323 research outputs found
linuswalter/WellPINN
No full textThis software computes the diffusion of fluid pressure p(x,y,t) in a 2D domain for a single injection well based on Physics Informed Neural Networks (PINN).The documentation is located in the README.md file.Pumping wells are a key component for modeling subsurface fluid flow, where their accurate representation is essential for reliable reservoir characterization through history matching of flow rates and well pressures, as well as for simulating operational scenarios. Physics-informed neural networks (PINNs) have recently emerged as a promising method for reservoir modeling, offering seamless integration of observational data and governing physical equations. However, existing PINN-based studies still face major challenges in capturing fluid pressure inference near pumping wells, particularly in the early phase after injection start. In response, we introduce 'WellPINN', a modeling workflow that combines the outputs of multiple sequentially trained PINN models. The first model covers the entire reservoir, while the size of each subsequently trained model equals the equivalent well radius of the previously trained model. This workflow allows us to iteratively approximate the radius of the equivalent pumping well to the actual pumping well dimensions.
Applied to a single pumping well in a two-dimensional domain, our forward model demonstrates how combining three PINNs can bridge a spatial scale spanning three orders of magnitude, from the reservoir boundary down to the well radius. The results showcase that our sequential training of superimposing networks around the pumping well is the first workflow that enables accurate inference of fluid pressure from pumping rates across the entire injection time, significantly advancing the potential of PINNs for inverse modeling and operational scenario simulations. Future research might enhance 'WellPINN' by employing it to an inverse problem. Additionally, further investigation could test its effectiveness for heterogeneous domains and multiple injection wells.LW acknowledges funding from the European Research Council (ERC) under the European Union’s Horizon 2020 Research and Innovation Program through the Starting Grant GEoREST (www.georest.eu) under Grant Agreement No. 801809.Peer reviewe
linuswalter/PINN-for-Subsurface-Flow
No full textThis software computes the diffusion of fluid pressure p(x,t) in a 1D domain based on the concept of Physics Informed Neural Networks (PINN) for a heterogeneous modeling domain.GNU General Public License 2 or later (GPL-2.0)Field-data assimilation to calibrate rock properties in numerical physics-based reservoir models of hydrogeological applications is challenging. Recently, Artificial Neural Networks (ANN) have emerged as a promising alternative to handle noisy data seamlessly. However, in hydrogeology, even though data are abundant over time, observation wells are sparse over space, which results in insufficient data to train ANN. Here, we propose Physics-Informed Neural Networks (PINN) to bridge the gap of sparse spatial data by imposing physical conditions. We test ANN and PINN on a synthetic dataset for fluid pressure diffusion p(x,t) through a low-permeable porous medium that hosts a high-permeability equivalent fracture material. By comparing the ANN and the PINN as a function of the number of observation wells, we find that the PINN model outperforms the ANN when having less than 14 wells. Adding noise to the training data reveals the advantage of PINN to be more robust for random measurement errors or ambient noise. We finally test the applicability and limitations of PINN to represent fractures as a thin equivalent material while having no observations inside the domain and find that accuracy can be maintained when reducing the thickness of the equivalent material at the expense of increasing the computation time. Given that hydrogeological applications count with a limited number of observation wells, PINN appears as a more suitable machine learning tool than purely data-driven ANN for reservoir modeling. Therefore, we consider this work a starting point for developing more realistic reservoir models with heterogeneous material distribution based on the PINN architecture.LW acknowledges funding from the European Research Council (ERC) under the European Union’s Horizon 2020 Research and Innovation Program through the Starting Grant GEoREST (www.georest.eu) under Grant Agreement No. 801809.Peer reviewe
linuswalter/Real-Time-Forecast-IBDP
No full text[Short description what the software is about] This software computes a forecast for the target variables seismic event count and maximum seimsic moment magnitude based on operational data (fluid pressure, flow rate) and a seismic catalog based on a random forest machine learning algorithm.[Links or references to publications or other documentation] https://github.com/linuswalter/Real-Time-Forecast-IBDPGeological carbon storage (GCS) can significantly reduce emissions from hard-to-abate industries. However, large-scale deployment is hindered by the risk of induced seismicity, which has led to multiple project shutdowns. Mitigating such events would require unprecedented reliable seismicity forecasting. We present an adaptive near-real-time workflow that forecasts seismic variables. Using a random forest trained on sequentially partitioned data, we forecast seismicity rate and maximum magnitude.
Applied to the Illinois Basin Decatur Project, the model receives operational parameters and the seismic catalog as input.
Our results identify well pressure, past event counts, and prior maximum magnitudes as key predictors.
However, accuracy declines over time as fluid pressure diffuses into the far-field, highlighting the need to incorporate spatiotemporal pressure diffusion for improved long-term forecasts.
Our framework is the first to offer near-real-time forecasting from the operation start with uncertainty quantification, providing a foundation for next-generation seismic hazard mitigation in geo-energy operations.LW acknowledges funding from the European Research Council (ERC) under the European Union’s Horizon 2020 Research and Innovation Program through the Starting Grant GEoREST (www.georest.eu) under Grant Agreement No. 801809.File List: 01_IBDP_RF_Model.py 02_IBDP_RF_Postproc.pyPeer reviewe
Effects of pendent phenol functional groups on secondary coordination spheres of heme like Fe-salen complexes
No full textSince the beginning of industrial revolution, burning of fossil fuels has mainly led to increase in atmospheric concentration of CO2 , a Green House Gas (GHG), from 250 ppm to 400 ppm between 1800 and 2012. One way to reduce the burning of fossil fuels and CO2 emission rate is to explore alternative carbon free fuels to meet the energy demand.This project aims at the synthesis and study of metal complexes inspired by biological models that will help better design catalysts to perform water oxidation more effectively.This poster won the Dean, Faculty of Science award (2020). Advisor: Dr. Linus Chiang, Departmen of Chemistry
21st-century scholarship and Wikipedia
No full textWikipedia, the world’s fifth most-used Web site, is a good illustration of the growing credibility of online resources. In his article in Ariadne earlier this year, “Wikipedia: Reflections on Use and Academic Acceptance”, Brian Whalley described the debates around accuracy and review, in the context of geology. He concluded that ‘If Wikipedia is the first port of call, as it already seems to be, for information requirement traffic, then there is a commitment to build on Open Educational Resources (OERs) of various kinds and improve their quality.’ In a similar approach to the Geological Society event that Whalley describes, Sarah Fahmy of JISC worked with Wikimedia and the British Library on a World War One (WWI) Editathon. There is a rich discourse about the way that academics relate to Wikipedia
Towards seamless integration of data and physics: machine learning workflows for subsurface flow and induced seismicity forecasting in geo-energy applications
No full text(English) Geo-energy applications, such as geothermal energy, geologic carbon capture and storage, and underground hydrogen storage, have the potential to address anthropogenic climate change by contributing to the decarbonization of the energy sector and in the so-called hard-to-abate-industries. However, the large-scale expansion of this technology is hampered by the problem of felt induced seismicity.
This poses a key technical and social challenge, as current mitigation protocols such as the Traffic Light System (TLS) have led to several severe failures and project cancellations. These setbacks underscore the need for more adaptive and physics-based approaches in the field of seismic hazard control.
This thesis addresses this challenge by developing machine learning workflows that combine physics-informed modeling and data-driven seismicity forecasting. The following key contributions of the thesis were derived from this overarching objective.
The conducted literature review analyses the existing mitigation schemes for seismic hazard and proposes a conceptual shift away from reactive TLS protocols toward adaptive, model-driven forecasting, and operational control. For this purpose, this chapter discusses the latest developments in machine learning with respect to real-time data processing, continuous reservoir characterization through physics-based surrogate modeling, and control theory.
The first modeling study in the thesis compares the modeling performance of data-driven versus physics-informed neural networks for fluid pressure diffusion in a heterogeneous porous medium. The PINN outperforms ANN for sparse and noisy spatial observations, offering a robust solution for hydrogeological contexts where data are limited. Subsequently, the second modeling study introduces WellPINN, a workflow based on physics-informed neural networks that accurately represents a pumping well on multiple spatial scales in a 2D reservoir domain. This model combines sequentially trained subnetworks to simulate pressure diffusion from the well to the reservoir boundary, laying the foundation for future inverse modeling of reservoir permeability.
The final work within this thesis presents a novel workflow for real-time forecasting of induced seismicity based on a data-driven machine learning approach. It forecasts seismic event rates and maximum magnitudes starting from early injection times, while providing performance insights by tracking feature importance over time. We apply our workflow to the seismic catalog of the Illinois Basin Decatur carbon storage site.
In general, this thesis addresses induced seismicity from two complementary angles: advances in physics-based reservoir modeling with PINNs and the development of adaptive, data-driven forecasting workflows. By presenting a novel workflow based on a literature review and by developing new methods for combining physics-based reservoir modeling and data-driven seismicity forecasting, this thesis contributes to the advancement of the next generation of mitigation frameworks for the control of induced seismic hazard.(Català) Les aplicacions geo-energètiques, com l'energia geotèrmica, la captura i emmagatzematge de carboni geològic, i l'emmagatzematge subterrani d'hidrogen, tenen el potencial d'abordar el canvi climàtic antropogènic contribuint a la descarbonització del sector energètic i en les anomenades indústries difícils de matar. No obstant això, l'expansió a gran escala d'aquesta tecnologia es veu obstaculitzada pel problema de la sismicitat induïda pel sentiment.
Això planteja un repte tècnic i social clau, ja que els actuals protocols de mitigació com el Sistema Lleuger de Trànsit (TLS) han provocat diverses fallades greus i cancel·lacions de projectes. Aquests contratemps subratllen la necessitat d'aproximacions més adaptatives i basades en la física en el camp del control de riscos sísmics.
Aquesta tesi aborda aquest repte mitjançant el desenvolupament de fluxos de treball d'aprenentatge automàtic que combinen la modelització informada per la física i la predicció de la sismicitat basada en dades. Les següents contribucions clau de la tesi es van derivar d'aquest objectiu general.
La revisió de la literatura realitzada analitza els esquemes de mitigació existents per al risc sísmic i proposa un canvi conceptual lluny dels protocols TLS reactius cap a la previsió adaptativa, basada en models i el control operatiu. Amb aquesta finalitat, aquest capítol analitza els últims desenvolupaments en l'aprenentatge automàtic respecte al processament de dades en temps real, la caracterització contínua de reservoris a través de la modelització de substituts basada en la física i la teoria de control.
El primer estudi de modelatge de la tesi compara el rendiment de modelatge de xarxes neuronals basades en dades versus xarxes neuronals informades per la física per a la difusió de pressió de fluids en un medi porós heterogeni. El PINN supera l'ANN per a observacions espacials disperses i sorolloses, oferint una solució robusta per a contextos hidrogeològics on les dades són limitades. Posteriorment, el segon estudi de modelatge introdueix WellPINN, un flux de treball basat en xarxes neuronals informades per la física que representa amb precisió un pou de bombament en múltiples escales espacials en un domini de reservori 2D. Aquest model combina subxarxes entrenades seqüencialment per simular la difusió de la pressió des del pou fins a la frontera del reservori, establint les bases per a la futura modelització inversa de la permeabilitat del reservori.
El treball final d'aquesta tesi presenta un nou flux de treball per a la predicció en temps real de la sismicitat induïda basat en un enfocament d'aprenentatge automàtic basat en dades. Pronostica taxes sísmiques d'esdeveniments i magnituds màximes a partir dels primers temps d'injecció, alhora que proporciona informació de rendiment mitjançant el seguiment de la importància de les característiques al llarg del temps. Apliquem el nostre flux de treball al catàleg sísmic del lloc d'emmagatzematge de carboni Decatur de la Conca d'Illinois.
En general, aquesta tesi aborda la sismicitat induïda des de dos angles complementaris: els avenços en la modelització de reservoris basada en la física amb PINN i el desenvolupament de fluxos de treball de predicció adaptatius i basats en dades. Mitjançant la presentació d'un nou flux de treball basat en una revisió de la literatura i el desenvolupament de nous mètodes per combinar la modelització de reservoris basada en la física i la predicció de la sismicitat impulsada per dades, aquesta tesi contribueix a l'avanç de la propera generació de marcs de mitigació per al control del risc sísmic induït.(Español) Las aplicaciones geoenergéticas, como la energía geotérmica, la captura y almacenamiento geológico de carbono y el almacenamiento subterráneo de hidrógeno, tienen el potencial de hacer frente al cambio climático antropogénico al contribuir a la descarbonización del sector energético y de las denominadas «industrias difíciles de reducir». Sin embargo, la expansión a gran escala de esta tecnología se ve obstaculizada por el problema de la sismicidad inducida. Esto plantea un reto técnico y social clave, ya que los protocolos de mitigación actuales, como el sistema de semáforos (TLS), han dado lugar a varios fallos graves y a la cancelación de proyectos. Estos contratiempos ponen de relieve la necesidad de enfoques más adaptables y basados en la física en el campo del control de los riesgos sísmicos. Esta tesis aborda este reto mediante el desarrollo de flujos de trabajo de aprendizaje automático que combinan la modelización basada en la física y la predicción de la sismicidad basada en datos. Las siguientes contribuciones clave de la tesis se derivan de este objetivo general. La revisión de la literatura presenta un cambio conceptual que se aleja de los protocolos TLS reactivos hacia la predicción adaptativa basada en modelos y el control operativo. Con este fin, este capítulo analiza los últimos avances en aprendizaje automático con respecto al procesamiento de datos en tiempo real, la caracterización continua de yacimientos mediante modelos sustitutivos basados en la física y la teoría de control. El primer estudio de modelización de la tesis compara el rendimiento de la modelización de redes neuronales basadas en datos frente a redes neuronales informadas por la física para la difusión de la presión de fluidos en un medio poroso heterogéneo. La PINN supera a la ANN en observaciones espaciales dispersas y ruidosas, ofreciendo una solución robusta para contextos hidrogeológicos en los que los datos son limitados. Posteriormente, el segundo estudio presenta WellPINN, un flujo de trabajo basado en redes neuronales informadas por la física que representa con precisión un pozo de bombeo en múltiples escalas espaciales en un dominio de yacimiento 2D. Este modelo combina subredes entrenadas secuencialmente para simular la difusión de la presión desde el pozo hasta el límite del yacimiento, sentando las bases para el futuro modelado inverso de la permeabilidad del yacimiento. El trabajo final de esta tesis presenta un novedoso flujo de trabajo para la predicción en tiempo real de la sismicidad inducida basado en un enfoque de aprendizaje automático basado en datos. Predice las tasas de eventos sísmicos y las magnitudes máximas a partir de los primeros momentos de la inyección, al tiempo que proporciona información sobre el rendimiento mediante el seguimiento de la importancia de las características a lo largo del tiempo. Aplicamos nuestro flujo de trabajo al catálogo sísmico del yacimiento de almacenamiento de carbono de Decatur, en la cuenca de Illinois. En general, esta tesis aborda la sismicidad inducida desde dos ángulos complementarios: los avances en el modelado de yacimientos basado en la física con PINN y el desarrollo de flujos de trabajo de predicción adaptativos y basados en datos. Al presentar un novedoso flujo de trabajo basado en una revisión de la literatura y desarrollar nuevos métodos para combinar el modelado de yacimientos basado en la física y la predicción de sismicidad basada en datos, esta tesis contribuye al avance de la próxima generación de marcos de mitigación para el control del riesgo sísmico inducido.DOCTORAT EN ENGINYERIA CIVIL (Pla 2012
Walter Benjamin (1936) “The Work of Art in the Age of Mechanical Reproduction”
No full textThis chapter delves into a pivotal question of media studies: “How do media impact us?" It examines how media technology moulds perception, experience, and cultural consumption. At the core of this exploration lies Walter Benjamin’s essay “The Work of Art in the Age of Mechanical Reproduction.” Written in the early 20th century, the essay primarily addresses archaic media technologies like silent film and photography. Yet, it offers an analysis that probes the dynamics between original and copy, the essence of perceiving an object as art, and the intersections of technology, culture, and politics - themes that endure and resurface with every introduction of new media. © 2024 selection and editorial matter, Stina Bengtsson, Staffan Ericson and Fredrik Stiernstedt; individual chapters, the contributors.</p
How Many Answers Are Enough? Optimal Number of Answers for Q&A Sites
No full textWith the proliferation of the social web, questions about information quality and optimization attract the attention of IS scholars. Question-answering (QA) sites, such as Yahoo!Answers, have the potential to produce good answers, but at the same time not all answers are good and not all QA sites are alike. When organizations design and plan for the integration of question answering services on their sites, identification of good answers and process optimization become critical. Arguing that ‘given enough answers all questions are answered successfully,’ this paper identifies the optimal number of posts that generate high quality answers. Based on content analysis of Yahoo! Answers’ informational questions (n=174) and their answers (n=1,023), the study found that seven answers per question are ‘enough’ to provide a good answer
Power Relations and Social Classes in Pengakuan Pariyem by Linus Suryadi AG: Reflection of Masculine Ideology
No full textThe study aims to explore power relations and social classes as the reflections of the masculine ideology of the author in the novel Pengakuan Pariyem by Linus Suryadi AG. The theories implemented in the study are van Dijk’s power relations and social classes theory and Connell’s masculinity theory. The study is qualitative descriptive and applies the Critical Discourse Analysis (CDA) method, used to dismantle the ideology that is produced and reproduced through the language within the novel. The research data are lingual units that indicate power relations and social classes which simultaneously reflect the notion of masculinity. The results of the study are as follows. First, Pariyem as the central character in the novel lives within a hierarchical and dualistic Javanese society. Her submission as the babu (housemaid) of a priyayi (noble) family does not only lead Pariyem to be dominated symbolically, but also legitimizes the priyayi (aristocrats) power over wong cilik (commoners). It is reinforced by the representations of the priyayis’ world views in terms of culture, aristocracy, bureaucracy, and education orientation. It shows that priyayis are culturally dominant. Secondly, since Pariyem is a character created by a male author, her behaviors and actions reflect the ideology of masculinity. Rather than voicing women, the power relations that Pariyem experiences through the events constructed in the novel show that she embodies the masculine ideology, or masculinity.
- …
