enforce exact equivariance, improving prediction and autoregressive rollouts in symmetry-breaking settings such as wall collisions, anisotropic confinement, and real driving trajectories.
- Sparsely-Supervised Data Assimilation via Physics-Informed Schrödinger Bridge
-
Addressing the bottleneck of unavailable full-field high-fidelity training supervision in partial differential equation systems, Physics-Informed Conditional Schrödinger Bridge (PICSB) introduces periodic surrogate-endpoint refresh and hard observation projection throughout sampling, training the transport velocity purely via physical PDE residuals to achieve sub-second zero-shot full-field assimilation.
- SPDEBench: An Extensive Benchmark for Learning Stochastic PDEs
-
Addressing the lack of unified benchmarks and controlled simulations for stochastic partial differential equations (SPDEs), SPDEBench introduces a rigorous testbed across 1D–3D regular and singular SPDEs with exact renormalization schemes, 7 evaluation metrics, and a renormalization-aware operator baseline (NSPDE-S).
- Stein Kernelized Molecular Dynamics for Active Learning of Interatomic Potentials
-
Addressing the challenges of metastability entrapment and loss of physical Boltzmann fidelity in active learning for interatomic potentials, this paper proposes Stein Kernelized Molecular Dynamics (SKMD) and an adaptive online stopping criterion, achieving Boltzmann-consistent enhanced sampling and state-of-the-art data acquisition efficiency across diverse potential energy surfaces.
- Structured Unitary Tensor Network Representations for Circuit-Efficient Quantum Data Encoding
-
The paper introduces TNQE, a circuit-efficient quantum data encoding framework based on structured unitary tensor network representations that maps classical data into shallow quantum circuits via quantized tensor train decomposition and learnable block unitaries, cutting circuit depth to 0.04× of amplitude encoding while maintaining downstream accuracy on real quantum processors.
- Symplectic Neural Operators for Learning Infinite Dimensional Hamiltonian Systems
-
Addressing the long-term energy drift and numerical instability of standard neural operators, this paper proposes the Symplectic Neural Operator (SNO), which enforces symplectic geometric constraints by composing gradient-based shear blocks parameterized by self-adjoint Fourier neural operators, establishing a rigorous modified-Hamiltonian long-time stability guarantee.
- The Spectral Amplitude Principle for Dynamics of Quantum Neural Networks
-
The paper theoretically and empirically establishes the "Spectral Amplitude Priority" principle in Quantum Neural Networks (QNNs), demonstrating that QNN optimization is prioritized by the spectral magnitude of residual components rather than their frequency indices, effectively breaking the fundamental low-frequency spectral bias of classical deep neural networks.
- Theory Guided and Interpretable Neural Operator Design for Partial Differential Equation Learning
-
AFDONet introduces the first top-down neural operator designed strictly from Adaptive Fourier Decomposition (AFD) theory, projecting latent representations onto an RKHS and reconstructing manifold PDE solutions via dynamic convolutional kernels to outperform prior neural operators by orders of magnitude on curved geometries and sharp gradients.
- Toward the Goldilocks: Blind Compression of Quantum States
-
This paper investigates blind single-copy compression of pure quantum states, proving that achieving information-theoretic optimal reconstruction over all CPTP channels requires sharply \(k\) encoder ancillas and \(n\) decoder ancillas, while establishing that isometric decoders are near-optimal for practical low-dimensional concentrated states.
- Winning Lottery Tickets in Neural Networks via a Quantum-Inspired Classical Algorithm
-
For sampling winning lottery tickets in shallow neural networks based on the ridgelet transform, this paper exploits the isometry of the discrete ridgelet transform to reduce high-dimensional dense matrix inversion to empirical data support, establishing a fully classical algorithm that achieves polynomial-time sampling and sparse subnetwork extraction without quantum hardware.
- ZNO: Stable Rational Neural Operators in the Z-Domain for Discrete-Time Dynamics
-
Addressing the continuous-time bias and numerical instability of conventional neural operators on discrete dynamics, ZNO parameterizes low-rank MIMO rational filters directly in the complex z-plane with smooth unit-disk stability guarantees and fused causal recurrence for linear-time discrete system identification.