Fernando Alves Rochinha
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Fernando Alves Rochinha has a total of 126 co-authors. Here are the top co-authors by number of shared publications:
- Alvaro L. G. A. Coutinho
Department of Civil Engineering, COPPE/Federal University of Rio de Janeiro, Rio de Janeiro, Brazil
12 shared publications - Eshan V. Dave
Univ. of New Hampshire
6 shared publications - Glaucio H. Paulino
Department of Civil and Environmental Engineering, Princeton University
6 shared publications Marek-Jerzy Pindera
6 shared publications- Renato Nascimento Elias
COPPE/Federal University of Rio de Janeiro
6 shared publications Robert H. Dodds
6 shared publications
Latest Publications
- Open Access
- 27 September 2026
- Published by Elsevier in SSRN Electronic Journal
Enabling Uncertainty Quantification in Seismic Images Employing Generative Multimodal Variational Autoencoders
Bayesian seismic imaging provides a natural framework for uncertainty quantification, although its practical implementation is limited by the high computational cost of posterior sampling using full-wave-equation simulations. Data-driven surrogate modeling offers an effective alternative in which a machine-learning model learns a functional mappingbetween inputs and outputs and replaces repeated PDE-based forward simulations. Although the training phase requires significant computation, the resulting surrogate enables fast, low-cost inference and repeated uncertainty-aware estimation. In this context, we propose a deep generative surrogate model based on a multimodalvariational autoencoder (mVAE) network that samples migrated seismic images from target areas of interest to support uncertainty quantification. Thus, rather than performing a new amortized inference, the mVAE learns the distribution of an ensemble of migrated images, enabling fast resampling of the underlying posterior at negligible cost. The mVAE implements a two-level generative process in which a continuous latent variable is sampled conditionally on a discrete (categorical) latent component. Learning the categorical latent variable encodes, unsupervisedly, the multimodality of the original seismic amplitudes across seismic image realizations. We train the mVAE model on a dataset of seismic images generated by a workflow designed for seismic imaging with quantified uncertainty. Our results demonstrate that the mVAE network accurately reproduces the multimodality observed in the reference seismic image amplitudes, thereby reflecting the system’s variability. This conclusion is supported by comparisons of the amplitude histograms of the reference and generated samples, together with uncertainty measures such as the variability index and the $p$-Wasserstein distance, with the mVAE reducing the pixel-wise Wasserstein distance by an order of magnitude relative to the traditional Gaussian VAE.
- 27 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEnabling Uncertainty Quantification in Seismic Images Employing Generative Multimodal Variational Autoencoders
Bayesian seismic imaging provides a natural framework for uncertainty quantification, although its practical implementation is limited by the high computational cost of posterior sampling using full-wave-equation simulations. Data-driven surrogate modeling offers an effective alternative in which a machine-learning model learns a functional mappingbetween inputs and outputs and replaces repeated PDE-based forward simulations. Although the training phase requires significant computation, the resulting surrogate enables fast, low-cost inference and repeated uncertainty-aware estimation. In this context, we propose a deep generative surrogate model based on a multimodalvariational autoencoder (mVAE) network that samples migrated seismic images from target areas of interest to support uncertainty quantification. Thus, rather than performing a new amortized inference, the mVAE learns the distribution of an ensemble of migrated images, enabling fast resampling of the underlying posterior at negligible cost. The mVAE implements a two-level generative process in which a continuous latent variable is sampled conditionally on a discrete (categorical) latent component. Learning the categorical latent variable encodes, unsupervisedly, the multimodality of the original seismic amplitudes across seismic image realizations. We train the mVAE model on a dataset of seismic images generated by a workflow designed for seismic imaging with quantified uncertainty. Our results demonstrate that the mVAE network accurately reproduces the multimodality observed in the reference seismic image amplitudes, thereby reflecting the system’s variability. This conclusion is supported by comparisons of the amplitude histograms of the reference and generated samples, together with uncertainty measures such as the variability index and the $p$-Wasserstein distance, with the mVAE reducing the pixel-wise Wasserstein distance by an order of magnitude relative to the traditional Gaussian VAE.
- 29 April 2026
- Published by Elsevier in Journal of Computational Physics
- Vol. 561
Hybrid DeepONet surrogates for multiphase flow in porous media
- 29 April 2026
- Published by Elsevier in Journal of Computational Physics
- Vol. 561
Hybrid DeepONet surrogates for multiphase flow in porous media
- Open Access
Hybrid DeepONet Surrogates for Multiphase Flow in Porous Media
The solution of partial differential equations (PDEs) plays a central role in numerous applications in science and engineering, particularly those involving multiphase flow in porous media. Complex, nonlinear systems govern these problems and are notoriously computationally intensive, especially in real-world applications and reservoirs. Recent advances in deep learning have spurred the development of data-driven surrogate models that approximate PDE solutions with reduced computational cost. Among these, Neural Operators such as Fourier Neural Operator (FNO) and Deep Operator Networks (DeepONet) have shown strong potential for learning parameter-to-solution mappings, enabling the generalization across families of PDEs. However, both methods face challenges when applied independently to complex porous media flows, including high memory requirements and difficulty handling the time dimension. To address these limitations, this work introduces hybrid neural operator surrogates based on DeepONet models that integrate Fourier Neural Operators, Multi-Layer Perceptrons (MLPs), and Kolmogorov-Arnold Networks (KANs) within their branch and trunk networks. The proposed framework decouples spatial and temporal learning tasks by splitting these structures into the branch and trunk networks, respectively. We evaluate these hybrid models on multiphase flow in porous media problems ranging in complexity from the steady 2D Darcy flow to the 2D and 3D problems belonging to the $10$th Comparative Solution Project from the Society of Petroleum Engineers. Results demonstrate that hybrid schemes achieve accurate surrogate modeling with significantly fewer parameters while maintaining strong predictive performance on large-scale reservoir simulations.
Open AccessHybrid DeepONet Surrogates for Multiphase Flow in Porous Media
The solution of partial differential equations (PDEs) plays a central role in numerous applications in science and engineering, particularly those involving multiphase flow in porous media. Complex, nonlinear systems govern these problems and are notoriously computationally intensive, especially in real-world applications and reservoirs. Recent advances in deep learning have spurred the development of data-driven surrogate models that approximate PDE solutions with reduced computational cost. Among these, Neural Operators such as Fourier Neural Operator (FNO) and Deep Operator Networks (DeepONet) have shown strong potential for learning parameter-to-solution mappings, enabling the generalization across families of PDEs. However, both methods face challenges when applied independently to complex porous media flows, including high memory requirements and difficulty handling the time dimension. To address these limitations, this work introduces hybrid neural operator surrogates based on DeepONet models that integrate Fourier Neural Operators, Multi-Layer Perceptrons (MLPs), and Kolmogorov-Arnold Networks (KANs) within their branch and trunk networks. The proposed framework decouples spatial and temporal learning tasks by splitting these structures into the branch and trunk networks, respectively. We evaluate these hybrid models on multiphase flow in porous media problems ranging in complexity from the steady 2D Darcy flow to the 2D and 3D problems belonging to the $10$th Comparative Solution Project from the Society of Petroleum Engineers. Results demonstrate that hybrid schemes achieve accurate surrogate modeling with significantly fewer parameters while maintaining strong predictive performance on large-scale reservoir simulations.
- Open Access
- 6 March 2025
- Published by Elsevier in Journal of Environmental Management
- Vol. 379, 124708
Bayesian uncertainty estimation of adsorption closure models in the computational simulation of contaminant transport
- 6 March 2025
- Published by Elsevier in Journal of Environmental Management
- Vol. 379, 124708
Open AccessBayesian uncertainty estimation of adsorption closure models in the computational simulation of contaminant transport
- Open Access
- 2 December 2024
- Published by Softaliza Tecnologias LTDA in Ibero-Latin American Congress on Computational Methods in Engineering (CILAMCE)
Ablation study of a non-intrusive data-driven surrogate model for the Rayleigh-Bénard convection problem
Non-intrusive data-driven models are promising class of methodologies to build surrogate models for challenging parametric engineering problems. POD-DL [1,2], a method from this class uses deep neural networks as the main data-driven ingredient. Following a first linear dimensionality reduction, based on a POD basis, it applies a second nonlinear dimension reduction, in this case a deep autoencoder, and a nonlinear regression using a forward neural network to account for the temporal and parametric coefficients. This setting generates a large number of hyperparameters, such as the number of modes of the POD basis, the number of layers and amount of neurons per layer for each network (the autoencoder one and the regression one), besides all typical hyperparameters of neural networks, as learning rate, batch size, etc. This all impacts the amount of time for the training process and the surrogate model accuracy.Based on our previous experience with this methodology for coupled transport-fluid problems, we decided to study the Rayleigh-Bénard convection problem [3] with a fixed Prandtl number, a well documented heat transfer problem coupled with fluid flow. Besides the temporal character of the problem, the Rayleigh number serves as a parameter, i.e., each different number generates a different dynamic that is learned by the surrogate model. Then, the regression neural network can predict the dynamics for an unseen Rayleigh number. In order to understand the contribution of each component of the method, we perform an ablation study in this controlled setup. [1] - M. Cracco et al., “Deep learning-based reduced-order methods for fast transient dynamics”, Arxiv Preprint 2212.07737, 2022.[2] - S. Fresca and A. Manzoni, “POD-DL-ROM: Enhancing deep learning-based reduced order models for nonlinear parametrized PDEs by proper orthogonal decomposition”, Comput. Methods Appl. Mech. Engrg., 2022.[3] - R. Kessler, “Nonlinear transition in three-dimensional convection”, Journal of Fluid Mechanics, 1987.
- 2 December 2024
- Published by Softaliza Tecnologias LTDA in Ibero-Latin American Congress on Computational Methods in Engineering (CILAMCE)
Open AccessAblation study of a non-intrusive data-driven surrogate model for the Rayleigh-Bénard convection problem
Non-intrusive data-driven models are promising class of methodologies to build surrogate models for challenging parametric engineering problems. POD-DL [1,2], a method from this class uses deep neural networks as the main data-driven ingredient. Following a first linear dimensionality reduction, based on a POD basis, it applies a second nonlinear dimension reduction, in this case a deep autoencoder, and a nonlinear regression using a forward neural network to account for the temporal and parametric coefficients. This setting generates a large number of hyperparameters, such as the number of modes of the POD basis, the number of layers and amount of neurons per layer for each network (the autoencoder one and the regression one), besides all typical hyperparameters of neural networks, as learning rate, batch size, etc. This all impacts the amount of time for the training process and the surrogate model accuracy.Based on our previous experience with this methodology for coupled transport-fluid problems, we decided to study the Rayleigh-Bénard convection problem [3] with a fixed Prandtl number, a well documented heat transfer problem coupled with fluid flow. Besides the temporal character of the problem, the Rayleigh number serves as a parameter, i.e., each different number generates a different dynamic that is learned by the surrogate model. Then, the regression neural network can predict the dynamics for an unseen Rayleigh number. In order to understand the contribution of each component of the method, we perform an ablation study in this controlled setup. [1] - M. Cracco et al., “Deep learning-based reduced-order methods for fast transient dynamics”, Arxiv Preprint 2212.07737, 2022.[2] - S. Fresca and A. Manzoni, “POD-DL-ROM: Enhancing deep learning-based reduced order models for nonlinear parametrized PDEs by proper orthogonal decomposition”, Comput. Methods Appl. Mech. Engrg., 2022.[3] - R. Kessler, “Nonlinear transition in three-dimensional convection”, Journal of Fluid Mechanics, 1987.
- Open Access
- 2 December 2024
- Published by Softaliza Tecnologias LTDA in Ibero-Latin American Congress on Computational Methods in Engineering (CILAMCE)
Improving accuracy of parametric surrogate model for turbidity currents using diffusion-based super-resolution
Turbidity currents are a sort of density-driven flow carrying particles that are generated between fluids with small density differences. They also are a mechanism responsible for the deposition of sediments on a seabed. A deep understanding of this phenomenon may help geologists on strategic knowledge in oil exploration. We simulate such currents using a stabilized finite element formulation in a Eulerian-Eulerian framework.This brings the challenge of a high computational cost for the evaluation of the high-fidelity model. We thus use a surrogate model composed of one linear (Proper Orthogonal Decomposition) and one non-linear (Autoencoder) reduction [1,2]. Once the surrogate is trained for a set of parameters with data generated by the high-fidelity model, one can use such a surrogate model for predicting the dynamics for unseen values of the parameter.Denoising Diffusion Probabilistic Models (DDPM) [3] is the method behind SOTA generative AI algorithms with very good performance in tasks like super-resolution. Based on such usage of diffusion for generative AI, the authors in [4] developed a framework for flow field reconstruction using a super-resolution task. Such a diffusion model is trained only using high-resolution data, without a pairing between low and high-resolution data, common in other techniques in the area.The present work aims to use the dataset of high-fidelity parametric simulations of turbidity currents for training both the surrogate and the diffusion model for the super-resolution task. The goal is: using the trained super-resolution model, improve the fields predicted by the reduced model for unseen points of the parametric space.[1] - M. Cracco et al., “Deep learning-based reduced-order methods for fast transient dynamics”, Arxiv Preprint 2212.07737, 2022.[2] - S. Fresca and A. Manzoni, “POD-DL-ROM: Enhancing deep learning-based reduced order models for nonlinear parametrized PDEs by proper orthogonal decomposition”, Comput. Methods Appl. Mech. Engrg., 2022.[3] - J. Ho, A. Jain and P. Abbeel, “Denoising Diffusion Probabilistic Models”, arXiv: 2006.11239, 2020.[4] - D. Shu, Z. Li, A. Barati Farimani, “A physics-informed diffusion model for high-fidelity flow field reconstruction”, Journal of Computational Physics, Volume 478, 2023
- 2 December 2024
- Published by Softaliza Tecnologias LTDA in Ibero-Latin American Congress on Computational Methods in Engineering (CILAMCE)
Open AccessImproving accuracy of parametric surrogate model for turbidity currents using diffusion-based super-resolution
Turbidity currents are a sort of density-driven flow carrying particles that are generated between fluids with small density differences. They also are a mechanism responsible for the deposition of sediments on a seabed. A deep understanding of this phenomenon may help geologists on strategic knowledge in oil exploration. We simulate such currents using a stabilized finite element formulation in a Eulerian-Eulerian framework.This brings the challenge of a high computational cost for the evaluation of the high-fidelity model. We thus use a surrogate model composed of one linear (Proper Orthogonal Decomposition) and one non-linear (Autoencoder) reduction [1,2]. Once the surrogate is trained for a set of parameters with data generated by the high-fidelity model, one can use such a surrogate model for predicting the dynamics for unseen values of the parameter.Denoising Diffusion Probabilistic Models (DDPM) [3] is the method behind SOTA generative AI algorithms with very good performance in tasks like super-resolution. Based on such usage of diffusion for generative AI, the authors in [4] developed a framework for flow field reconstruction using a super-resolution task. Such a diffusion model is trained only using high-resolution data, without a pairing between low and high-resolution data, common in other techniques in the area.The present work aims to use the dataset of high-fidelity parametric simulations of turbidity currents for training both the surrogate and the diffusion model for the super-resolution task. The goal is: using the trained super-resolution model, improve the fields predicted by the reduced model for unseen points of the parametric space.[1] - M. Cracco et al., “Deep learning-based reduced-order methods for fast transient dynamics”, Arxiv Preprint 2212.07737, 2022.[2] - S. Fresca and A. Manzoni, “POD-DL-ROM: Enhancing deep learning-based reduced order models for nonlinear parametrized PDEs by proper orthogonal decomposition”, Comput. Methods Appl. Mech. Engrg., 2022.[3] - J. Ho, A. Jain and P. Abbeel, “Denoising Diffusion Probabilistic Models”, arXiv: 2006.11239, 2020.[4] - D. Shu, Z. Li, A. Barati Farimani, “A physics-informed diffusion model for high-fidelity flow field reconstruction”, Journal of Computational Physics, Volume 478, 2023
- 6 March 2024
- Published by Springer Nature in Journal of the Brazilian Society of Mechanical Sciences and Engineering
- Vol. 46 (4) , 1-21
An embedded deep learning model discrepancy for computational combustion simulations
- 6 March 2024
- Published by Springer Nature in Journal of the Brazilian Society of Mechanical Sciences and Engineering
- Vol. 46 (4) , 1-21
An embedded deep learning model discrepancy for computational combustion simulations
- 15 November 2023
- Published by Wiley in International Journal for Numerical and Analytical Methods in Geomechanics
- Vol. 48 (2) , 588-606
A bi‐fidelity model for hydraulic fracturing
Hydraulic fracture, whether a natural phenomenon or an engineered process, consists of a highly pressurized flow evolving within a rock and, when necessary, fracturing such a medium. Particularly, in the context of oil and gas exploitation, a mixture of water and proppant is injected through a well to improve the oil extraction conditions by artificially increasing the local permeability. From a modeling standpoint, it is fundamental to acknowledge that this complex flow involves, among other disciplines, fracture mechanics, fluid dynamics, non‐Newtonian flow, solid mechanics, and flow in porous media. Therefore, explicitly or not, a model bears multiple scales and multiple physics, leading to a non‐linear moving boundary mathematical problem that is hard to solve through analytical techniques. Amongst them, the Pseudo‐3D, which computes fracture evolution in reservoirs confined by symmetric stress barriers, is frequently employed in the oil and gas industry domain. The different assumptions made to obtain the P3D model lead to inaccurate predictions, like the overestimation of fracture height for important operating conditions. We propose a model discrepancy term in a multi‐fidelity approach to correct such drawbacks. The main difficulty associated is constructing a mapping between the inputs and outputs that faithfully represents the error between high and low fidelity models. In this paper, we investigated the performance of a nonlinear multi‐fidelity approach relying on Gaussian Processes. We present a careful analysis to assess the good performance of such a proposition and also present results involving Uncertainty Quantification, which is made feasible by using the bi‐fidelity surrogate.
- 15 November 2023
- Published by Wiley in International Journal for Numerical and Analytical Methods in Geomechanics
- Vol. 48 (2) , 588-606
A bi‐fidelity model for hydraulic fracturing
Hydraulic fracture, whether a natural phenomenon or an engineered process, consists of a highly pressurized flow evolving within a rock and, when necessary, fracturing such a medium. Particularly, in the context of oil and gas exploitation, a mixture of water and proppant is injected through a well to improve the oil extraction conditions by artificially increasing the local permeability. From a modeling standpoint, it is fundamental to acknowledge that this complex flow involves, among other disciplines, fracture mechanics, fluid dynamics, non‐Newtonian flow, solid mechanics, and flow in porous media. Therefore, explicitly or not, a model bears multiple scales and multiple physics, leading to a non‐linear moving boundary mathematical problem that is hard to solve through analytical techniques. Amongst them, the Pseudo‐3D, which computes fracture evolution in reservoirs confined by symmetric stress barriers, is frequently employed in the oil and gas industry domain. The different assumptions made to obtain the P3D model lead to inaccurate predictions, like the overestimation of fracture height for important operating conditions. We propose a model discrepancy term in a multi‐fidelity approach to correct such drawbacks. The main difficulty associated is constructing a mapping between the inputs and outputs that faithfully represents the error between high and low fidelity models. In this paper, we investigated the performance of a nonlinear multi‐fidelity approach relying on Gaussian Processes. We present a careful analysis to assess the good performance of such a proposition and also present results involving Uncertainty Quantification, which is made feasible by using the bi‐fidelity surrogate.
- 1 October 2023
- Published by Elsevier in Journal of Applied Geophysics
- Vol. 218
Horizon extraction using absolute seismic phase obtained by unwrapping techniques
- 1 October 2023
- Published by Elsevier in Journal of Applied Geophysics
- Vol. 218
Horizon extraction using absolute seismic phase obtained by unwrapping techniques