Papers & Surveys

Must-read benchmark papers and technical surveys (no news or industry reports)

20 resources

Does Transolver Really Need a Transformer?

CFD
This paper provides an empirical and theoretical analysis of Transolver’s physics-attention mechanism for three-dimensional fluid dynamics, examining how unstructured mesh points are assigned to slices, tokens interact through self-attention, and results are broadcast back to the points.
See LICENSE file neural operatorsphysics-attentionfluid dynamics

Physics-Informed Learning of Feedback-Linearizing Representations

ScientificML
The paper proposes a cascaded physics-informed neural network (PINN) framework that numerically solves the PDEs governing feedback-linearizing coordinate transforms, whose well-posedness follows from the Lie-algebraic Frobenius theorem, making them tractable for large-dimensional nonlinear control systems.
See LICENSE file Physics-Informed Neural NetworkFeedback LinearizationNonlinear Control

PINNMorph: Evolving Online Adaptation Policies for Physics-Informed Neural Networks

Scientific ML
PINNMorph is an online adaptation framework for physics-informed neural networks that addresses partial differential equation solving by evolving network architecture and major training mechanisms during optimization.
See LICENSE file Physics-Informed Neural NetworksPartial Differential EquationsOnline Adaptation

Uncovering Flame Physics with Machine Learning: Application to the Reaction Rate in Hydrogen Flames

CFD
This paper uses convolutional neural networks as analytical tools to investigate relationships between the progress variable field and local chemical source term in lean premixed hydrogen flames. The approach analyzes spatially resolved fields from direct numerical simulations to study combustion reaction-rate physics.
See LICENSE file Hydrogen FlamesCombustion Reaction RateConvolutional Neural Networks

Architecture Acceleration of Machine Learning Gravitational Waveform Surrogate Models

Geophysics
This paper presents differentiable and hardware-accelerated implementations of the machine-learning surrogate models mlgw and mlgw-bns for gravitational-wave waveform modeling, enabling JIT compilation, vector parallelization, GPU acceleration, and automatic differentiation within JAX.
See LICENSE file gravitational waveswaveform surrogate modelsGPU acceleration

Distance-Residual Physics-Informed Neural Networks: A Deep Learning Framework for Differential and Partial Differential Inclusions

ScientificML
Distance-Residual Physics-Informed Neural Networks provide a physics-informed deep learning framework for ordinary and partial differential inclusions. The method uses the squared distance from a differential operator to an admissible set-valued map instead of the classical pointwise PDE or ODE residual.
See LICENSE file Physics-Informed Neural NetworksDifferential InclusionsScientific Machine Learning

Equation Discovery with Bayesian Tree-Adjoining Grammars

ScientificML
A Bayesian tree-adjoining grammar framework for equation discovery and structure identification in general nonlinear systems.
See LICENSE file equation discoveryBayesian modelingtree-adjoining grammars

Gradient Surgery for Physics-Informed Neural Networks

ScientificML
This paper studies gradient conflicts and multi-task optimization in Physics-Informed Neural Networks, with implications for more stable convergence in stiff and high-frequency partial differential equations.
See LICENSE file Physics-Informed Neural NetworksGradient ConflictsMulti-Task Optimization

NEXT: Physics-Informed Neuro-Spectral Exponential Time Differencing Architectures

ScientificML
NEXT presents physics-informed neuro-spectral exponential time-differencing architectures for time-dependent PDEs, targeting the spectral bias and causality limitations of PINNs and the numerical instability of NeuSA for stiff differential equations in neural representations.
See LICENSE file physics-informed neural networksneuro-spectral architecturesexponential time differencing

On the Limits of Univariate Deep Learning for Significant Wave Height Forecasting

Fluid Mechanics
This study evaluates five deep learning architectures for univariate significant wave height forecasting using buoy observations. Across a 47-buoy, 37-year evaluation corpus, the model families converge to nearly identical performance, with an between-family standard deviation of 0.0014 square meters.
See LICENSE file significant wave height forecastingocean buoysdeep learning comparison
141–160 / 364
1 …6 7 8 9 10 …19
Go to