Scientific machine learning

topic9 events

papersTODAY 04:00 UTC

Study identifies derivative-fidelity failure mode in physics-informed neural networks

A new arXiv paper argues that physics-informed neural networks can match target function values while still producing inaccurate derivatives. The authors describe this as a distinct failure mode and provide strengthened benchmark evidence for it, based on models trained only on function values. The work suggests that evaluating PINNs by function agreement alone can mask errors in the derivative terms central to solving differential equations.

pinns/benchmarks/research/autodiff

papersTODAY 04:00 UTC

Sparse autoencoders used to probe physics-informed neural network internals

A new arXiv paper introduces PhysSAE, a method that applies sparse autoencoders to inspect what hidden layers in physics-informed neural networks actually represent. The authors aim to determine whether these networks learn localized, physically meaningful features tied to the PDE residuals they are trained on. The work falls within mechanistic interpretability research for scientific machine learning.

interpretability/sparse-autoencoders/pinns/research

papersTODAY 04:00 UTC

Pullback-corrected auxiliary variable optimizer targets multi-term scientific ML losses

A new arXiv paper proposes a pullback-corrected scalar auxiliary variable (PB-SAV) optimizer that adds momentum and adaptive mobility. The method is aimed at scientific machine learning objectives that combine several loss terms, such as the residual, boundary, initial, and data losses used in physics-informed neural networks. The abstract frames the work as addressing optimization challenges specific to these composite objectives.

optimization/scientific-ml/pinns/arxiv

papersTODAY 04:00 UTC

Graph Transformer Approach Reconstructs Detonation Flow Fields on Meshes

A new arXiv preprint presents a mesh-based super-resolution method that uses graph transformers to reconstruct multiscale detonation flow data. The authors argue such data-driven reconstruction is useful for subgrid closure modeling, faster spatiotemporal forecasting, compression, and as an upsampling step in simulations. The work appears as a cross-listed revision in the cs.AI and cs.LG categories.

graph-transformers/super-resolution/fluid-dynamics/arxiv

papersTODAY 04:00 UTC

Gradient Repair Method Aimed at Stabilizing Neural ODE Training

A new arXiv paper introduces GradRepair-ODE, a technique that certifies and repairs gradients when neural ordinary differential equations are trained. Because neural ODEs embed a numerical solver in the training loop, solver choices shape both the forward trajectory and the gradients sent to the optimizer, which the authors flag as a reliability issue for scientific machine learning. The approach is presented as a way to keep those gradients trustworthy during training.

neural-ode/training/research/scientific-ml

papersSEP 12 04:00 UTC

PINN Framework Infers Perpendicular Heat Conductivity in Stellarator Scrape-Off Layer

Researchers present an inverse physics-informed neural network that estimates how the scrape-off layer's perpendicular heat conductivity varies with plasma density and temperature in stellarator devices. The approach embeds physical constraints into the learning process rather than relying solely on labeled data, allowing the conductivity function to be recovered from available measurements. This is an arXiv preprint on fusion plasma modeling and has not yet been peer reviewed.

physics/neural-networks/fusion/scientific-ml

papersSEP 10 04:00 UTC

Physics-Guided Machine Learning Extrapolation Framework Validated on Diffusion Benchmark

A new arXiv paper introduces a physics-guided machine learning framework designed to make reliable predictions outside the limited operating ranges in which engineering models are typically trained. The authors argue that extrapolation, rather than interpolation, is the central challenge for applied ML, and they validate their approach using a classical transient diffusion problem as a benchmark.

physics-informed/extrapolation/machine-learning/research

papersSEP 10 04:00 UTC

Researchers build physics-informed surrogate model for Mars' nightside thermosphere

A new arXiv paper introduces a multi-task surrogate model that combines physical constraints with machine learning to simulate the Martian nightside thermosphere. The problem is difficult because direct measurements are sparse and transport, magnetic, and seasonal effects interact strongly, so purely data-driven approaches can produce unphysical outputs such as reversed density trends. Embedding physics into the training process aims to keep the model's predictions consistent with known atmospheric behavior.

space/physics-informed/machine-learning/research

papersSEP 10 04:00 UTC

Convolutional autoencoder and neural ODE framework for transient counterflow flame modeling

Researchers propose a reduced-order modeling framework that combines a convolutional autoencoder with a neural ordinary differential equation to serve as a surrogate for simulating transient two-dimensional counterflow flames. The approach extends autoencoder–neural ODE techniques, previously applied to homogeneous reactive systems, to spatially resolved combustion problems. Such surrogates can cut the computational cost of modeling reactive flows.

scientific-ml/neural-ode/combustion/surrogate-models