Chalmers Open Digital Repository
Not a member yet
26247 research outputs found
Sort by
Fine Tuning a Large Language Model for Tactical Decision Making in Level 3 Autonomous Trucks
No full textThis thesis investigates whether a Large Language Model (LLM) can be adapted to serve as the tactical brain of a Level-3 autonomous truck through supervised fine-tuning (SFT). We first generated highway driving scenarios in the SUMO simulator, pairing each coded scenario with high-level maneuvering decisions, which include ACC set speed, time gap, lane change intent, generated by a powerful LLM. The resulting scenario-decision pairs constitute a domain-specific dataset that captures a variety of safety-critical interactions between a self-propelled truck and surrounding traffic. Three open-source modelsMeta-Llama-3.1-8B, Qwen 2.5-14B, and DeepSeek-R1-Distill-Llama-8B-are then fine-tuned with Low-Rank Adaptation (LoRA). A modular control stack separates the LLMs high-level reasoning from a low-level Intelligent Driver Model (IDM) that executes longitudinal and lateral motion, mirroring real-world practice.
Evaluation of SUMO episodes showed that fine-tuning improved the quality of decisions. All models improve the achieve a high success rate. Despite the fact that the fine-tuned LLMs achieved a high success rate, we discovered that the LLMs does not fully learn a perfect set of driving strategies. The LLMs does not completely learn the truck’s lane changing strategy. As a result, the LLMs behaved somewhat clumsily in some scenarios. After fine-tuning, some unsafe decisions were eliminated, which confirms the improvement of logical consistency. The models also generate concise natural language rationales, improving the interpretability and compliance of the system. This study shows that when equipped with a tailored driving dataset and efficient LoRA fine-tuning, a modestly sized LLM can provide a degree of safe, efficient, and interpretable but not perfect tactical decisions for self-driving trucks
Applying and Evaluating Large Language Models for Triage at a Paediatric Emergency Department in a Swedish Hospital
No full textThis thesis aims to explore and evaluate a software system using an LLM at the Paediatric Emergency Department (PED) at Sahlgrenska University Hospital. Approximately 60,000 patients visit the PED annually, while reports of decreasing staff availability and increases in burnout are observed. LLMs have shown potential in medical tasks, however, there is limited knowledge on how they would perform in a real setting. This thesis explored LLMs for streamlining the triage process to address this problem. Design Science Research was applied through three iterations involving interviews, prompt engineering, system-level simulations and human evaluations with nurses and voluntary patients.
15 functional and 10 non-functional requirements covering aspects such as accuracy, relevancy, usability and regulatory compliance were elicited from the stakeholders: nurses, the head of section, data scientists, and infrastructure providers. These were translated into a prototype using four instances of Llama 3.3 70B Instruct with Retrieval-Augmented Generation (RAG), each handling tasks such as generating follow-up questions, suggesting clinical controls and tests, or summarising information. The prototype demonstrated potential to support the triage process in 80% of the cases, showed particularly promising results in terms of accuracy when suggesting controls and generating relevant questions. However, it also exhibited certain limitations. Implementing LLM systems in a PED requires further research, especially on validating information completeness and how the RAG document structure and content affect accuracy
Dusty Star-Forming Galaxies at High-Redshift A study on their star formation with dust and CO emission
No full textUtveckling av AR-applikation med AI-glasögon för förbättrad bilkvalitetsuppfattning
No full textDetta projekt undersöker ett alternativt sätt att interagera med en AI-klient utvecklad av startupföretaget Intended Future, som analyserar bilder på bilar och returnerar en uppskattning av upplevd kvalitet. Det nuvarande sättet att använda klienten innebär att användaren manuellt tar ett foto, till exempel med en mobiltelefon, och laddar upp det via ett webbgränssnitt för att få AI:ns textbaserade tolkning. För att möjliggöra en mer sömlös och handsfree-interaktion har projektet utvärderat att istället använda AR via AI-glasögon som ett alternativt sätt att interagera med en sådan AI-klient.
En mobilapplikation utvecklades för att kommunicera med Frame, ett par open-source AI-glasögon från Brilliant Labs. Applikationen parar med Frame via Bluetooth och låter användaren ta bilder med glasögonens inbyggda kamera. Bilderna visas i mobilens gränssnitt, skickas till ett API, och den resulterande texten returneras och kan presenteras via Frames display, via mobiltelefonens gränssnitt, och/eller genom text-to-speech. Användaren kan även välja mellan olika interaktionssätt, som att ta bilder genom att ta på Frame eller använda en knapp på mobiltelefonens gränssnitt, välja mellan att använda Frames kamera eller mobilkameran, eller välja om texten ska visas i Frames display eller inte.
Syftet är att utvärdera användbarheten av denna typ av interaktion genom användartester. Testresultatet visade en preferens för att ta bild genom att ta på glasögonen framför att använda en knapp på mobilapplikationens UI, samt en preferens att ta del av svarsinformationen i form av text via glasögonens display
Enabling a Broader Use of Time-Coupled Building Information Modeling: Understanding current barriers and investigating takt planning as a catalyst for implementation
No full textTime-coupled Building Information Modeling (BIM) is a powerful tool for integrat ing BIM and production planning. This paper reviews the literature and interviews
conducted with practitioners in the construction industry to test the benefits men tioned in the literature. The interviews showed that StreamBIM has the potential
to integrate BIM into existing workflows and tools. It also enables the development
of automated methods for this integration. In addition, the interviews show that
BIM’s flexibility and ability to visualize 3D for production planning can be improved
by combining BIM with a takt-time plan. The results from the interviews suggest
that iterative and continuous usage of time-coupled BIM needs to be done to make
on-boarding easier for both clients, designers, and contractors
Predictive AI for Hepatic Safety: A dual analysis of CYP450 time-dependent inhibition and trapping assays using supervised learning models
No full textThis work explores the development and evaluation of machine learning models for predicting toxicity-related endpoints, focusing on time-dependent inhibition of cytochrome P450 enzymes and reactivity in trapping assays (glutathione, potassium cyanide, and methoxylamine). A variety of modeling strategies were assessed, including decision trees and Chemprop neural networks in both single-task and multitask configurations. Model performances were estimated using temporally split datasets to better reflect real-world prediction scenarios. While tree-based models consistently delivered more stable and balanced results, Chemprop models showed greater sensitivity to class imbalance, data partitioning, and representation. Attempts to mitigate these issues using data resampling techniques, additional molecular descriptors, and scaffold-based data reduction led to limited improvements. Further analysis of feature distributions and chemical space connectivity highlighted key challenges, such as weak class separation in descriptor values and structural isolation of test compounds, especially under temporal splits. In the case of trapping assays, multitask learning failed to improve generalization, likely due to the biological heterogeneity of the endpoints. Overall, results emphasize that data limitations are the primary bottleneck. Enhancing chemical diversity, improving feature representations, and tailoring models to specific endpoint properties appear critical for achieving more robust predictions in toxicity modeling
Design and Optimization of Exhaust Manifold For Volvo Penta D6
No full textThe Volvo Penta D6 is a high-performance diesel engine specifically designed for marine applications. Its efficiency relies heavily on the turbocharger, making it crucial
to retain as much energy as possible from the exhaust gases. However, the current water-cooled exhaust manifold reduces the available thermal energy by cooling
down the exhaust gases, which affects turbocharger efficiency. The water-cooling of the exhaust manifold is necessary due to regulations of 220 ◦C maximum temp for
engine surfaces. The limit is important because engines often operate in tight compartments, sometimes with multiple units nearby, where there is a risk of diesel fuel
leaking directly onto hot engine parts. The water-cooling of the exhaust manifold is also problematic for implementation of future after-treatment systems such as SCR,
as the current exhaust gas temperatures are to low.
This thesis presents a 1-D simulation study in GT-Power to explore how reducing heat loss in the exhaust manifold could enhance engine performance. An engine
model of the D6 engine was analyzed and then used to simulate different scenarios such as changes in manifold geometry, levels of heat loss and turbo chargers.
A few important conclusions are listed below:
• Only changing the geometry of the exhaust manifold while keeping the cooling unchanged is beneficial.
• Reducing the cooling of the original exhaust will not improve the performance, but rather make the engine perform worse. However, combining reduced
cooling with a more capable turbocharger is beneficial.
The simulations showed that changing from a single exhaust pipe to a two-pipe system, without making any changes to the water cooling, improved performance
and fuel efficiency. It resulted in a 4.5% increase in peak power and a 2.5% reduction in Brake Specific Fuel Consumption (BSFC).
Further simulations indicated that reducing the heat lost in exhaust manifold with 78% from the original watercooled exhaust to a insulated dry exhaust system led to
a worse performance, with peak power decreasing by 1.7% and a less desired torque curve with a large dip around 2400 rpm. This occurred because the increased exhaust
pressure and mass flow pushed the turbocharger out of its optimal efficiency zones. Based on those discoveries, a new larger turbocharger was simulated with
the same insulated system, showing increase in performance of +4.3% and BSFC reduction of 2.7%, just by supplying more air efficiently.
Further simulations explored how waste-gating could successfully increase exhaust gas temperature, and how changes to injection timing and amount of fuel complimented
the benefits of increased exhaust enthalpy. A final configuration of the engine that combined a two bank system with exhaust flow from three cylinders in
each bank to get a pulse divided exhaust manifold with less cooling together with a GT45 turbo, advanced timing and a increase in fuel it became possible to success fully meet
Volvo Penta’s target of 550 hp without increasing the NOx emissions.
This thesis demonstrates the possibility to unlock performance potential by reducing the heat being currently lost in cooling. It mentions the technical modifications
required to harness that energy while maintaining emission levels and safety requirements
Parameter Study on a Fully Trimmed Body: Opening Distortion Fingerprint of a Finite Element Model imported to Multibody Dynamics through a Modal Neutral File Compared to Test Data
No full textA complete vehicle model in a multibody dynamics simulation is considered constituting
a key enabler for virtual development within the automotive industry. In
collaboration with Zeekr Technology Europe, a multibody dynamics model of a
robot-taxi has been studied. The multibody dynamics model includes a so-called
fully trimmed body of a finite element model through a so-called modal neutral file.
In particular, the opening distortion fingerprint consisting of diagonals over body
openings and cross-sections, has been used as a measure to evaluate how well the
simulation model correlate to test data. The main topic of this master’s thesis has
been to perform a parameter study in order to investigate what parameters affect
the correlation.
In addition to simulations, a literature review has been conducted in order to understand
the impact from the parameters related to the modal neutral file. From
the literature review, examples of component mode synthesis by Guyan and Craig-
Bampton on a beam are presented. The examples illustrate that the definition of
the interface between the finite element model and the rigid components as well
as the modes included in the synthesis have an impact on how well eigenpairs can
be represented. Additionally, a physical beam has been utilized in order to understand
the impact from rigidbody motions in simulations and on test data. The
physical beam example illustrates that the accelerometers used measure the gravity
field when being rotated. These rotations must be taken into account for lower
frequencies.
The outcomes of the parameter study are that the diagonals over the side doors
and the A-pillar section show a poor correlation to the test data, while the remaining
diagonals show a similar fingerprint to the one obtained from the test data.
Furthermore, none of the studied parameters has shown a significant impact on the
fingerprint over the side doors. Nevertheless, the considered frequency range used to
obtain the fingerprint has shown to impact the correlation over the side doors. However,
all diagonals require further investigation for the fingerprint to be correlated.
The further investigation may include a review of the simulation models, a study
on combined parameter effects, studying animations of the eigenmodes, dive deeper
into the component mode synthesis or review the parameter settings, in particular
the frequency range, of the fingerprint