Interpolación de Funciones de Flameo con Regresión Simbólica

Conocer adecuadamente los fenómenos que el viento produce en estructuras es un punto clave en la optimización de su diseño. Las funciones de flameo (flutter functions) describen la vibración del puente a una velocidad de viento concreta. Debido a que, en ciertos casos, la amortiguación puede ser negativa, puede dar lugar a un fallo completo del puente.

Normalmente se explora el espacio de diseño con simulaciones de dinámica de fluidos (CFD), pero su alto coste computacional hace a menudo inviable el análisis de todas las variaciones posibles en las variables de diseño. Una posible aproximación es utilizar Regression Simbólica para su interpolación en base a un subconjunto de simulaciones CFD, pudiendo dar lugar a nuevos descubrimientos en términos de interacción de fluidos.

La Regresión Simbólica es un tipo de análisis de regresión que busca en el espacio de expresiones matemáticas un modelo óptimo para un dataset concreto, en nuestro caso interpolación de funciones de flameo.

Interpolation of Flutter Functions with Symbolic Regression

Understanding the effects of wind on structures is a key point in optimizing their design. Flutter functions describe the vibration of a bridge at a specific wind speed. Since, in certain cases, damping can be negative, it may lead to a complete bridge failure.

The design space is typically explored using Computational Fluid Dynamics (CFD) simulations. However, their high computational cost often makes it unfeasible to analyze all possible variations in the design variables. One potential approach is to use Symbolic Regression for interpolation based on a subset of CFD simulations, which could lead to new insights into fluid interactions.

Symbolic Regression is a type of regression analysis that searches the space of mathematical expressions to find an optimal model for a specific dataset—in our case, interpolation of flutter functions.

https://m2lines.github.io/L96_demo/notebooks/symbolic_methods_comparison.html

https://astroautomata.com/PySR

Suite de benchmaking para a predición de fluxos en CFD

A falta de ferramentas estandarizadas de benchmarking no campo da dinámica de fluídos computacional (CFD) para a predición de fluxos presenta un reto constante para a investigación de novos métodos de predición. Actualmente, os modelos raramente se proban en conxuntos de datos comúns, e o código fonte a miúdo non se comparte, o que xera problemas de reproducibilidade.

Este proxecto propón a creación dunha suite de benchmarking completa que facilite a avaliación de novos modelos de predición de fluxos. A suite incluirá conxuntos de datos de CFD validados experimentalmente, ferramentas para preprocesamento de imaxes, pipelines de post-procesamento estandarizadas, e unha aplicación web en liña que proporcionará acceso ás ferramentas, aos conxuntos de datos e a un leaderboard para comparar o rendemento dos modelos.

CFD flow prediction benchmarking suite

The lack of standardized benchmarking tools in the field of computational fluid dynamics (CFD) flow prediction presents a persistent challenge for research into new prediction methods. Currently, models are rarely tested on common datasets, and source code is often not shared, leading to issues with reproducibility.

This project proposes the creation of a comprehensive benchmarking suite to facilitate the testing of new flow prediction models. The suite will include experimentally validated CFD datasets, tools for image preprocessing, standardized post-processing pipelines, and an online web application providing access to the tools, datasets, and a leaderboard for model performance comparisons.

https://fluidfoam.readthedocs.io

https://www.cambridge.org/core/journals/journal-of-fluid-mechanics/article/datadriven-prediction-of-unsteady-flow-over-a-circular-cylinder-using-deep-learning/0DB7E8CFF3BF3D5AB4C73A9F38150316

Análise de datos distribuído a gran escala e sobre a marcha con Ray

A fenda entre a capacidade de procesamento e o rendemento da E/S en contornas HPC non fixo máis que incrementarse nos últimos tempos. As grandes simulacións numéricas xeran cantidades de datos coas que é difícil lidiar cunha aproximación tradicional de análise baseada no posproceso, almacenando todos os datos que se desexan analizar durante a simulación para a seu posterior análise.

Neste traballo desenvolverase unha aproximación baseada en RAY para a análise distribuída e sobre a marcha dos datos producidos por unha simulación numérica paralela. RAY é un framework Python distribuído que permite traballar con referencias futuras e manter o estado dos procesos de análise. O traballo estenderá outro previo (ver material relacionado) e analizará como aproveitar algunhas das vantaxes doutro framework Python distribuído, Dask, para optimizar o uso de RAY para a análise sobre a marcha a gran escala.

Large-scale Distributed Data Analysis on the Fly with Ray

The gap between processing capacity and I/O performance in HPC environments has only increased in recent times. Large numerical simulations generate data volumes that are difficult to handle with a traditional post-processing analysis approach, which involves storing all the data to be analyzed during the simulation for subsequent analysis.

In this work, we will develop an approach based on Ray for the distributed and on-the-fly analysis of data produced by a parallel numerical simulation. Ray is a distributed Python framework that allows working with future references and maintaining the state of the analysis processes. The work will extend previous research (see related stuff and some pointers) and analyze how to leverage some advantages of another distributed Python framework, Dask, to optimize the use of Ray for large-scale on-the-fly analysis.

Xico Fernández, Bruno Raffin, Julien Bigot, Emilio J. Padrón. Análisis distribuido de datos sobre la marcha para simulaciones numéricas con Ray Jornadas Sarteco 2024 - XXXIV Jornadas de Paralelismo DOI: https://doi.org/10.5281/zenodo.12205022

Distributed on-the-fly Data Analysis for Parallel Numerical Simulations using Ray by Xico Fernández Lozano co-advised by Bruno Raffin (Inria Rhône-Alpes, Datamove team), Julien Bigot (Maison de la Simulation, France) and Emilio J. Padrón (UDC, CITIC).

-> pdf (en): https://gac.udc.es/~emilioj/project/archive/files/FernandezLozano_Xico_TFM_2023_en.pdf

-> pdf (gl): https://gac.udc.es/~emilioj/project/archive/files/FernandezLozano_Xico_TFM_2023.pdf

-> https://github.com/xicko7/reisa -> https://github.com/xicko7/reisa_tests

Redución de ruído (denoising) en render volumétrico

Direct Volume Rendering (DVR) con Volumetric Path Tracing (VPT) é unha técnica de visualización científica que simula o transporte de luz con modelos de iluminación físicos. Aínda que o trazado de camiños de Monte Carlo (MC) adoita empregarse en modelos de superficies, a súa aplicación en modelos volumétricos é complexa, debido á dificultade de integrar camiños de luz en medios volumétricos sen límites materiais claros ou suaves. Ademais, a alta presenza de ruído dificulta a efectividade dos denoisers en tempo real, que dependen de información libre de ruído para preservar os detalles da imaxe.

Nesta área acumulamos xa certa experiencia e temos algún traballo publicado, incluíndo o desenvolvemento de denoisers propios (ver traballo relacionado). Contamos, ademais, con varias ideas para mellorar o estado da arte, tanto en técnicas de redución de ruído en procesado offline como no deseño dun novo denoiser en tempo real.

Denoising for volume rendering

Direct Volume Rendering (DVR) with Volumetric Path Tracing (VPT) is a scientific visualization technique that simulates light transport using physically-based lighting models. Although Monte Carlo (MC) path tracing is commonly used for surface models, its application to volumetric models is complex due to the difficulty of integrating light paths in volumetric media without clear or smooth material boundaries. Additionally, high noise levels reduce the effectiveness of real-time denoisers, which rely on noise-free information to preserve image details.

In this area, we have already accumulated some experience and published related work, including the development of custom denoisers (see related stuff and some pointers). Furthermore, we have several ideas to advance the state of the art, both in offline noise reduction techniques and in designing a new real-time denoiser.

Real-Time Denoising of Volumetric Path Tracing for Direct Volume Rendering http://www.j4lley.com/content/publications/2020-ieee-tvcg-mcdvr-denoising

Neural James-Stein Combiner for Unbiased and Biased Renderings http://www.j4lley.com/content/publications/2022-sigasia-neuraljs

Protocolo de comunicacións compatible con OpenXR para sesións RV multiusuario nunha aplicación de visualización médica

CineM3D é unha ferramenta inmersiva de visualización 3D de datos médicos desenvolvida no noso grupo de investigación. Esta ferramenta incorpora varias técnicas que son estado da arte para a interacción nunha contorna de realidade virtual (RV) con modelos 3D realistas de render volumétrico obtidos a partir de CTs ou resonancias magnéticas. O obxectivo deste proxecto sería incorporar en CineM3D un protocolo de comunicación que permita traballar a varios usuarios nunha mesma contorna de realidade virtual e manexar coherentemente a visualización médica do modelo co que se estea a traballar.

Communication Protocol Compatible with OpenXR for Multi-User VR Sessions in a Medical Visualization Application

CineM3D is an immersive 3D visualization tool for medical data developed by our research group. This tool integrates several state-of-the-art techniques for interaction within a virtual reality (VR) environment using realistic 3D volumetric renderings from CT or MRI scans. The goal of this project is to incorporate a communication protocol into CineM3D that allows multiple users to collaborate within the same virtual reality environment, ensuring coherent and synchronized medical visualization of the model being analyzed.

https://cinem3d.citic.udc.es http://j4lley.com/cinem3d

https://docs.godotengine.org/en/stable/tutorials/networking/high_level_multiplayer.html

https://gafferongames.com/post/what_every_programmer_needs_to_know_about_game_networking

Reconstrución volumétrica 3D en realidade virtual con Gaussian splatting

A idea do proxecto sería traballar con Gaussian splatting nun contexto de visualización de imaxe médica (CTs, resonancias magnéticas, etc.) en Realidade Virtual. O alcance do proxecto a desenvolver non está aínda completamente definido, pero podería ser implementar un sistema do tipo descrito no recente traballo Multi-Layer Gaussian Splatting for Immersive Anatomy Visualization (arxiv).

3D Volumetric Reconstruction in Virtual Reality with Gaussian Splatting

The goal of this project is to work with Gaussian splatting in a medical imaging visualization context (CT, MRI, etc.) in Virtual Reality. The full scope of the project is not yet completely defined, but it could involve implementing a system similar to that described in the recent work Multi-Layer Gaussian Splatting for Immersive Anatomy Visualization (arxiv).

Simple description of Gaussian splatting and related methods: https://aws.amazon.com/blogs/spatial/3d-gaussian-splatting-performant-3d-scene-reconstruction-at-scale

3D Gaussian Splatting for Real-Time Radiance Field Rendering https://arxiv.org/abs/2308.04079

Multi-Layer Gaussian Splatting for Immersive Anatomy Visualization https://arxiv.org/abs/2410.16978

Detección automática de píxeles defectuosos en pantallas LED

Aplicación de técnicas de visión artificial e procesado de imaxe para detectar píxeles mortos en pantallas LED. Posibles extensións e/ou alternativas:

• deseño dun sistema low-cost de calibración de pantallas LED (visión artificial + aprendizase automática)
• aplicación para a xestión do envío de contidos ás pantallas LED (probablemente implique a programación a baixo nivelspecific dataset—ao nivel da API gráfica— para controlar o envío do sinal ao dispositivo que controla as pantallas, un controlador VX1000 no noso caso).

Automatic Detection of Defective Pixels on LED Screens

Application of computer vision and image processing techniques to detect dead pixels on LED screens. Potential extensions and/or alternatives:

• Design of a low-cost LED screen calibration system (computer vision + machine learning)
• Application for managing content delivery to LED screens (likely involving low-level programming—at the graphics API level—to control signal transmission to the screen controller, a VX1000 controller in our case).

Simulación de axentes en cadeira de rodas en vivendas reformadas

A idea, a grandes rasgos, sería analizar, mediante axentes (probablemente baseados en Aprendizaxe por reforzo), o grao de accesibilidade dunha vivenda para unha persoa con cadeira de rodas, antes e despois dunha reforma. Así, a partir da representación 3D dunha vivenda, no seu estado actual e coa reforma proposta, o simulador permitiría xerar datos de mobilidade dunha persoa cunha discapacidade física nesas contornas pechadas: zonas máis ou menos accesibles, mellores opcións para chegar a puntos estratéxicos da vivenda, puntos problemáticos que se resolveron coa reforma, etc. A análise destes datos permitiría analizar a reforma proposta, así como tomar decisións ao respecto de como reorganizar espazos nunha potencial nova iteración da reforma.

Simulation of Wheelchair Agents in Renovated Homes

The general idea is to analyze the accessibility of a home for a wheelchair user before and after a renovation, using agents (likely based on Reinforcement Learning). Starting from a 3D representation of the home in its current state and with the proposed renovation, the simulator would generate mobility data for a person with a physical disability in these enclosed environments. This data would identify more or less accessible areas, optimal paths to strategic points in the home, and problem areas resolved by the renovation, among other insights. Analyzing these data would allow for evaluating the proposed renovation and making decisions on how to reorganize spaces in a potential new iteration of the renovation.

Godot RL Agents https://github.com/edbeeching/godot_rl_agents https://edbeeching.github.io/papers/gdrl.html https://huggingface.co/learn/deep-rl-course/en/unitbonus3/godotrl

Visualización de datos volumétricos previamente segmentados con cinematic volume rendering (CTs, RM, neurociencia)

WiP

Visualization of Pre-Segmented Volumetric Data Using Cinematic Volume Rendering (CT, MRI, Neuroscience)

WiP

Cinematic rendering (by Siemens) https://www.siemens-healthineers.com/es/digital-health-solutions/cinematic-rendering

CineM3D: our cinematic rendering approach for Virtual Reality https://cinem3d.citic.udc.es

Optimization of Snow Simulations

WiP

3D to 2D indoor reconstruction

WiP

Optimización de simulacións de neve

WiP

Optimization of Snow Simulations

WiP

Prashant Goswami. Snow and Ice Animation Methods in Computer Graphics https://www.siemens-healthineers.com/es/digital-health-solutions/cinematic-rendering

Prashant Goswami. Webbinarium: Real-time Snow and Cloud Animation in Computer Graphics https://humanfactorsnetwork.se/aktivitet/webbinarium-prashant-goswami