DSpace@RPI (Rensselaer Polytechnic Institute)
Not a member yet
6809 research outputs found
Sort by
Customizable Knowledge Graph Visualization using the Whyis Knowledge Explorer
No full textNetwork visualization over large knowledge graphs suffers from multiple challenges: graphs have varying and sometimes multiple ways to represent what people expect a "link" to be - everything from direct triples to complex chemical interactions, social constructs, and OWL property restrictions can be considered a link. Additionally, large knowledge graphs cannot be usefully visualized as a whole because they are simply too large and complex, an any patterns are lost in the noise when there is enough computational ability to represent them. The Whyis Knowledge Explorer is a component of the Whyis knowledge graph development framework that addresses these issues. It allows for fast, customizable network visualization of large scale knowledge graphs. By providing a "starting point" with any specific node, users can explore the graph piece by piece, building a view up by expanding selected nodes on demand, making it easier to explore locally. By using “data views”, the component provides a consistent user interface over a wide range of entity types that can handle both simple and complex relationships between entities. These data views publish a consistent output from multiple templates and can be extended through plugins as well as by the implementing Knowledge Graph App (KGApp). Entity types can also be assigned custom styles through CSS using Cytoscape.js styling. Additionally, links can be qualified with certainty values, showing more probable links as having greater weight. We also use the same interface to provide a summary view of the knowledge graph by automatically generating concept maps of instantiated types, allowing users to see and explore overall usage patterns in the knowledge graph, highlighting both intended design and knowledge curation issues. This component has been a key part of many Whyis-based projects and is mature and scalable
Essays on subjective time in strategy
No full textAugust 2024School of ManagementTime is inherently important for strategy research. It could be experienced by firms in subjectively different and heterogenous ways, such as temporal focus, time horizon, temporal depth, etc. Although the subjective time and its antecedents and consequences have attracted more and more attention, this literature can be further enriched theoretically and empirically both in terms of temporal concepts studied and phenomena explained. In terms of temporal concepts, prior studies about subjective time cast almost exclusive attention to general temporal orientations that manifest in firms’ daily operations and regard them as “persistent and relatively stable”. Strategy scholars have not considered much about subjective time's dynamic features with the change of the environment. To fill this gap, we study how firms construct and perceive external shocks using the COVID-19 pandemic as the context. Results indicate in our sample, 76.7% firms perceived the pandemic to be a relatively long-term event with little change in these perceptions over time. In contrast, 23.3% firms perceived the pandemic to be relatively short term, while revising these perceptions to long term by the end of the observation period. Drawing on two notions in time related research—temporal structures and temporal work—we find firms’ initial time horizons and slack resources influence temporal duration perceptions. Our research highlights how shocks are subjectively perceived and construed by firms including in terms of their duration, while drawing attention to temporal sense-making processes within firms.
In terms of research contexts, prior studies mainly focus on single firms’ subjective time and its antecedents or consequences but overlook multiple firms’ subjective time interaction. Aiming at this gap, we explored the role of subjective time alignment between the acquirer and the target in their merger and acquisition processes. Moving beyond the traditional resource-fit perspective, this study suggests that cognition fit in temporal terms between acquirers and targets is a fundamental factor driving acquisitions, since it affects whether the two firms have similar thinking mode and coordination potential. Accordingly, we propose that certain level of temporal focus cognition alignment needs to be met to ensure smooth collaboration and coordination, thereby allowing the positive impacts of other factors on M&A performance to materialize. Using a sample of US public M&A transactions and adopting choice model and Necessary Condition Analysis, our study finds: (1) Firms with different temporal focuses are less likely to merge; (2) Acquirers and targets with different temporal focuses require more time to complete their transactions; (3) Having a similar temporal focus between the acquirer and the target is necessary (although not sufficient) for achieving high M&A performance. By introducing a necessity-based element to the cognitive fit literature in M&As, our research offers new insights into the intricate interplay between resource fit, cognition fit, and M&A efficiency.Ph
Can AI have common sense? Finding out will be key to achieving machine intelligence
No full textSince their public release less than two years ago, large language models (LLMs) such as those that underlie ChatGPT have unleashed exciting and provocative progress in machine intelligence. Some researchers and commentators have speculated that these tools could represent a decisive step towards machines that demonstrate ‘artificial general intelligence’ — the range of abilities associated with human intelligence — thereby fulfilling a 70-year quest in artificial-intelligence (AI) research
Graph neural networks for power grid stability
No full textAugust 2024School of EngineeringThe electric power grid is a very large and complicated system comprised of numerous machines and interconnections. Due to the growing complexity of these connections and increased penetration of renewable energy sources, maintaining stability is becoming an increasingly difficult problem. In recent years, data driven control methods are gaining more attention due to their versatility. Graph neural networks however, offer added stability, computational efficiency and robustness by exploiting the underlying graphical nature of the power grid. This thesis shows applications of graph neural networks to power system analysis. The first application considers the power system security assessment problem. Exploiting the structure of the grid, a classifier to trained to classify safe and unsafe states using a very small network and more efficient training. The classifier is also shown to perform efficiently in cases of limited observability, with missing data at some of the buses in the network. Case studies are shown on the IEEE 68-bus system and the NPCC 140-bus power system. The second application considers under-frequency load shedding using reinforcement, with graph networks improving the speed and computational efficiency of the training process. Results are shown on the IEEE 68-bus power system.Ph
Towards efficient multi-disciplinary optimization of electric aircraft motors
No full textMay 2024School of EngineeringThe use of numerical optimization techniques is becoming an instrumental tool used in the design of new aircraft. As the aviation industry explores novel configurations of electrically powered aircraft, new multi-disciplinary analysis tools will be needed that are able to analyze and optimize candidate designs. In particular, the electric motors used in these new aircraft's propulsion systems require special attention: the complex interaction between the motor's electromagnetic and thermal behavior can have significant impacts on the motor's performance and operational safety. Simply put, if a motor gets too hot, its performance will degrade and it may even fail catastrophically. Further, additional system efficiency gains can be found if the motor is designed simultaneously with the rest of the aircraft. In this thesis, I introduce the first high-fidelity fully-coupled analytically-differentiated electro-thermal motor analysis and optimization framework. The framework efficiently computes derivatives through the multi-disciplinary forward analysis with the coupled adjoint method. These efficient derivative computations enable the use of scalable gradient-based optimization methods for electric motor problems, all while accurately capturing the coupled electro-thermal physics. Next, I use this electro-thermal motor optimization framework to explore the importance of modeling the electro-thermal physics with a fully coupled feedback model.I perform a multi-objective electric motor optimization study that seeks to minimize motor mass and maximize efficiency, constrained by motor power and maximum motor winding and magnet temperatures.
In the study, I consider a feedback-coupled model as well as two feedforward-coupled models; each using a different reference temperature when computing the electromagnetic performance.
The results of this optimization study show that feedback coupling is indeed important to model, particularly for cases when it is not obvious to a practitioner how they should trade motor mass and efficiency. Subsequently, I further demonstrate the merit of the coupled electro-thermal motor analysis framework by including it in a conceptual aircraft design study. This study combines the analytically differentiated electric motor model with models for an aircraft's inverter, gearbox, and propeller, and explores how the aircraft's system-level parameters (e.g. cruise velocity) impact its maximum range. This study illustrates how an engineer may use analytically differentiated analysis tools in a conceptual design framework, as it allows them to perform high-level trade studies of optimal designs. Finally, the last part of this thesis introduces a novel multi-fidelity optimization algorithm. While we as engineers often want to be able to use the highest fidelity analysis models available, such models are often cost-prohibitive to use early in the design process when several different concepts and configurations may be considered. To that end, I have developed a multi-fidelity optimization algorithm that uses the estimated error between the low- and high-fidelity analyses to efficiently globalize the optimization. I compare the algorithm to existing state-of-the-art methods on a series of benchmark problems where I show that it performs favorably.Ph
Designing two-dimensional materials with novel spin degrees of freedom
No full textAugust 2024School of ScienceThe central aim of this thesis is to explore the interplay between reduced dimensions in two-dimensional (2D) materials and magnetism. How reduced dimensionality of 2D materials affects their spin properties is an active area of research. In this thesis, I explore this topic from the perspective of materials discovery. That is, I investigate which 2D materials exhibit various magnetic states and propose strategies for the accelerated discovery of novel 2D magnets with desirable properties. In the Chapter 3, the magnetic and thermodynamic properties of transition metal dichalcogenides of the form A2X4, based on monolayer Mn2Se4, are investigated using data analytics. In particular, by combining first-principles calculations with machine learning methods the microscopic origins of the magnetocrystalline anisotropy in these materials are elucidated. A large number of candidate transition metal dichalcogenides are explored by varying the chemical compositions of the transition metal (A) sites and the chalcogen (X) sites. The magnetocrystalline anisotropy is investigated by studying the transition between in-plane and out-of-plane magnetization. Using data analytics, we demonstrate that the interplay between the spin-orbit interactions of the chalcogen and transition metal atoms can impact the magnetic behavior. Finally, this investigation resulted in the identification of several novel transition metal dichalcogenides with large anisotropies that are chemically stable. In Chapter 4, magnetic ordering in two-dimensional (2D) materials is investigated using state-of-the-art machine learning models that use a graph-theory framework. A method for predicting the ground state collinear ordering of 2D magnets using machine learning is presented. We find that representing materials as graphs allows us to better learn structure-property relationships by leveraging both the chemical properties of the constituent atoms and the connectivity between those atoms. Graph neural network models are capable of predicting global properties of crystal structure (i.e. graph-wise properties) and local properties of the constituent atoms (i.e. node-wise properties). Physical constraints are embeded into the model by simultaneously making predictions of local and global properties. In particular, the Atomistic Line Graph Neural Network (ALIGNN) architecture is used.
The ALIGNN model is trained on data comprising local and global magnetic moments of 314 2D structures of the form CrAiiBiBiiX6, based on monolayer Cr2Ge2Te6, calculated from first-principles.
By learning the relationships between both local and global magnetic properties, we demonstrate an improvement over models that only consider global magnetic properties. In Chapter 5, noncollinear spin configurations called spin textures are investigated in magnetic materials that break inversion symmetry. More specifically, Janus monolayers are explored for their potential to host skyrmions, a class of topologically protected magnetic textures. High-throughput first-principles calculations are used to screen for Janus monolayers that exhibit large chiral interactions, which can stabilize skyrmions. In addition, the thermodynamic behavior of spin textures in promising candidates are investigated using Monte Carlo (MC) simulations. Further to this, in Chapter 6, a machine learning framework for predicting the Heisenberg model parameters of magnetic materials, that is, the parameters that define the magnetic interactions in the material, is presented. This framework uses equivariant neural networks, a type of symmetry-aware machine learning model. The advantages of using these types of models is discussed, as well as a model architecture that enforces the symmetry rules of magnetic interactions. Preliminary results for a model that predicts the strength of isotropic magnetic interactions trained on the materials investigated in Chapter 5 are presented.Ph
Evaluating virtual levee failures: a tool for demonstrating engineering judgment
No full textMay 2024School of EngineeringThis thesis explores an innovative way to develop engineering judgment in undergraduates through the use of virtual simulations, specifically focusing on levee failure scenarios. The Levee Inspection module of the larger GeoExplorer virtual environment (game) is under development for widespread use in undergraduate geotechnical engineering courses and is designed to simulate real-world scenarios where engineers face challenges and decision-making processes encountered in practice. The development and effectiveness of the Levee Inspection module are evaluated through two distinct case studies involving two different classes at Rensselaer Polytechnic Institute (RPI): the Foundation Engineering course and the Introduction to Geotechnical Engineering course. The design and implementation of the Levee Inspection module are informed by these case studies, drawing parallels to the existing Cone Penetration Test (CPT) module in terms of scalability and instructional design. This thesis refrains from making broad claims about the module’s efficacy, focusing instead on specific, measurable outcomes and the insights gained from two implementations. The findings from this research highlight the potential of virtual simulation (game-based learning) tools like the Levee Inspection module in the development of engineering judgment among undergraduate students, providing them with a safe, controlled environment to explore, observe, and document the complexities of levee systems and their failure mechanisms. This thesis, based on case studies conducted at RPI, showcases the practical use of the Levee Inspection module in educational settings, while the insights from the CPT module's implementation across US campuses inform the potential scalability and applicability of the Levee Inspection module in similar courses nationwide.M
On connections between mean field games and deep generative models
No full textMarch 2024School of ScienceDeep generative models have exploded in popularity and made frequent headline appearances in the past few years. From generating photo-realistic images and contest-winning artworks with diffusion models, to discovering millions of promising new materials with graph neural networks, and finally training large language models such as GPT-4 that exhibits glimmers of artificial general intelligence. The field of generative modeling has seen unprecedented levels of accelerating growth in its application and is rapidly transforming various aspects of everyday life. Nevertheless, there remains much need for a systematic theoretical understanding of generative approaches. While we understand each model well in isolation, it is generally difficult to compare different methods as they are derived from distinct motivating principles. Hence, we need to contextualize different generative approaches under an unifying framework to make them commensurable. Mean-field games (MFG), a versatile framework for modeling density evolution under customizable preferences, has emerged as a promising candidate for this purpose. In this thesis, we will explore and formalize the connection between MFG and normalizing flows, a prominent family of generative models, then outline extensions to other methods such as diffusion models. With this insight, we introduce transport costs to regularize NF optimization and demonstrate its effectiveness at controlling the Lipschtiz constant of the trained flows. On the other hand, MFG is also a powerful modeling tool that finds application in game theory, economics, finance, and industrial planning. Devising algorithmic solutions for MFG is thus interesting in its own right. With a bridge between MFG and generative modeling, we harness advancements in scalable and expressive neural architectures to solve high-dimensional MFG accurately and efficiently, a case that is especially challenging for classic optimization techniques. Starting with solving single instance MFG with flexible flow parametrizations, we then take it one step further to learn mappings that outputs optimal trajectories for distinct MFGs without re-training. Our proposed approach leverages attention-based layers to build sampling-invariant parametrizations for continuous operators and is the pioneering work for the unsupervised learning of high dimensional MFG solution maps. Finally, we study the inverse MFG problem and develop the first learning based framework for its solution. Our approach leverages bilevel optimization to simultaneously infer optimal agent trajectories and the unseen obstacle in the MFG setup. Our method is robust across various complexity levels and serves as an effective regularization for trajectory likelihood estimation in data-scarce scenarios.Ph
Model-based guidance and control of tail-sitter transitioning unmanned aerial systems
No full textMay 2024School of EngineeringTail-sitter unmanned aerial systems (UAS) are vehicles that are capable of operating in and switching between the vertical take-off and landing (VTOL) and fixed-wing flight regimes through a 90 degree rigid body rotation. This capability lends tail-sitter UAS several unique characteristics and abilities that allow them to outperform pure VTOL or fixed-wing UAS in a variety of civil and military applications. The key design challenge regarding achieving autonomy for tail-sitter UAS lies in the characterization of the transition flight regime, or the flight regime between VTOL and fixed-wing flight. The complexity of the transition regime makes it particularly difficult to design unified guidance (path-planning) and control architectures for tail-sitter UAS that are applicable to all flight regimes (VTOL, fixed-wing, and transition), with typical approaches to control relying on either avoiding or overpowering the aerodynamics of the transition regime. This can result in overly conservative controllers capable of only stabilizing a tail-sitter under very specific conditions. Furthermore, guidance methods for tail-sitter UAS rarely consider the aerodynamics of the transition regime, often relying on heuristic or geometric methods for generating state trajectories for various flight missions. To address these challenges, this thesis aims to develop a unified approach to a guidance and control for tail-sitter UAS autonomy. Specific considerations are given to (1) tail-sitter path-planning methodologies that rely on transition model-based optimization techniques used to generate missions that can fulfill a specific objective (specified by the user) as well as aerodynamic state estimations for the wings that are useful for controller design, (2) model-based control methodologies that strategically use any wing aerodynamic state estimates generated from mission planning to improve position and attitude tracking performance, and (3) the derivation of an analytical guarantee of controller stability that both shows the existence of a convergence region of position and velocity error states and estimates the size said convergence region based on a error-state dependent bound on the uncertainty in the aerodynamic forces estimate and maximum moment. All contributions are validated in the context of controlling a tail-sitter configuration known as a quadrotor biplane tail-sitter.Ph
ARCLIGHT: Automated Clustering and Curriculum Learning Guided by Human Training
No full textARCLIGHT is an AI fusion system that leverages Large Language Models, perception learning, knowledge graphs, and human guidance to describe high-level concept instances with lower-level attributes and affordances. By combining structured models and unsupervised exploration, ARCLIGHT discovers attributes and affordances in both known and unknown objects, entities, or activities. This enables automated novelty detection, curation of a symbolic knowledge graph, and a dialogue agent that asks discriminating questions. The system’s perception component utilizes Bayesian models to recognize unknown and novel concepts, flag regions of high epistemic uncertainty, and update the knowledge graph based on user interactions. ARCLIGHT can potentially improve human-machine collaboration and advance artificial intelligence in various fields