Proposes a unified scalable framework combining structured decomposition with SDP certificates and small-gain analysis for robust stability and performance analysis of feedback systems with high-dimensional NNs and unmodeled dynamics.
See LICENSE fileneural network verificationrobust controlsemidefinite programming
A method to predict how force correction changes energy exchange and structural statistics in molecular dynamics by measuring leading response coefficients on shared reference trajectories.
See LICENSE filemolecular dynamicsforce correctionresponse coefficients
Proposes AffordanceWAM, an affordance-aware generative World Action Model that leverages both action-labeled robot videos and egocentric human videos for generalizable robot manipulation.
See LICENSE filerobot manipulationaffordance learningworld model
Diff-NekRS is an open-source framework that enables automatic differentiation through the spectral element CFD solver NekRS, enabling gradient backpropagation across multi-timestep rollouts to train neural correction networks for error mitigation.
See LICENSE fileCFDDifferentiable FrameworkMulti-Timestep Training
This paper investigates physical diagnosis for robots facing unfamiliar conditions, defining six ordered empirical conditions required for behavior improvement, and validates the chain in controlled and public environments.
See LICENSE filerobot diagnosisphysical adaptationtest-time inference
This research applies Bayesian optimization and surrogate modeling to investigate urban canopy parameter ranges for heat mitigation, analyzing trade-offs among daytime air temperature, nighttime air temperature, and Universal Thermal Climate Index (UTCI) in urban environments.
See LICENSE fileUrban Heat IslandBayesian OptimizationSurrogate Model
This paper investigates the acoustic dip phenomenon in the effective temperature of hot Brownian particles in compressible solvents, revealing how finite sound speed modifies the fluctuation-dissipation relation in the kinetic regime.
See LICENSE filehot Brownian motionacoustic effectfluctuation-dissipation
Proposes TASE-stabilized Time-Evolving Natural Gradient (TENG) method that projects time-discrete target states onto the neural-network manifold via local natural-gradient iterations to address stability constraints in stiff diffusion-dominated PDEs.
See LICENSE fileTime-Evolving Natural GradientDiffusion-Dominated PDEsNeural PDE Approximation
Introduces a computational framework for mechanical systems combining reservoir encoding and active inference, enabling adaptive learning in changing environments beyond traditional passive design and offline optimization constraints.
See LICENSE fileReservoir ComputingActive InferenceCognitive Behavior
Proposes test-time trained flow matching for Bayesian filtering, enabling continuous distribution representation for state estimation in high-dimensional physical systems.
See LICENSE filestate estimationflow matchingBayesian filtering
Presents a method using sparse differentiation to compute exact molecular Hessians from machine learning interatomic potentials, reducing computational complexity from O(N²) to O(N), overcoming limitations for large atomistic systems.
See LICENSE filemolecular dynamicsinteratomic potentialsHessian matrix
Establishes asymptotic-preserving property of exponential Euler method for stochastic reaction-diffusion-advection equations in R² under fast Hamiltonian advection with multiple critical points, reducing dynamics to a stochastic PDE on a noncompact metric graph.
See LICENSE fileexponential Euler methodasymptotic preservationstochastic PDE
Proposes NeSTR method that learns the S-Transform on a finite dimensional subspace of white noise test space to reconstruct stochastic solutions for Gaussian SPDEs.
See LICENSE fileSPDENeural S-TransformGaussian Process
This paper presents an integrated CFD-machine learning framework for optimizing pod geometry in supersonic evacuated tube transport, using composite cubic Bezier curves with seven design variables and Latin Hypercube Sampling to evaluate 200 configurations across subsonic and supersonic regimes.
See LICENSE fileevacuated tube transportsupersonic aerodynamicspod optimization
Proposes D3DWA, an adaptive DWA framework using Dueling Double Deep Q-Network to adjust weights and prediction horizon for improved robot navigation in varied environments.
See LICENSE filerobot navigationdeep reinforcement learningpath planning
Proposes neural network-based pulse deconvolution for FCC-ee's high-granularity dual-readout calorimeter, leveraging timing separation of Cherenkov and scintillation signals to improve energy resolution.
See LICENSE fileParticle DetectorCalorimeterPulse Deconvolution
DeepFEAv2 applies deep learning surrogate models to transient finite element analysis on unstructured meshes, enabling joint prediction of node-based and element-based outputs to reduce computational cost in structural mechanics.
See LICENSE fileFinite Element AnalysisDeep LearningTransient Simulation
DeViGrasp introduces methods to systematically examine how occlusion, segmentation-mask dropout, depth noise, and target-localization jitter affect grasp reasoning and target tracking for quadruped manipulators, improving whole-body control robustness under degraded perception.
See LICENSE filequadruped robotrobust graspingdegraded perception
Proposes a subcell-refined entropy-residual-driven limiter for high-order DG methods on Legendre-Gauss-Lobatto nodes, enabling robust simulation of under-resolved nonlinear hyperbolic systems while preserving accuracy via nearest-neighbor pairwise dissipation.
See LICENSE fileDiscontinuous Galerkinentropy residualhigh-order methods