Repository for Publications and Research Data
News from the ETH Library
Recently Added
RAG-Stack: Co-Optimizing RAG Serving Performance and Quality
Item type: Master Thesis
Haiqiang, Zhang (2026)
Retrieval-augmented generation (RAG), which augments large language model (LLM) generation with information retrieved from databases, has become a widely used approach for knowledge-intensive applications. Modern RAG systems, however, expose many configuration choices, such as retrieval indexes, model selections, and how models invoke retrieval. Each configuration yields a different trade-off between answer quality and serving performance, making it challenging to choose the optimal setting for a specific application deployment. We present RAG-Stack, a framework for efficiently discovering quality-performance Pareto frontiers across diverse RAG applications and serving systems. RAG-Stack consists of RAG-PE, an iterative design-space exploration algorithm that selects the next RAG configuration to evaluate; RAG-IR, a workload abstraction for diverse RAG algorithms; and RAG-CM, a performance model that predicts the optimal deployment and serving performance on the given hardware. Together, these components allow RAG-Stack to search the joint algorithm-system configuration space without deploying every candidate and to transfer an existing Pareto frontier to a new serving system. Given the same number of optimization iterations across diverse datasets, the Pareto frontiers found by RAG-Stack cover 52.5% to 153.2% more of the normalized quality-performance space than those found by state-of-the-art configuration-search methods evaluated over the same RAG design space.
Making Space for the Single Woman: Housing, Gender and Social Reform in early 20th Century Europe
Item type: Journal Article
von Wyl J. (2026)
This paper examines how the material need for independent space for single working women in the early 20th century was addressed through a distinct form of architecture: purpose-built women’s residences. At the time, a personal room and independent income were scarce privileges for women, with housing options often limited to shared lodgings, tenement barracks or religious institutions. The shift in female labour patterns due to industrialisation led to an increase in women working in cities. This gave rise to the concept of the 'surplus woman'—a single, working woman viewed as both a modern paradox and an undesired societal phenomenon. The housing issue, exacerbated by economic inequality and limited opportunities for ownership, prompted the development of alternative residential typologies. The paper explores the architectural characteristics and social implications of women’s housing colonies in Switzerland, Germany, and the United Kingdom from the late 19th century to the 1930s through a comparative analysis of case studies. It argues that these residences played a pivotal role in advancing women’s emancipation by offering spaces for visibility and autonomy, while also reflecting broader tensions between reform, respectability, and control. Revisiting these housing models further contributes to current debates on single households, gender and urban planning.
Fast process-level screening of metal–organic frameworks for adsorption separation using 3D classical density functional theory
Item type: Journal Article
Granderath M.; Rehner P.; Gross J.; et al. (2026)
Adsorption-based separation using metal–organic frameworks (MOFs) represents a promising, energy-efficient alternative to conventional separation technologies. While the vast design space of MOFs allows for precise property tuning to specific applications, the sheer scale of the design space renders exhaustive experimental testing impractical. Consequently, materials discovery relies on large-scale computational screening. Current screening workflows typically employ Grand-Canonical Monte Carlo (GCMC) simulations to predict adsorption properties; however, the high computational cost of GCMC often limits the number of materials screened. In this work, we introduce a process-level screening paradigm leveraging the computational speed of GPU-accelerated 3D classical density functional theory (cDFT). Validation against state-of-the-art GCMC simulations demonstrates that cDFT accurately predicts process key performance indicators (KPIs). The KPIs calculated via cDFT generally deviate by less than 5% from GCMC results while reducing computational costs by two to four orders of magnitude. Leveraging this increased efficiency, we screen the publicly available CoRE MOF 2025 database for methane/nitrogen separation using a temperature-swing adsorption process. By evaluating performance across a wide range of feed compositions, from industrial gas streams to dilute methane sources, the study identifies MOF candidates with robust performance profiles. The entire screening of over 5000 MOFs, requiring 460 000 adsorption calculations, was completed in only five days using two GPUs. The work establishes cDFT as a reliable, high-throughput approach for process-informed material discovery.
ATRP depolymerization of lignin-derived polymethacrylates
Item type: Journal Article
Mountaki S.A.; Whitfield R.; Parkatzidis K.; et al. (2026)
The chemical recycling of lignin-derived polymers synthesized by atom transfer radical polymerization (ATRP) is demonstrated, yielding depolymerization efficiencies up to 99%. Key to this work is the complete elimination of lactonization, a deleterious side reaction that has previously been shown to limit the depolymerization yield of petroleum-based analogues.
Accelerating Automotive Perception: An Outlook on Smart Sensing and Edge AI Trends
Item type: Conference Paper
Capogrosso L.; Magno M. (2026)
The transition to highly automated and autonomous driving systems is leading to an explosion in the volume of data generated by automotive sensors. Although modern vehicles are equipped with diverse cameras, radars, and LiDARs to perceive complex dynamic environments, processing this massive influx of data - exceeding 40 Gb/s of raw aggregate bandwidth - presents a significant computational challenge. The existing literature frequently surveys the hardware specifications of these sensors or the high-level algorithmic fusion strategies used for perception. This article offers a different perspective by explicitly focusing on Edge Artificial Intelligence (Edge AI) within the automotive sensing pipeline. We examine how advanced neural network architectures are adapted for deployment directly on sensor nodes or localized Electronic Control Units (ECUs) under stringent automotive constraints. These restrictions include ultra-low latency (under 100 ms end-to-end response time, with 10-30 ms for critical collision avoidance), limited power budgets (addressed by highly efficient accelerators over 100 TOPS/W), and rigorous ISO 26262 safety certifications. By reorganizing the perception taxonomy around compute location (in-sensor, zonal, and centralized) and resource-aware Machine Learning (ML) techniques like INT8/INT4 quantization and structured pruning, this article highlights the critical transition from traditional signal processing to deep learning at the edge. Furthermore, noting that current state-of-the-art techniques achieve only 40-60% of the target capabilities required for autonomous deployment, we outline open research directions in on-device continual learning, self-diagnostics, and sensor fusion, underscoring the need for hardware-software co-design in next-generation automotive perception.