A5012913879- Model Reduction and Neural Networks
- Probabilistic and Robust Engineering Design
- Lattice Boltzmann Simulation Studies
- Computational Physics and Python Applications
- Neural Networks and Applications
- Fluid Dynamics Simulations and Interactions
- Topology Optimization in Engineering
- Non-Destructive Testing Techniques
- Additive Manufacturing and 3D Printing Technologies
- Meteorological Phenomena and Simulations
- Generative Adversarial Networks and Image Synthesis
- Infrastructure Maintenance and Monitoring
- Advanced Numerical Methods in Computational Mathematics
- Injection Molding Process and Properties
- Fluid Dynamics and Turbulent Flows
- Advanced Multi-Objective Optimization Algorithms
- Risk and Safety Analysis
- Computer Graphics and Visualization Techniques
- Visual Attention and Saliency Detection
- Digital Transformation in Industry
- Gaussian Processes and Bayesian Inference
- Markov Chains and Monte Carlo Methods
- Electrical Contact Performance and Analysis
- Seismology and Earthquake Studies
- Railway Systems and Energy Efficiency
Nvidia (United States)
2021-2025
Nvidia (United Kingdom)
2020
University of Illinois Urbana-Champaign
2016-2019
George Washington University
2014
Abstract Physics‐informed neural networks (PINNs) are a class of deep that trained, using automatic differentiation, to compute the response systems governed by partial differential equations (PDEs). The training PINNs is simulation free, and does not require any data set be obtained from numerical PDE solvers. Instead, it only requires physical problem description, including governing laws physics, domain geometry, initial/boundary conditions, material properties. This usually involves...
Abstract To optimize mitigation, preparedness, response, and recovery procedures for infrastructure systems, it is essential to use accurate efficient means evaluate system reliability against probabilistic events. The predominant approach quantify the impact of natural disasters on systems Monte Carlo approach, which still suffers from high computational cost, especially when applied large systems. This article presents a deep learning framework accelerating seismic analysis, transportation...
Abstract Accurately predicting the dynamics of complex systems governed by partial differential equations (PDEs) is crucial in various applications. Traditional numerical methods such as finite element (FEMs) offer precision but are resource‐intensive, particularly at high mesh resolutions. Machine learning–based surrogate models, including graph neural networks (GNNs), present viable alternatives reducing computation times. However, their accuracy significantly contingent on availability...
Accurate near-term passenger train delay prediction is critical for optimal railway management and providing passengers with accurate arrival times. In this work, a novel bi-level random forest approach proposed to predict delays in the Netherlands. The primary level predicts whether will increase, decrease, or remain unchanged specified time frame. secondary then estimates actual (in minutes), given predicted category at level. For validation purposes, model has been compared several...
We propose the geometry-informed neural operator (GINO), a highly efficient approach to learning solution of large-scale partial differential equations with varying geometries. GINO uses signed distance function and point-cloud representations input shape operators based on graph Fourier architectures learn operator. The handles irregular grids transforms them into from regular latent which can be efficiently applied. is discretization-convergent, meaning trained model applied arbitrary...
The objective of this paper is to develop a straightforward, robust, stable, and accurate mesh-free numerical technique for modeling the dynamic behavior free surface, incompressible, multiphase granular flows. This method (henceforth, MPS method) based on fully Lagrangian moving particle semi-implicit (MPS). provides approximations strong form PDEs basis integral interpolants. fluid represented with particles, motion each calculated interactions neighboring particles by means kernel...
In this paper, we introduce a physics-driven regularization method for training of deep neural networks (DNNs) use in engineering design and analysis problems. particular, focus on prediction physical system, which addition to data, partial or complete information set governing laws is also available. These often appear the form differential equations, derived from first principles, empirically-validated laws, domain expertise, are usually neglected data-driven systems. We propose approach...
Numerical simulations play a critical role in design and development of engineering products processes. Traditional computational methods, such as CFD, can provide accurate predictions but are computationally expensive, particularly for complex geometries. Several machine learning (ML) models have been proposed the literature to significantly reduce computation time while maintaining acceptable accuracy. However, ML often face limitations terms accuracy scalability depend on significant mesh...
This paper introduces a data-driven algorithm for modeling and compensating shape deviations in additive manufacturing (AM), addressing challenges geometric accuracy batch production. While traditional methods, such as analytical models metrology, laid the groundwork precision, they are often impractical large-scale Recent advancements machine learning (ML) have improved compensation but issues remain generalizing across complex geometries adapting to position-dependent variations. We...
Abstract Climate change has intensified extreme weather events, with floods causing significant socioeconomic and environmental damage. Accurate flood forecasting is crucial for disaster preparedness risk mitigation, yet traditional hydrodynamic models, while precise, are computationally prohibitive real‐time applications. Machine learning surrogates, such as graph neural networks (GNNs), improve efficiency but often lack physical consistency interpretability. This paper introduces...
Abstract We propose a novel method for solving partial differential equations using multi-fidelity physics-informed generative adversarial networks. Our approach incorporates physics supervision into the optimization process to guide learning of generator and discriminator models. The has two components: one that approximates low-fidelity response input another combines generate an approximation high-fidelity responses. identifies whether input–output pairs accord not only with actual...
Metal Sintering is a necessary step for Injection Molded parts and binder jet such as HP's metal 3D printer. The sintering process introduces large deformation varying from 25 to 50% depending on the green part porosity. In this paper, we use graph-based deep learning approach predict deformation, which can speed up simulation substantially at voxel level. Running well-trained inferencing engine only takes range of seconds obtain final value. tested accuracy example complex geometry achieves...
A Multi-Resolution Weakly Compressible Moving-Particle Semi-Implicit (MR-WC-MPS) method is presented in this paper for simulation of free-surface flows. To reduce the computational costs, as with multi-grid schemes used mesh-based methods, there also a need particle methods to efficiently capture characteristics different flow regions levels complexity spatial resolutions. The proposed MR-WC-MPS allows use particles sizes domain, analogous multi-resolution grid grid-based methods. evaluate...
We present SimNet, an AI-driven multi-physics simulation framework, to accelerate simulations across a wide range of disciplines in science and engineering. Compared traditional numerical solvers, SimNet addresses use cases - coupled forward without any training data, inverse data assimilation problems. offers fast turnaround time by enabling parameterized system representation that solves for multiple configurations simultaneously, as opposed the solvers solve one configuration at time. is...
Topology Optimization is the process of finding optimal arrangement materials within a design domain by minimizing cost function, subject to some performance constraints. Robust topology optimization (RTO) also incorporates effect input uncertainties and produces with best average structure while reducing response sensitivity uncertainties. It computationally expensive carry out RTO using finite element Monte Carlo sampling. In this work, we use neural network surrogates enable faster...
In this paper, we propose the Adaptive Physics-Informed Neural Networks (APINNs) for accurate and efficient simulation-free Bayesian parameter estimation via Markov-Chain Monte Carlo (MCMC). We specifically focus on a class of problems which computing likelihood function requires solving PDE. The proposed method consists of: (1) constructing an offline PINN-UQ model as approximation to forward model; (2) refining approximate fly using samples generated from MCMC sampler. APINN constantly...